Part IIOrchestration in Action

The Playbook, Applied

By Jeff Nash

January 14, 2026

8 real-world scenarios where orchestration transforms chaos into compounding.

InvestigationJob HuntTravelMovingPersonal ResetCar BuyingEventsEmergencies
Remember: Power ≠ Autonomy. The motor pedals. You steer.

Example 1 of 8

The 'Vibe Discovery' Investigation

5,000 emails → a prosecution-grade fact pattern timeline

The Problem

A zip file of ~5,000 internal emails. There's a suspicion that "Project Blue" went over budget due to fraud, but there's no proof. You need a chronological "Fact Pattern" timeline showing who knew what, when, with traceable support from the emails.

Why This Is an Orchestration Problem

"AI, read these emails and tell me if there was fraud."

You get a confident story with missing evidence, invented causality, and no chain of custody

Then you spend 12 hours arguing with the model about specifics (Oscillation Trap)

Tools vs Vibes

You don't ask for conclusions. You build a pipeline that produces auditability.

Before AI touches anything, establish a "forensic posture." Not because it's dramatic, but because without this, your output becomes hand-wavy vibes and collapses under scrutiny.

Supervisor Checklist (Human)

  • Work from a copy of the zip, not the original
  • Keep originals read-only
  • Assign a stable ID to every email once parsed (e.g., E-000412)
  • Preserve metadata (timestamps, sender, recipients, Message-ID, thread references)
  • Log every transformation step ("we extracted, normalized dates, deduped")
Output Artifact
METHOD_NOTE.md
Never Let Go of the Handlebars

This is "never let go of the handlebars" applied to data.

Meta-Problem Prompt (Supervisor → AI)

You are the Investigation Supervisor.

We have ~5,000 emails. We suspect Project Blue overran budget due to possible fraud, but we do not assume guilt.

Your job in Act I is NOT to conclude fraud. Your job is to define a methodology to produce a prosecution-grade "fact pattern" timeline.

Output MUST include:

  • • Working definitions: what counts as "fraud indicators" vs normal dysfunction
  • • What counts as a "fact" vs "inference" vs "hypothesis"
  • • A minimal set of fields each timeline entry must contain
  • • A plan for parallel tracks + reconciliation
  • • A plan for contradiction handling
Output Artifacts
INVESTIGATION_SPEC.md
Avoiding the Oscillation Trap

This prevents Ron Burgundy from "reading forward" into conclusions.

Wheel Spinners dump 5,000 emails into a model and pray. Orchestrators index.

Step 1: Parse to Structured Dataset

Every email becomes a record with fields:

email_idthread_idtimestamp_utcfromtoccsubjectbody_textattachments

Step 2: Build Indices

So you don't overload context:

  • • Full-text search index (keywords)
  • • Semantic index (embeddings)
  • • Entity index (people, vendors, amounts, dates)
Output Artifacts
emails.parquet
CORPUS_INDEX.md
Context Economics

This is "tools vs vibes" in your Steering Playbook.

Here's the orchestration move: split by lens, not by random chunks. Each track is a specialist. None is allowed to "solve the case." They produce structured evidence packets.

Track A: Timeline Scout

Identify top 30-60 event candidates: approvals, scope changes, vendor changes, budget updates, escalations.

EVENT_CANDIDATES.md
Track B: Money & Approval Trail

Extract all money flows: budgets, invoices, purchase orders, approvals, reallocations, change orders.

MONEY_TRAIL.md
Track C: Vendor & Relationship Map

Build entity map: vendors, internal owners, approvers. Identify unusual patterns.

VENDOR_MAP.md
Track D: Policy/Compliance Signals

Scan for compliance language: policy exceptions, audit concerns, approval bypasses.

COMPLIANCE_SIGNALS.md
Track E: The Innocence Lens

Build the strongest legitimate explanation for the overrun. Anti-bias track.

BENIGN_CAUSAL_MODEL.md
Track F: Thread Reconstruction

Reconstruct threads and identify missing context: forwarded snippets, missing attachments, gaps.

MISSINGNESS_REPORT.md
Track G: Decision Moments Extractor

Identify decision points where someone chose a path: selecting vendor, approving spend, changing scope.

DECISION_LOG.md

This is your core requirement: merge parallel tracks into one final result.

Reconciler Prompt

You are the Reconciler.

Inputs: EVENT_CANDIDATES, MONEY_TRAIL, VENDOR_MAP, COMPLIANCE_SIGNALS, BENIGN_CAUSAL_MODEL, MISSINGNESS_REPORT, DECISION_LOG.

Produce ONE canonical FACT_PATTERN_TIMELINE.md with:

  • Chronological Timeline — Each entry: timestamp, event title, participants, classification (Fact/Inference/Hypothesis), confidence level, evidence links
  • Contradiction Ledger — Every disagreement across tracks with resolution outcome
  • Actor Dossiers — For key people/vendors: first appearance, role in decisions, key threads
  • Findings Summary — What emails establish, what they don't, what must be investigated next

Quality Gates (Non-negotiable)

  • No timeline entry without evidence IDs
  • No "fraud conclusion" unless evidence meets the defined bar
  • Every major claim must survive the Innocence Lens challenge or be labeled as hypothesis
Parallel Tracks Must Reconcile

This is orchestration: managed intelligence, not "ask for truth."

Final Bundle
METHOD_NOTE.md (how you did it)
FACT_PATTERN_TIMELINE.md (the product)
CONTRADICTION_LEDGER.md
ACTOR_DOSSIERS.md
OPEN_QUESTIONS.md (next steps)
EXHIBIT_INDEX.md (email IDs → quoted snippets)

The Orchestration Insight

You don't ask for conclusions. You build a pipeline that produces auditability. The final FACT_PATTERN_TIMELINE.md is something a lawyer, auditor, or board can actually use because every claim has evidence IDs, confidence levels, and a contradiction ledger.

Principles demonstrated: Tools vs Vibes (deterministic preprocessing), Never Let Go (forensic posture), Parallel Tracks (7 specialist agents), Reconciliation (single truth document).

Example 2 of 8

Resume Optimization + Targeting

Messy folder of old resumes → 50 hyper-customized applications that pass ATS and impress humans

The Problem

You have a messy folder of old resumes and want to apply to ~50 distinct jobs without spamming generic applications. You need 50 hyper-customized applications that pass ATS filters, feel obviously tailored to humans, remain truthful and internally consistent, and are produced efficiently without losing your mind.

Why This Is Orchestration

Paste job description → "rewrite my resume"

Accept output → submit

Repeat 50 times

End up with contradictions, inflated claims, and random keyword soup

Context Economics

You don't ask for "a resume." You design a factory line.

This is the Delegation Paradox: AI can't define "good" unless you define what's true.

Supervisor Prompt

You are my Resume Orchestration Supervisor.

First, you will build a canonical "Truth Core" from my messy resumes and any notes I provide. You must NOT embellish. No inventions. If something is unclear, mark it as unknown.

Output:

  • • MASTER_FACTS.md (roles, dates, projects, tools, measurable outcomes)
  • • ACHIEVEMENT_BANK.md (bullet library with evidence tags)
  • • SKILLS_CANON.md (skills list with proficiency + proof)
  • • ROLE_THEMES.md (3–5 positioning angles I can credibly claim)
Never Let Go of the Handlebars

This is you staying on the handlebars. Without this, Ron Burgundy will happily "fill in" your life story.

ATS and hiring managers care about signal. Job descriptions are noisy.

Job Card Schema

Company + role title
Top 10 keywords (ATS-critical)
Top 5 responsibilities
Top 5 required qualifications (hard)
Top 5 preferred qualifications (soft)
"Implied archetype" (builder, operator, strategist)
Red flags (unrealistic requirements, mismatch)
Your match score (initial)
Output Artifacts
jobs/COMPANY_ROLE/JOB_CARD.md (50 of these)
Garbage Collection

This is compaction + garbage collection: no giant threads, just clean specs.

For EACH job, run multiple specialist agents. They each output parts, not the final resume.

Track A: Keyword & ATS Optimizer

Keywords to include, where to include them, which keywords NOT to force (would look fake).

ATS_PLAN.md
Track B: Human Narrative Builder

2 positioning narratives: technical/operator angle AND strategy/leadership angle with proof points.

POSITIONING.md
Track C: Bullet Selection & Rewrite

Choose best 8-12 bullets from ACHIEVEMENT_BANK, rewrite to mirror job language while staying true.

BULLETS.md
Track D: Gap Handler

Top 3 gaps between JOB_CARD and MASTER_FACTS. Mitigation tactics for each.

GAP_PLAN.md
Track E: Formatting & ATS Compliance

ATS-safe formatting rules, section order optimized for role type, page count decision.

FORMAT_SPEC.md
Track F: Cover Letter Builder

Echoes positioning, references 2-3 proof bullets, avoids generic fluff.

COVER_LETTER.md
Track G: Recruiter Screen Prep

30-second pitch, 3 "tell me about a time…" stories using STAR, top 8 likely questions + answers.

SCREEN_PACK.md

This is the most important part: you do not submit fragments.

Reconciler Prompt (Per Job)

You are the Application Reconciler.

Inputs: JOB_CARD, ATS_PLAN, POSITIONING, BULLETS, GAP_PLAN, FORMAT_SPEC, COVER_LETTER, SCREEN_PACK.

Produce ONE final application packet with:

  • • RESUME_FINAL.md in ATS-safe format
  • • COVER_LETTER_FINAL.md
  • • SUBMISSION_NOTES.md (which narrative chosen, top keywords, any caution claims)

Contradiction Handling

  • If ATS_PLAN wants keywords that conflict with truth, reject them
  • If POSITIONING conflicts with BULLETS, resolve by selecting the most evidenced narrative
  • If job asks for something you don't have, GAP_PLAN must decide: mitigate, disclose, or omit
Stochasticity Management

This is steering. You're forcing structure so the model can't fail silently.

Now you add a global supervisor layer, because 50 jobs creates drift.

Global "Consistency Auditor"

Scans all RESUME_FINAL outputs and detects inconsistencies: job titles/date mismatches, tool claims that appear in some but not others, conflicting seniority signals.

→ CONSISTENCY_REPORT.md

Global "Outcome Tracker"

Job → status → which narrative used → callbacks → notes. Weekly: sample 10 applications' outcomes, adjust weighting rules and narratives.

→ That's compounding.

Truth Core

Job Cards

Parallel Tracks

Reconcile

Audit

The Punchline

This works because you stopped treating AI like Google ("write my resume"), and started treating it like an electric motor you steer: context is money, structure prevents drift, stochasticity is managed (multiple drafts, evaluator gates), and outputs are reconciled into a single truth packet every time.

Principles demonstrated: Context Economics (Truth Core + Job Cards), Stochasticity Management (parallel drafts), Never Let Go (no invented facts), Garbage Collection (consistency audits).

Example 3 of 8

The "Zero-Meltdown Trip" Orchestrator

A travel OS that turns chaos into a plan that survives reality

The Problem (Wheel Spinner Version)

You're planning a trip and you do the classic 2026 failure mode: paste a giant blob of thoughts into ChatGPT, ask "make me the perfect itinerary," and get a beautiful story that ignores half your constraints, teleports between neighborhoods, schedules a 9am museum after a red-eye, assumes restaurants with no reservations, and crams every day until the trip becomes a punishment.

Then reality hits → you crash → you panic edit everything manually. That's the Oscillation Trap: Magic mode → crash → micromanage mode. You get the worst of both worlds.

"AI, plan my perfect trip to Tokyo"

Get a beautiful but impossible schedule

Discover constraints were ignored on day 2

Spend the trip manually fixing everything

Avoiding the Oscillation Trap

Stop asking for answers, start designing processes.

A Concrete Scenario

Destination

Tokyo + Kyoto (10 days)

You + partner + your dad (70, knee issues)

Hard Constraints

  • • No more than 12k steps/day for Dad
  • • One "anchor" activity per day max
  • • 2 food allergies (shellfish + peanuts)
  • • Nothing before 9:30am
  • • Budget: $800/day total (soft cap)

Before anything else, create a tight "Trip Spec" artifact that becomes the only input most agents need.

Supervisor Prompt (Trip Program Manager)

You are my Trip Orchestration Supervisor.

Your job is NOT to generate an itinerary yet. Your job is to capture constraints and define what "success" means so the trip cannot fail silently.

Ask me at most 12 questions, then output TRIP_SPEC.md with:

  • • Travelers (names, ages, mobility/food constraints)
  • • Hard Constraints (Non-negotiable)
  • • Strong Preferences / Soft Preferences
  • • Budget Model (hard cap vs soft cap, where we splurge)
  • • Pace Model (start times, max steps, buffer rules)
  • • Must-Do / Must-Avoid (per traveler, ranked)
  • • Scoring Weights (rest, novelty, food, cost, convenience)
  • • Definition of Done (what a "great trip" looks like)
  • • Red Flags (what would ruin the trip)
Never Let Go of the Handlebars

If you skip this, the model becomes Ron Burgundy: it will confidently "continue the vibe" into a plan that violates reality.

Wheel Spinners start with "Day 1: do 18 things." Orchestrators start with constraints + geometry.

Track A: Logistics Realist

Arrival/departure times, hotel candidates by neighborhood, transit friction warnings, jet lag strategy.

LOGISTICS_BASELINE.md
Track B: Mobility & Energy Modeler

Daily step budget, rest blocks, "energy curve," accessibility assumptions, fatigue trigger protocol.

PACE_RULEBOOK.md
Track C: Budget Controller

Fixed vs variable costs, per-day target spend bands, budget drift alarms, splurge categories.

BUDGET_BASELINE.md
Track D: Food Constraint Guardian

Safe-food strategy, communication templates, fallback meal patterns, "don't gamble hungry" rule.

FOOD_SAFETY_PLAN.md
Track E: Experience Architect

Top 3 anchor moments per traveler, surprise delight slots, no-plan day design, conflict forecast.

JOY_MAP.md

All of these run in parallel. None is allowed to create the full itinerary yet.

This is where most people fail: they collect outputs and drown.

Reconciler Prompt

You are the Trip Reality Reconciler.

Inputs: LOGISTICS_BASELINE, PACE_RULEBOOK, BUDGET_BASELINE, FOOD_SAFETY_PLAN, JOY_MAP, TRIP_SPEC.

Output ONE canonical TRIP_OPERATING_SYSTEM.md with:

  • Constraint Truth Table — every constraint and how we enforce it
  • Scheduling Laws — "anchor max/day," "buffer after transit," "no early starts"
  • Trip Architecture — Base Camps + day types (anchor/recovery/roam/transit)
  • Risk Register + Mitigations — top 10 risks with prevention
  • Open Questions + Experiments — what must be verified
Garbage Collection

This is garbage collection. Instead of letting 5 threads rot, you compact decisions into a single clean state.

Step 1: Itinerary Skeleton

Day 0–10 labeled by day type, base camp assignment, one anchor slot per day max, roam blocks + rest blocks, explicit buffer placement.

No restaurant names, no train minutiae yet.

Step 2: Map/Reduce Day Modules

Each day becomes a module built in parallel: morning window (2 options), midday food plan, afternoon roam block, rest block, anchor moment, escape hatch.

Step 3: Deterministic Tool Routing

Verify the fragile parts using tools: opening hours, ticket availability, transit time, accessibility notes, reservation requirements.

Key idea: you don't let the LLM "be right." You force it to prove.

One itinerary is one dice roll. Orchestrators sample + evaluate.

Generate 3 Complete Itineraries

Plan A

Ambitious within constraints

Plan B

Balanced

Plan C

Chill + recovery heavy

Final Reconciled Output
FINAL_ITINERARY.md
BOOKING_CHECKLIST.md (what to book, when, priority, fallback)
CONTINGENCY_PLAYBOOK.md (rain day, fatigue day, missed train)
ONE_PAGE_SUMMARY.md (printable, minimal cognitive load)

This is the part most people forget: the trip doesn't end when you press "generate."

Daily Micro-Steering Prompt

Inputs: yesterday fatigue (1–10), today weather, budget status, any new constraints.

Output:

  • • Today's plan (Option 1: normal, Option 2: low energy)
  • • Confirm anchor moment
  • • Confirm buffers
  • • Confirm food safety plan
  • • If budget alarm triggered, apply budget mitigation rule

Keep changes local to TODAY. Do not rewrite the whole trip.

Never Let Go of the Handlebars

This is exactly "never let go of the handlebars": the motor pedals, you steer continuously.

Why This Example Is The Exact Thesis

Power ≠ autonomy: LLM can generate an itinerary, but it won't self-drive your constraints.
Oscillation Trap avoided: no magic → crash → panic. You never let it run wild.
Context economics: small canonical artifacts prevent drift.
Tools vs vibes: verification steps force truth.

Example 4 of 8

"Move Like a Logistics Company" Orchestrator

Declutter → sell → pack → relocate, without turning your life into an open-world survival game

The Problem (Wheel Spinner Version)

Moving is the perfect 2026 failure mode because it looks like one big task ("move"), but it's actually 15 intertwined systems: inventory, deciding what stays/goes, selling/donating, packing standards, fragile handling, address changes, utilities, movers/truck scheduling, cleaning, keys + access, pets/kids stress, and 300 micro-decisions that drain your brain.

"AI, help me move."

You get a generic checklist

You do random tasks out of order

You pack the wrong stuff first

You create 40 unlabeled boxes ("misc 1–40")

You forget the one thing you needed on move day

Avoiding the Oscillation Trap

You're not "asking for a moving plan." You're building a Move Operating System.

Concrete Scenario

Space

2-bedroom apartment + storage closet

Timeline

Move date in 21 days

Constraints

Fragile electronics, sentimental clutter, 1 panicky cat

Investigation had "evidence hygiene." A move has "constraint hygiene." If you skip this, every downstream plan becomes vibes: wrong box order, wrong truck size, wrong day-of access, fragile losses, and move-day chaos.

Supervisor Checklist (Human)

  • Confirm move date + time window (keys available, elevator slot, truck pickup/return)
  • Confirm origin + destination access constraints (stairs, parking, loading dock, gate codes)
  • Define budget ceiling + stress ceiling (money vs time vs sanity tradeoffs)
  • List special items requiring protocol (TVs, monitors, art, instruments, plants, cat)
  • Pick definition of done (what success looks like 7 days after move)
  • Set operating rules (no unlabeled boxes; open-first kit never goes on truck)
Supervisor Prompt (Move Program Manager) — Specific Output Required

You are my Move Orchestration Supervisor.

Your job is NOT to generate a generic checklist. Your job is to build a system that prevents silent failure.

Ask me at most 10 questions. Then output MOVE_SPEC.md using the exact sections below. Use concise bullets. If any field is unknown, mark it UNKNOWN and add a "How to resolve" line.

MOVE_SPEC.md MUST include:

  • • Move Parameters: move date/time window, origin, destination
  • • Access Restrictions: parking, elevator reservations, stairs, loading constraints, building rules
  • • Budget + Tradeoffs: max spend, max time, max stress; what you're optimizing for
  • • Special Items Register: item, risk level, handling protocol, who owns it on move day
  • • Hard Constraints (Non-negotiable) + Soft Constraints (Preferences)
  • • Definition of Done (6 bullets, measurable)
  • • Timeline Phases (T-21 → T-0 → T+7) with gates
  • • Critical Path Guess (what likely blocks everything)
  • • Risk Register (Top 12) + mitigations
  • • Operating Rules (no unlabeled boxes, one room at a time, daily compaction)
Output Artifacts
MOVE_SPEC.md
SPECIAL_ITEMS_REGISTER.md

MOVE_SPEC.md — Example Fields (Code-Style)

move:
  date: 2026-06-14
  origin: "2BR apartment, 3rd floor walk-up"
  destination: "1BR condo, elevator required"
constraints:
  budget_max_usd: 1800
  stress_ceiling: "medium"
access:
  origin:
    parking: "street only"
    stairs: 3
  destination:
    elevator_slot: "10:00–12:00"
    loading_dock: true
special_items:
  - name: "OLED TV 65""
    risk: "high"
    protocol: "box + foam corners; ride in car"
  - name: "cat"
    risk: "high"
    protocol: "safe room + carrier + vet calming plan"
Never Let Go of the Handlebars

This is you refusing to treat "move" like an autonomous task. You're building handlebars.

Wheel Spinners try to "declutter by feeling." Orchestrators build a rough map first. Your inventory doesn't need perfection. It needs categories, rough counts, risk flags, and location.

Tools vs Vibes (Recommended)

Use your phone to do a 90-second video sweep of each room, or quick photo bursts per zone. You're not creating art. You're creating evidence the system can reference.

Inventory Builder Prompt — Specific Output Required

You are the Inventory Systems Engineer for my move.

Inputs: room-by-room photos/video sweeps + my answers to questions about high-risk items.

Output INVENTORY_MAP.md with:

  • • Room list + zones (e.g., "Kitchen: cabinets, pantry, under-sink")
  • • Per-zone rough counts and volume estimate (small/medium/large)
  • • Risk flags: FRAGILE, HEAVY, LIQUIDS, HAZMAT, HIGH-VALUE, SENTIMENTAL
  • • "Do Not Pack Yet" list + reasons (needs daily use, move-day kit, pet supplies)
  • • "Special handling" items that must ride in car vs truck
  • • Unknowns + follow-up questions (max 8)

Constraint: This is a rough map. Do not invent items. Use UNKNOWN rather than guessing.

Inventory Structure (Code-Style)

inventory:
  - room: "Kitchen"
    zones:
      - name: "Upper cabinets"
        volume: "medium"
        flags: ["FRAGILE"]
      - name: "Under sink"
        volume: "small"
        flags: ["LIQUIDS"]
  - room: "Bedroom"
    zones:
      - name: "Closet"
        volume: "large"
        flags: []
pack_exclusions:
  - item: "coffee maker"
    reason: "daily use until T-1"
car_only:
  - item: "laptop + chargers"
    reason: "continuity"
Output Artifacts
INVENTORY_MAP.md
DO_NOT_PACK_YET.md

Split by lens, not by random tasks. Each track produces a structured module with explicit assumptions, inputs, and quality gates — so reconciliation can merge them without hand-waving.

Track B: Declutter & Disposition

Decision rules for keep/sell/donate/trash, sentimental handling protocol, sell-fast playbook with pricing.

DISPOSITION_SYSTEM.md
Track C: Packing Systems Engineer

Labeling standard (room code + box number + priority tag), box taxonomy, "No Orphans" rule.

PACKING_PROTOCOL.md
Track D: Logistics & Cost Modeler

3 complete plans: Full-service, Hybrid, DIY truck. Each with cost, schedule, risk, and failure points.

LOGISTICS_OPTIONS.md
Track F: Admin Eliminator

USPS change of address, utilities transfer, subscriptions, insurance, employer/bank/DMV priority list.

ADMIN_CHECKLIST.md
Track G: Turnover & Access Planner

Cleaning plan, landlord/keys/inspection schedule, move-day hour-by-hour runbook draft.

TURNOVER_PLAN.md
Track H: Stress & Continuity

Pet protocol, keep-life-running zones, meal plan for move week, "decision fatigue reduction" defaults.

SANITY_PLAN.md

Supervisor Quality Gates (Per Track)

  • Each track output must start with: Inputs used, assumptions made, and what's UNKNOWN
  • Every recommendation must map to constraints in MOVE_SPEC.md (or be flagged as in conflict)
  • Every plan must include failure points + mitigations
  • Each output must provide 5–10 concrete next actions with dependencies
Parallel Tracks Prompt — Specific Output Required

You are running as a specialist track in a move orchestration system.

Inputs: MOVE_SPEC.md + INVENTORY_MAP.md (and any relevant artifacts). Do not invent facts.

Your output file MUST include:

  • • Summary (5 bullets max)
  • • Decisions + rules (what we will do repeatedly)
  • • Checklist (10–25 items) with owner + due window (T-21/T-14/T-7/T-2/T-0)
  • • Dependencies: "cannot start until…"
  • • Risks + mitigations
  • • Open questions (max 7)
Stochasticity Management

For logistics, you don't pick one plan blind. Generate 3 variants (Full-service, Hybrid, DIY), then run an evaluator to score each on cost, stress, risk, time, reliability. That's orchestration maturity.

You now have 6–8 modules. Wheel Spinners stop here and drown. Orchestrators reconcile into one master plan with gates.

How Tracks Merge (Reconciliation Mechanics)

Step 1: Normalize

  • • Convert every track into: decisions, tasks, dependencies, risks
  • • Map tasks to timeline windows (T-21/T-14/T-7/T-2/T-0/T+7)
  • • Attach owner (you, partner, mover, building, landlord)

Step 2: Gate

  • • Identify blockers that unlock downstream work
  • • Declare non-negotiable quality gates (labels, fragile protocol, car-only kit)
  • • Reject contradictions or force a decision with tradeoffs
Reconciler Prompt (Move Reconciler) — Specific Output Required

You are the Move Reconciler.

Inputs: MOVE_SPEC.md, INVENTORY_MAP.md, DISPOSITION_SYSTEM.md, PACKING_PROTOCOL.md,LOGISTICS_OPTIONS.md, ADMIN_CHECKLIST.md, TURNOVER_PLAN.md, SANITY_PLAN.md.

Produce ONE canonical MOVE_OPERATING_SYSTEM.md with:

  • • Constraint Truth Table (hard/soft + who enforces)
  • • Workstreams + owners + single source of truth pointers
  • • A dependency graph: tasks as nodes, gates as edges
  • Gating Dependencies (Critical!) — what must happen before what
  • • Critical Path (highlighted) + "time bombs" (things with deadlines)
  • • Timeline (T-21 → T+7) with daily steering checkpoints
  • • Contradiction Ledger + resolutions (explicit decisions)
  • • Quality Gates (non-negotiable) + acceptance criteria

Constraints: Do not add new tasks. Only merge, dedupe, sequence, and label. If a task is missing but required, add it to MISSING_TASKS.md instead.

Example Gating Dependencies

  • "Cannot pack kitchen essentials until open-first kit assembled"
  • "Cannot schedule movers until elevator/parking confirmed"
  • "Cannot sell couch until new place confirmed fits / arrival time fixed"

Quality Gates (Non-negotiable)

  • No box goes on truck without a label (room code + box # + priority tag)
  • No fragile electronics move without handling protocol + photo baseline
  • Open-first kit rides with you (never on truck)
  • Move-day access confirmed in writing (elevator slot / parking / keys)

Dependency Visualization (Mini Timeline)

Gate AAccess confirmed (origin + destination)unlocks → movers booking, load plan
Gate BDisposition rules setunlocks → packing order, box count
Gate COpen-first kit assembledunlocks → pack essentials
Output Artifacts
MOVE_OPERATING_SYSTEM.md
CONTRADICTION_LEDGER.md
MISSING_TASKS.md
Parallel Tracks Must Reconcile

This is managed intelligence. Reality doesn't come from one perfect answer. It comes from controlled merging + explicit conflict resolution.

MOVE_DASHBOARD.md Sections

Today's Top 3 (only 3)
Blocking issues
Workstream status (0–100%)
Next gate to unlock
Donations/sales scheduled
Admin tasks due
Move-day readiness score

Readiness Timeline (Simple Visualization)

T-21
Gates + top risks
T-14
Gates + top risks
T-7
Gates + top risks
T-2
Gates + top risks

Use this to force sequencing: what gate must be unlocked by each checkpoint to avoid move-day improvisation.

Daily Steering Prompt (15 minutes) — Specific Output Required

Based on MOVE_DASHBOARD.md, choose TODAY's Top 3 tasks that unlock the most downstream progress. Do not choose tasks that feel good. Choose tasks that reduce risk or unlock gates.

Output MUST include:

  • • Today's Top 3 with owners + estimated minutes
  • • Blockers (if any) + the next smallest action to unblock
  • • Next gate to unlock + acceptance criteria
  • • One compaction action (delete/close/merge stale tasks)

Wheel Spinners improvise and pay the labor penalty. Orchestrators precompute.

MOVE_DAY_RUNBOOK.md must include
T-2 hours: staging + safe room for pet + open-first kit in car
Load order strategy: heavy first / fragile last (or vice versa)
"Open-first" always travels with you, never on truck
New place setup order: bed + bathroom + kitchen basics first
Failure playbook: mover late, rain, truck too small, elevator broken

This is where people lose weeks: half-unpacked boxes become permanent debt.

Week 1 Protocol

  • Day 1: sleep + bathroom + basic kitchen + Wi-Fi
  • Day 2–3: clothing + work setup
  • Day 4–7: remaining boxes by room
  • Donation/trash sweep
  • Update addresses not completed
Garbage Collection

Weekly Compaction Prompt: Summarize what's done, what's blocked, and delete/close stale tasks. Convert leftover chaos into a new clean plan.

Orchestration Insight

Moving is a multi-domain orchestration problem with hard deadlines and real consequences. The wheel spinner treats it as "pack and hope." The orchestrator treats it as parallel workstreams (logistics, admin, setup) with clear reconciliation points and a single source of truth.

Principles demonstrated: Parallel Tracks (logistics + admin + setup streams), Never Let Go (human-verified timing), Garbage Collection (post-move compaction), Context Economics (canonical MOVING_SPEC).

Example 5 of 8

The "Personal Reset OS"

Energy, sleep, focus, movement — without vibes, without burnout, without pretending your life is a lab you control

(Not medical advice. This is a behavior + workflow system. If you have medical conditions, meds, injuries, or mental health concerns, loop in a clinician.)

The Problem (Wheel Spinner Version)

You decide "I'm fixing my life" and then do the classic 2026 oscillation:

Magic Mode

"AI, make me the perfect routine."

It gives you a heroic schedule with: 5am wakeups, cold plunges, meal prep, 60-minute workouts, meditation, journaling, no sugar, no phone, 10,000 steps, a new personality.

Reality hits: Work happens. You sleep poorly once. You miss a workout. You eat something random. The plan breaks because it wasn't built for drift.

Over-ambitious plan → crash → all-or-nothing self-control

Pay the manual labor penalty: guilt, fatigue, and chaos

Treat habits as a performance you must maintain perfectly

Never Let Go of the Handlebars

AI is the motor. You are the handlebars. Your job isn't to suffer or 'work' harder in the pre-AI sense, it's to steer with precision.

This section exists to stop the model from writing a heroic plan that assumes you have infinite sleep, infinite willpower, and no obligations.

Reality Baseline (Snapshot)

Sleep: average bedtime / wake time + variability
Energy: average 1–10 + worst days
Focus: hours of deep work + top distraction
Stress: 1–10 + typical spike windows
Movement: minutes/week + pain/injury flags
Nutrition: meal source reality (home, takeout, office, delivery)

Supervisor Checklist (Human)

  • Write constraints in plain language (no optimizing yet)
  • Name non-negotiables (work, family, commute, health care)
  • Define time budgets (weekday vs weekend)
  • Define failure modes (bad sleep, travel, sickness, deadline week)
  • Agree on metrics that matter (sleep, energy, focus, movement)

RESET_SPEC schema (field view)

Treat this like a config file. The goal is a stable spec that downstream tracks can obey.

constraintsschedule_windowshealth_flagsnon_negotiablessuccess_metricstime_budgetpreferencesrefusals
Supervisor Prompt (Personal Reset Program Manager)

You are my Personal Reset Supervisor.

Your job is NOT to give me a perfect routine. Your job is to define "good," capture constraints, and prevent silent failure.

You must stay inside reality: work schedule, family obligations, commute, health conditions, meds, and environmental constraints (kitchen access, gym access, budget).

Ask me at most 12 questions:

  • • Sleep schedule (actual, not ideal), wake times, insomnia patterns
  • • Caffeine/alcohol timing
  • • Current movement (even if it's "none")
  • • Pain/injuries/limitations
  • • Typical workday + stress spikes
  • • Food reality (where do meals come from?)
  • • Focus problems (doomscrolling? procrastination?)
  • • What I refuse to do (non-negotiables)
  • • Minimum time budget per day and per week

Then output RESET_SPEC.md with:

  • • Reality Baseline snapshot (sleep/energy/focus/movement)
  • • Constraint Truth Table (hard vs soft constraints)
  • • Time budget (weekday/weekend) + protected windows
  • • Success metrics (2–4) with target ranges
  • • Refusals + preferences (to avoid hidden non-compliance)
Output Artifact
RESET_SPEC.md
Context Economics

LLMs are powerful, but they don't get to decide what your real constraints are. You do.

Wheel Spinners redesign everything on day 1. Orchestrators first measure the current system.

Baseline Prompt (Reality Logger)

You are my Reality Logger for 7 days.

Goal: capture a minimal dataset that reveals drift patterns (sleep → energy → focus → food → movement) without taking more than 60 seconds/day.

Output requirements:

  • • Create `BASELINE_WEEK.md` with 7 entries (Day 1–7)
  • • Each entry must include: bedtime, wake time, energy (1–10), focus (1–10), caffeine after 2pm (Y/N)
  • • Each entry must include: movement minutes + one sentence on biggest friction
  • • At end: compute simple averages + identify top 3 patterns (e.g., late screen → later bedtime)
  • • Produce `BASELINE_SUMMARY.md` as a 1-page synthesis with constraints carried forward

7-Day Tracking (60 seconds/day)

Sleep start / wake time
Energy rating (1–10)
Caffeine after 2pm? (Y/N)
Movement minutes
Late screen use (Y/N)
"Biggest friction today"

Daily record (example)

day: 3
bedtime: 00:40
wake: 07:20
energy_1_10: 5
focus_1_10: 4
caffeine_after_2pm: Y
movement_min: 12
late_screen: Y
friction: "late meeting → doomscroll"

This prevents "I started five habits and I don't know what worked." You just need directional signal, not lab-grade truth.

Output Artifacts
BASELINE_WEEK.md
BASELINE_SUMMARY.md

In this act, you don't ask one model to solve your whole life. You run parallel specialist tracks and force each one to output a tight, reusable module.

Module Packet Schema (shared)

Every specialist must produce the same shape of output so reconciliation is possible.

purposedefaultsrulesfallbacksquality_gatesexceptionsnext_actions
Sleep Architect

Bedtime/wake window, ramp strategy, pre-sleep shutdown routine (10-min + 30-min), caffeine rule, bad night recovery rule.

SLEEP_PROTOCOL.md
Nutrition Minimalist

Default meals (2 breakfasts, 2 lunches, 3 dinners), convenience ladder, snack protocol, grocery list, eating-out rules.

FOOD_SYSTEM.md
Movement Programmer

Weekly structure, minimum viable workout (10 min), standard workout (30-45 min), recovery protocol, progression rule.

MOVEMENT_PLAN.md
Focus & Friction Engineer

Top 5 friction points, environment tweaks, daily focus blocks, interrupt protocol, "low energy day" focus plan.

FOCUS_SYSTEM.md
Sustainability Guardian

Burnout red flags, rest rules, overreach protocol, self-talk constraints, travel/sickness exception plan.

SUSTAINABILITY_GUARDRAILS.md
Specialist Prompt (Template for Each Track)

You are a specialist agent for one track: sleep OR nutrition OR movement OR focus OR sustainability.

Inputs: RESET_SPEC.md + BASELINE_SUMMARY.md. You may propose options, but you must not contradict constraints.

Output requirements:

  • • Produce exactly ONE file named for your track (e.g., SLEEP_PROTOCOL.md)
  • • Include: defaults (what to do on normal days), fallbacks (bad day / low energy), exceptions (travel/sick)
  • • Include: 2–4 quality gates with thresholds (ex: sleep window adherence, caffeine cutoff compliance)
  • • Include: a "don't make it worse" rule (prevent downstream contradictions)
  • • Keep it executable: max 10 bullets per section, no essays
Tools vs Vibes

This is "don't ask Ron Burgundy to be perfect." You build rails: bad-night protocol prevents one bad night from breaking the whole system.

Now we have multiple modules that can conflict: movement wants mornings but sleep says no; food plan requires prep but schedule says impossible; focus blocks conflict with workouts.

Reconciliation Map (parallel → integrated)

Sleep sets the ceiling: bad sleep triggers a gentler movement + focus day
Nutrition defaults reduce decision fatigue and stabilize energy for focus blocks
Movement is programmed around schedule windows, not motivation windows
Focus blocks protect the highest-value work and prevent doomscroll rebound
Sustainability guardrails veto any plan that creates burnout debt
All tracks converge into one daily protocol with fallbacks + gates

Supervisor Checklist (Human)

  • No rule allowed without an owning track (sleep/food/movement/focus/sustainability)
  • Every rule must have a fallback (minimum viable version)
  • Resolve contradictions explicitly (don't leave silent conflicts)
  • Define gating dependencies (sleep stabilizes before intensity increases)
  • Confirm the daily schedule is actually executable in real calendar windows

Progress Visualization (simple)

You don't need a dashboard UI. You need a small number of weekly signals you can steer with.

Sleep
trend
Nutrition
trend
Movement
trend
Focus
trend

Integrated Daily Protocol (field view)

The merge result is a single protocol that a human can execute without running five separate mental threads.

today_protocol:
  morning:
    - sleep_recovery_rule: "if < 6h sleep → no intensity"
    - caffeine_window: "before 2pm"
    - focus_block: "1 x 50min (highest value)"
  midday:
    - default_meal: "Lunch A"
    - movement: "10–30min depending on sleep gate"
  evening:
    - shutdown: "10-min + 30-min ramp"
    - screens_cutoff: "last 45 min"
quality_gates:
  - sleep_window_adherence: ">= 5/7 nights"
  - movement_minimum: ">= 3 days"
  - focus_blocks: ">= 4 blocks"
Reconciler Prompt (Reset Reconciler)

You are the Reset Reconciler.

Inputs: RESET_SPEC.md, BASELINE_SUMMARY.md, SLEEP_PROTOCOL.md, FOOD_SYSTEM.md, MOVEMENT_PLAN.md, FOCUS_SYSTEM.md, SUSTAINABILITY_GUARDRAILS.md.

Output ONE canonical RESET_OPERATING_SYSTEM.md:

  • • Constraint Truth Table
  • • Daily Architecture (morning / midday / evening)
  • • Weekly Architecture (exact days for movement, focus blocks, meal strategy)
  • • Gating Dependencies (e.g., "don't intensify workouts until sleep stabilizes for 10 days")
  • • Contradiction Ledger + resolutions
  • Minimum Viable Day — the fallback plan that keeps momentum even when life is chaos
  • • Quality Gates (2–4 metrics + thresholds)

Reconciliation requirements (no hand-waving):

  • • Show how each track changes the daily protocol (sleep gate → movement intensity, etc.)
  • • Produce an explicit merge table: track rule → where it lives (morning/midday/evening/weekly)
  • • Include a "veto" rule: sustainability guardrails override everything
  • • Ensure the Minimum Viable Day still satisfies the top 2 quality gates
Output Artifacts
RESET_OPERATING_SYSTEM.md
CONTRADICTION_LEDGER.md
Garbage Collection

Instead of letting a dozen habits rot in your brain, you compact into one clean state.

"One plan" is one dice roll. For humans, it's worse: it's one fantasy.

Plan A

Aggressive (within constraints)

Plan B

Balanced

Plan C

Gentle (lowest friction)

Evaluator Prompt (Reality Critic)

Score Plan A/B/C using these criteria:

  • • Compliance with constraints
  • • Sustainability risk
  • • Time budget fit
  • • Expected impact on success metrics

No vibes; tie each claim to baseline patterns.

Output Artifacts
PLAN_EVAL.md
RESET_FINAL.md (chosen plan + first 7 days + Minimum Viable Day + Bad Week fallback)

Daily Micro-Steering (2 min)

Inputs: yesterday's sleep, energy score, schedule constraints today. Output: today's plan (normal), today's plan (minimum viable), one thing to avoid.

Keep changes local to TODAY.

Weekly Compaction (10 min)

Summarize: what worked, what didn't, what to remove, what to keep. Update RESET_FINAL and dashboard. Delete stale rules.

Never Let Go of the Handlebars

This is the "never let go of the handlebars" loop. The motor does the work. You steer with tiny corrections.

Why This Example Is the Thesis, in Human Form

  • Power ≠ autonomy: AI can generate a routine, but it can't self-drive your life constraints.
  • Oscillation Trap avoided: no heroic plan that collapses; minimum viable day prevents spiral.
  • Context economics: spec + baseline reduce noise and drift.
  • Tools vs vibes: simple tracking gives real signal.
  • Stochasticity management: multiple plans + evaluator prevents fantasy lock-in.
  • Garbage collection: weekly compaction prevents "habit thread rot."

Example 6 of 8

"Buy a Used Car Without Regret" Orchestrator

Turn an emotional, high-stakes purchase into a managed intelligence pipeline

(Not financial advice. This is a decision + verification system. For legal/inspection specifics, rely on local professionals.)

The Problem (Wheel Spinner Version)

Buying a used car is a perfect "power ≠ autonomy" trap because it feels like: "Find me a good car under $X."

"AI, find me the best used car under $18k"

Get a pretty list, pick one emotionally

Discover hidden maintenance issues, salvage titles, mismatched trims, dealership fees

Either buy anyway because exhausted, or become paranoid and stall for months

Avoiding the Oscillation Trap

You're not "shopping." You're building a procurement + verification pipeline.

Concrete Scenario

Budget & Use

  • • $18k cap (hard), $350/mo cap (soft)
  • • Daily commute + occasional road trips
  • • Must-haves: safety, reliability, CarPlay, decent cargo

Dealbreakers

  • • Salvage title, flood damage, major accidents
  • • "Project car" vibes
  • • Risk tolerance: conservative (predictability over performance)

Before you look at a single listing, lock the system constraints. This is the difference between “I found a car” and “I executed a purchase program.”

Buyer Profile (Hard Constraints)

  • Budget hard cap (all-in): purchase price + tax + registration + fees
  • Monthly payment cap (if financing) and max term (e.g., 48–60 months)
  • Minimum safety bar (e.g., airbags/ESC + target safety rating)
  • Reliability bar (you would rather be bored than surprised)
  • Non-negotiable features (e.g., CarPlay, backup camera) vs nice-to-haves
  • Deal-breakers: salvage/flood, major accidents, missing title, undisclosed lien
  • Time constraints: how quickly you need a vehicle and how far you can travel

Financing Constraints (If Applicable)

Decide in advance:

  • • Down payment range
  • • Credit union / pre-approval plan
  • • Max APR you will accept
  • • Max monthly + max term

No-go rules for financing:

  • • “Four-square” games / payment-only focus
  • • Add-on products you didn’t request
  • • OTD price changes without justification
Supervisor Prompt (Vehicle Purchase Program Manager)

You are my Vehicle Purchase Supervisor.

Your job is NOT to recommend a car yet. Your job is to define what "good" means and prevent silent failure.

Ask me at most 10 questions. Then output CAR_SPEC.md that includes:

  • • Hard Constraints (Non-negotiable) with numeric caps where possible
  • • Must-haves vs Nice-to-haves (explicitly separated)
  • • Deal-breakers + why (what failure they prevent)
  • • Use Case Profile (commute/road trips/cargo/climate/parking)
  • • Risk Tolerance (conservative/balanced/aggressive) + what it means operationally
  • • Scoring Weights (reliability, safety, cost, comfort, efficiency, features) totaling 100
  • • Definition of Done (1 sentence) + checklist
  • • No-Go Rules (salvage, flood, missing title, etc.)
  • • Verification Gates (what must be verified before purchase)
  • • Evidence Rules: what counts as proof vs seller claims

Output format requirements: include a “Constraints” table, a “Weights” table, and a “Gates” numbered list.

Never Let Go of the Handlebars

Without a spec, you're just asking for vibes.

Used cars are not "good" or "bad." They're risk bundles. Orchestrators don't ask "Is this a good car?" They ask: "What are the failure modes, how do we detect them, and what's the cost if they happen?"

Track A: Market Shortlist Builder

8-12 models that match constraints, 2-3 recommended year ranges each, why it fits, what to avoid.

SHORTLIST.md
Track B: Reliability & Failure Modes

Common failure modes, early warning signs during test drive, severity, "walk-away triggers."

RISK_TABLE.md
Track C: Ownership Cost Modeler

Monthly cost components, maintenance reserve rule, how to compare two cars honestly, hidden costs.

TCO_MODEL.md
Track D: Title/History/Fraud Risk

History checks to run, common dealership tricks, private-party risks, evidence requirements.

TITLE_HISTORY_PROTOCOL.md
Track E: Test Drive + Inspection

Pre-call questions script, on-site checklist, test drive script, PPI checklist interpretation.

INSPECTION_PROTOCOL.md

Evidence Capture (Per Listing)

Make each candidate computable with a consistent record:

listing_idsourceurllocationpricefeesodometervinyear_make_model_trimseller_typeseller_claimsmaintenance_docs

Rule: if you can't fill a field, you don't “assume” it—ask for it or downgrade the listing.

Supervisor Checklist (Inspection / PPI)

  • Never skip VIN capture and title status confirmation
  • Test drive includes city + highway + braking + steering at speed
  • Scan for warning lights and verify all electronics
  • Look for mismatch indicators: paint, panel gaps, odd tire wear
  • If serious issue appears, stop: don't negotiate against your own safety
  • PPI summary must be translated into: immediate fixes, 12-month risk, walk-away triggers
Track Prompt Template (Applies to All Tracks)

You are one specialist track in a used-car purchase program.

Inputs: CAR_SPEC.md + the track's assigned scope. Do not recommend a specific listing yet.

Output requirements (mandatory):

  • • A structured markdown artifact with headings matching the TrackGrid output filename
  • • A table with at least 8 rows of concrete entries (models, risks, costs, checks)
  • • A “Walk-away triggers” section (explicit stop conditions)
  • • An “Evidence required” section (what proof is needed, what does NOT count)
  • • A “Known unknowns” section (what to ask next)

Style constraint: cite sources by name (e.g., “NHTSA recall database”, “owner forum consensus”), but avoid invented links.

Tools vs Vibes

LLMs are good at structuring and synthesizing. Determinism is how you verify reality (history reports + inspection + PPI). You're building brakes.

This is the step wheel-spinners skip: you don't “pick a car.” You merge independent research tracks into a single, auditable decision system.

How Tracks Merge (Reconciliation Map)

Inputs → intermediate tables

  • • SHORTLIST → TARGETS table (model/year/trim guidance)
  • • RISK_TABLE → FAILURE_MODES table (severity + triggers)
  • • TCO_MODEL → COST_MODEL table (monthly + reserve)
  • • TITLE_HISTORY_PROTOCOL → HISTORY_CHECKLIST table (required proofs)
  • • INSPECTION_PROTOCOL → INSPECTION_SCORECARD table (what to record)

Intermediate tables → final matrix

  • • Merge keys: model + year range + trim, plus “candidate packet” ID
  • • Apply CAR_SPEC weights to normalize tradeoffs
  • • Any failed gate becomes a hard reject regardless of score
  • • Contradiction Ledger records disagreements (e.g., model reputation vs recall data)

Decision Matrix Schema (What You Must Produce)

target_idmodel_year_trimfit_scorereliability_scoresafety_scoretco_scorefeatures_scorerisk_flagsgate_statusevidence_requirednotes

This matrix is what you use to choose targets and to evaluate actual listings later.

Reconciler Prompt (Car Decision Reconciler)

You are the Car Purchase Reconciler.

Inputs: CAR_SPEC, SHORTLIST, RISK_TABLE, TCO_MODEL, TITLE_HISTORY_PROTOCOL, INSPECTION_PROTOCOL.

Output ONE canonical CAR_DECISION_SYSTEM.md that includes:

  • • Constraint Truth Table (hard constraints + whether each target can satisfy them)
  • • Top 3–5 Recommended Targets (model + year range + trim guidance + why)
  • • Walk-Away Rules (non-negotiable) and which track justified each rule
  • • Verification Gates (must pass in order) + the evidence required at each gate
  • • Contradiction Ledger + resolution outcomes
  • • Decision Rubric (exact scoring formula using weights from CAR_SPEC)
  • • A Decision Matrix table (at least 6 rows) using the schema above
  • • Scripts & Templates (seller message, phone call, negotiation outline)

Hard rule: if a target fails a gate (e.g., title status), it is marked REJECT even if the score is high.

Verification Gates (Must Pass in Order)

Gate 1: Listing Sanity

Gate 2: Seller Call

Gate 3: Title/History

Gate 4: Test Drive

Gate 5: PPI Pass

Quality Gates (Non-negotiable)

  • No target recommendation without citing which constraints it satisfies
  • No “good deal” claim without total out-the-door estimate and 12-month maintenance reserve
  • Every walk-away rule must map to a failure mode (what it prevents)
Parallel Tracks Must Reconcile

You're not "trusting the model." You're forcing structure and conflict resolution.

"One car" is one dice roll. Orchestrators sample and evaluate.

Listing Triage Agent

Classify each listing: PASS / MAYBE / REJECT. Cite which gate it failed. List questions to ask for MAYBE items.

→ LISTING_TRIAGE.md

Evaluation Packets (Per Candidate)

Listing summary, seller answers, history check results, test drive notes, PPI results, score using rubric.

→ CANDIDATE_PACKET_{id}.md
Listing Triage Prompt (Supervisor → AI)

You are the Listing Triage Supervisor.

Inputs: CAR_SPEC.md + CAR_DECISION_SYSTEM.md + a batch of listings (each with URL + VIN if available).

Output LISTING_TRIAGE.md with:

  • • A table of listings with PASS/MAYBE/REJECT
  • • For every MAYBE/REJECT: cite the exact gate or constraint violated
  • • A “questions to ask” list per MAYBE listing (max 5 questions)
  • • A “what evidence to request” list (VIN photo, maintenance records, title photo, etc.)
Candidate Packet Prompt (Per Candidate)

You are building a candidate evaluation packet for ONE listing.

Inputs: listing details + seller call notes + title/history report summary + test drive notes + PPI notes.

Output CANDIDATE_PACKET_{id}.md with:

  • • Listing summary (fields) + a single-sentence “why it's on the table”
  • • Gate status (Gate 1–5) with PASS/FAIL and evidence
  • • Risk flags: failure modes likely + cost ranges (if known)
  • • Score using the Decision Rubric + show the arithmetic
  • • Next actions (what to verify next) or walk-away rationale

Wheel Spinners negotiate based on vibes. Orchestrators negotiate based on evidence.

Negotiation Strategist Prompt

You are the Negotiation Strategist.

Input: CANDIDATE_PACKET + CAR_DECISION_SYSTEM

Output NEGOTIATION_PLAN.md that MUST include:

  • • Target offer range justified by evidence (comps, defects, market time-on-lot if known)
  • • Concessions list (what you'll accept) + what each concession is “worth”
  • • Walk-away point (OTD) and the exact reason (gate / constraint / cost)
  • • Script for dealer vs private party (tone + order of asks)
  • • Fee defense plan (doc fee, reconditioning, add-ons)
  • • How to handle "today only" pressure (pre-written refusal + next-step scheduling)

Output format: start with “OTD target”, then “talk track”, then “fallbacks”.

Context Economics

Urgency is a UX hack. Orchestrators route around it with rules.

Non-negotiable Gates Before Purchase

  • Title is clean enough for your No-Go rules
  • PPI is completed (unless explicitly waived for extremely low-risk)
  • You understand the total out-the-door cost
  • You can explain the purchase in one sentence without rationalizing
Final Gate Prompt (Human Supervisor Checklist)

Before you sign, produce PURCHASE_DECISION.md with the final proof packet.

Required sections:

  • • Candidate summary (VIN, price, mileage, seller type)
  • • Gate evidence (Gate 1–5) with citations to your notes/reports
  • • OTD breakdown (tax/fees/etc.) + monthly payment if financing
  • • 12-month maintenance reserve plan
  • • Why this purchase matches CAR_SPEC in 1 sentence (no rationalizing)

Most people "win" by buying a car and then lose by neglecting the first 90 days.

POST_PURCHASE_PLAN.md
Immediate maintenance items (based on inspection)
Baseline fluid checks
Tire/brake status
Documentation to keep (receipts, inspection)
90-day monitoring checklist
"If X symptom appears, do Y"
Garbage Collection

If you don't compact what you learned into a plan, you pay later.

Tools vs Vibes

LLMs are good at structuring and synthesizing. But determinism is how you verify reality. History reports + PPI are the truth layer. Orchestrators don't handwave gates. You don't buy until you can explain the purchase in one sentence without rationalizing.

Principles demonstrated: Tools vs Vibes (Carfax + PPI = deterministic verification), Context Economics (structured candidate comparison), Never Let Go (walking away = steering), Garbage Collection (90-day plan).

Example 7 of 8

"Once-in-a-Lifetime Event" Orchestrator

Wedding / surprise party / memorial / reunion — a system that survives emotions, vendors, and time

This is the kind of task that destroys Wheel Spinners because it mixes: logistics (hard constraints, deadlines, contracts), people (feelings, expectations, politics), and performance(it happens live, no "undo"). That combination is exactly where power ≠ autonomy matters most.

"AI, plan my wedding / surprise party / memorial."

Get a beautiful schedule and Pinterest vibe board

Discover venue has restrictions, vendor contract has traps, timeline is impossible

Emotional beats are mis-timed, guests confused, microphone fails during the one speech that mattered

Concrete Scenario: Memorial Celebration

Event Details

  • • 80 guests
  • • Budget cap: $6,000
  • • Venue: community hall (strict end time, limited AV)
  • • Accessibility required (elder guests)

Emotional Goal

  • • "Warm, grateful, not depressing"
  • • Non-negotiables: slideshow, 3 short speeches, music playlist, small ritual
  • • Risks: family dynamics, tech failures, timing overruns
Avoiding the Oscillation Trap

The same pipeline applies to weddings and surprises; memorials just highlight how important emotional pacing is.

The fastest way to blow up an event plan is to skip Act 0 and jump to an itinerary. This is where you turn "vibes" into constraints.

EVENT_PARAMETERS.json (code-style fields)

These are the guardrails every other module must obey:

event_date_windowstart_timehard_end_timeguest_count_estimatebudget_cap_totalvenue_restrictionsaccessibility_requirementsnon_negotiablesmust_avoidstakeholdersdecision_owner

Supervisor Checklist (Human)

  • Confirm venue contract: end time, load-in/out, noise/alcohol rules
  • Confirm budget cap includes taxes/gratuity/fees
  • List 3–5 non-negotiable moments (and what can be cut)
  • Name a single decision owner (prevents family deadlock)
  • Write 3 anti-vibe words (what we must avoid)
Supervisor Prompt (Event Program Manager)

You are my Event Orchestration Supervisor.

Your job is NOT to spit out an itinerary. Your job is to create a spec that prevents silent failure and protects the emotional purpose.

Ask up to 12 questions, then output two artifacts:

  • EVENT_PARAMETERS.json — fill ALL fields; if unknown, set to null and add to OPEN_QUESTIONS
  • EVENT_SPEC.md — purpose, success criteria, failure modes, non-negotiables, and quality gates

Question areas (keep answers short and testable):

  • • Purpose: what do we want guests to feel leaving the event?
  • • Guest list rough count + key stakeholders
  • • Date constraints: fixed date vs range, start time flexibility, hard end time
  • • Venue limitations: capacity, AV restrictions, setup/cleanup windows
  • • Budget constraints: cap, must-haves, must-not-spend categories
  • • Accessibility needs
  • • Non-negotiable moments (speeches, ritual, slideshow, toasts)
  • • Vibe words (3–5) and anti-vibe words (3–5)
  • • Family/politics constraints
  • • What would "ruin" the event

Format requirements:

  • • Provide a Constraint Truth Table (hard vs soft vs preference)
  • • Provide an OPEN_QUESTIONS section with owner + due date
Output Artifacts
EVENT_PARAMETERS.json
EVENT_SPEC.md
Never Let Go of the Handlebars

Without an Event Spec, the model will confidently narrate an event that violates your venue and your people.

Wheel Spinners jump to details ("which caterer?"). Orchestrators start with structure.

EVENT_ARCHITECTURE.md includes:

  • Event phases (arrival → welcome → peak → release → close)
  • Attention curve design (where guests are emotionally & physically)
  • Transition strategy (how people know what to do next)
  • Buffer placement rules
  • Content pacing rules (no speech pileups)

Do NOT pick vendors yet. Define the shape.

Timeline Visualization (conceptual)

Arrival
Find seats, settle
Welcome
Set tone
Peak
Speeches / ritual
Release
Food / mingling
Close
Thanks + exit

Each phase must declare: start signal, end signal, transition owner, and "what gets cut first" if behind.

Architecture Prompt (Event Architect)

You are the Event Architect.

Inputs: EVENT_PARAMETERS.json and EVENT_SPEC.md. You are not allowed to select vendors or invent venue capabilities.

Output EVENT_ARCHITECTURE.md with strict requirements:

  • • 5 phases with estimated durations + buffer policy
  • • An attention curve (what guests feel + what they do)
  • • A transition plan: signage/emcee cues/volunteer cues per transition
  • • A "cut list" ranked by emotional damage (what can be shortened vs never cut)
  • • A rehearsal plan: what must be tested (AV + speakers + ritual)
Output Artifact
EVENT_ARCHITECTURE.md
Context Economics

This is "stop asking for the final output, force structure." You're making the model build bridges step-by-step.

Every module must produce a packet that can be merged. That means: named owners, explicit assumptions, and a schema (not prose).

Module Packet Schema (shared)

module_idinputs_usedassumptionsconstraints_acknowledgedrisksdecisions_neededdeliverablesdependenciesquality_checks

This is how you prevent two modules from quietly disagreeing about end time, guest count, or AV reality.

Parallel Tracks Prompt (Supervisor → Specialists)

You are coordinating six specialist modules for one event.

Inputs: EVENT_PARAMETERS.json, EVENT_SPEC.md, EVENT_ARCHITECTURE.md. Each specialist must output one markdown packet using the Module Packet Schema.

Hard rules:

  • • Use explicit time math; include buffers
  • • Every handoff has a named owner
  • • If you assume something (e.g., "venue has projector"), log it in assumptions + request verification
  • • Output must include a "Dependencies" section that other modules can link to
Module A: Run-of-Show Architect

Minute-by-minute schedule with buffers, responsibility for each transition, speech time limits, Plan A (tight) and Plan B (relaxed).

RUN_OF_SHOW.md
Module B: Guest Experience Designer

Entry flow, seating strategy (accessibility, family dynamics), signage plan, "friction points" list + fixes.

GUEST_EXPERIENCE.md
Module C: AV / Tech Reliability

Mic plan (wired vs wireless + backups), slideshow playback plan, "single point of failure" list + redundancy, failover procedures.

AV_PLAN.md
Module D: Content & Ritual Curator

Speech guidelines, who speaks when (emotional pacing), slideshow arc, ritual design, "anti-cringe" rules, emcee script.

CONTENT_PLAN.md
Module E: Vendor & Budget Controller

Budget table with caps per category, vendor shortlist criteria, contract red flags, cost creep alarms.

BUDGET_VENDOR_PLAN.md
Module F: Disaster & Contingency

Top 15 failure modes (weather, late speakers, AV failure), decision tree for day-of emergencies, "protect the emotional goal" rule.

CONTINGENCY_PLAYBOOK.md
Parallel Output Artifacts
RUN_OF_SHOW.md
GUEST_EXPERIENCE.md
AV_PLAN.md
CONTENT_PLAN.md
BUDGET_VENDOR_PLAN.md
CONTINGENCY_PLAYBOOK.md
Stochasticity Management

This is the 2026 upgrade: you don't rely on one "God Model" to hold everything. You use specialists and then reconcile.

Reconciliation is where parallel tracks become real. You don't "combine documents" — you resolve contradictions and produce one plan that can be executed.

How Tracks Merge (explicit)

Venue/Space (Guest module) ↔ Run-of-Show: transitions must fit the room and flow
Catering (Budget/Vendor) ↔ Run-of-Show: service times lock the schedule
AV (Tech) ↔ Content: slideshow + mic choices determine who can speak where
Guest Management ↔ Contingency: family dynamics + crowding are failure modes
Budget ↔ Everything: if cost forces cuts, the cut list must protect emotional anchors
Architecture ↔ Run-of-Show: phases must match attention curve and buffer rules
Reconciler Prompt (Event Binder Reconciler)

You are the Event Binder Reconciler.

Inputs: EVENT_PARAMETERS.json, EVENT_SPEC.md, EVENT_ARCHITECTURE.md, RUN_OF_SHOW.md, GUEST_EXPERIENCE.md, AV_PLAN.md, CONTENT_PLAN.md, BUDGET_VENDOR_PLAN.md, CONTINGENCY_PLAYBOOK.md.

Output ONE canonical EVENT_BINDER.md with strict requirements:

  • Constraint Truth Table — hard vs soft; include source (which input artifact) for each constraint
  • Master Run-of-Show — Plan A and Plan B; every transition has an owner; time math must sum and include buffers
  • Roles & Responsibilities — who owns setup, emcee, AV, catering, cleanup
  • Integration Points — explicit handoffs: e.g., "slideshow start cue," "mic handoff," "food served cue"
  • Contradiction Ledger — list every disagreement across modules (with resolution and decision owner)
  • Risk Register — top risks with triggers + mitigations + contingency action
  • One-Page Day-Of Sheet — printable cheat sheet

Reconciliation method:

  • • Build a "Master Plan Table" mapping each module deliverable to binder sections
  • • Resolve conflicts by re-checking EVENT_PARAMETERS first (constraints win)
  • • If unresolved, log as decision-needed with owner + deadline

Quality Gates

  • No single points of failure without backups
  • Every transition has a human owner
  • Total content time fits the schedule with buffers (time math shown)
  • Emotional goal is protected (what gets cut first is predefined)
  • Every module assumption is either verified or recorded as risk/decision
Output Artifacts
EVENT_BINDER.md (master)
DAY_OF_ONE_PAGER.md (printable)
Parallel Tracks Must Reconcile

You don't "hope" the plan works; you force reconciliation and embed brakes.

Events fail on details. You verify like an engineer.

Verification Checklist (Must Be Executed)

Test mic + speaker coverage in venue
Run slideshow offline + backup on USB
Confirm end time + overtime policy
Confirm parking/load-in rules
Confirm vendor arrival windows
Print 2 copies of Day-Of sheet
Stage a "go box" (tape, adapters, batteries, scissors)
Tools vs Vibes

This is your "route truth to deterministic systems" applied to real life.

Day-Of Supervisor Prompt (5 minutes before start)

Inputs: current delay status, who has arrived, any tech issues.

Output:

  • • Updated run-of-show for the next 30 minutes only
  • • What to cut if behind
  • • Confirm anchor emotional beats remain intact

Keep changes local. Preserve buffers if possible.

This prevents panic-mode thrashing mid-event.

Wheel Spinners collapse after the event and lose everything they learned.

EVENT_POSTMORTEM.md
What worked
What failed
What to change next time
Vendor notes
Updated templates

Why This Example Is the Thesis

Power ≠ autonomy: AI can generate a schedule; it can't run your event.
Never let go of handlebars: owners, gates, buffers, and contingency trees.
Context economics: spec → architecture → modules prevents drift.
Tools vs vibes: AV is deterministic; verify, don't assume.

Example 8 of 8

Family Emergency Readiness OS

Earthquake, wildfire, storm → an operational system you can execute under stress

The Problem

You want to be prepared for earthquakes, wildfires, storms, or extended outages. Not just "buy supplies": you need an actual operational system: who does what, where things are, how to communicate, and drills that prove you can execute under stress.

Why This Is Orchestration

Buy random survival gear from Amazon lists

Stuff it in a closet, feel "prepared"

When reality arrives: can't find things, no plan, freeze

Having supplies ≠ having a system

Tools vs Vibes

Buying gear is vibes. Having a system you've drilled is deterministic verification.

Before you buy anything or write checklists, define the household you're actually responsible for. Emergency planning fails when it assumes ideal mobility, perfect health, and zero pets.

Supervisor Checklist (Human)

  • List every person who might be present (including kids, elders, roommates)
  • Note mobility constraints (stairs, wheelchair, injuries, car access)
  • Note medical constraints (prescriptions, devices, allergies)
  • List pets + carriers/leashes and who handles them
  • Identify time-of-day variance (work/school schedules) and who is usually home
  • Identify hard constraints: 'must not separate', 'must keep insulin cold', etc.
Household Assessment Prompt (Supervisor → AI)

You are my Emergency Preparedness Supervisor.

I will provide household details. Your job is to produce a constraints-first assessment that all later modules must satisfy.

Inputs I will provide:

  • • People: names/roles, ages, special needs
  • • Mobility: stairs/elevators, vehicles, who can carry what
  • • Medical: prescriptions, allergies, devices (CPAP, inhaler, EpiPen)
  • • Pets: species, carriers, meds, who handles them
  • • Home layout: exits, sleeping locations, utility shutoffs known/unknown

Output MUST include:

  • • A concise household roster
  • • A constraints list (hard vs soft constraints)
  • • A critical-dependencies list (meds, devices, documents)
  • • A "single point of failure" list (e.g., only one driver, only one pet carrier)
  • • Questions to resolve before Act I threat modeling (missing info)

Write:

HOUSEHOLD_PROFILE.mdHOUSEHOLD_CONSTRAINTS.md
Output Artifacts
HOUSEHOLD_PROFILE.md
HOUSEHOLD_CONSTRAINTS.md
Never Let Go of the Handlebars

This is "never let go of the handlebars" applied to your own constraints.

Different locations, different threats. Your preparation should match your actual risks.

Threat Assessment Prompt (Supervisor → AI)

You are my Emergency Preparedness Supervisor.

Use the household constraints from HOUSEHOLD_CONSTRAINTS.md and the location [city/region]. Identify the top 5 most likely emergency scenarios I should prepare for.

For each scenario, output MUST include:

  • • Warning time (none/minutes/hours/days)
  • • Expected duration (hours/days/weeks)
  • • Primary risks (injury, smoke, power loss, water loss, displacement)
  • • Decision triggers (what makes us shelter vs evacuate)
  • • Constraints impact (how our household constraints change the plan)
  • • A "first 60 minutes" action sketch

Write:

THREAT_MODEL.mdDECISION_TRIGGERS.md
Output Artifacts
THREAT_MODEL.md
DECISION_TRIGGERS.md

Split by function, not by random shopping. Each module has its own spec, checklist, and maintenance schedule.

Module Spec Prompt (Supervisor → AI)

You are my Emergency Preparedness Supervisor.

Using THREAT_MODEL.md + HOUSEHOLD_CONSTRAINTS.md, produce 7 module specs. Each spec must be executable under stress and must define what "done" means.

For each module, output MUST include:

  • • Purpose + scope boundaries (what's in/out)
  • • Minimal viable checklist (≤ 20 items) for "ready state"
  • • Storage/locations model (where things live)
  • • Ownership (who maintains it)
  • • Maintenance cadence (monthly/quarterly/annual)
  • • Failure modes (what breaks first under stress)

Write the following files (one per module):

SUPPLIES_SYSTEM.mdEVAC_SYSTEM.mdCOMMS_PLAN.mdMEDS_LOGISTICS.mdDIGITAL_GOBAG.mdHARDENING_PLAN.mdMAINTENANCE_SCHEDULE.md
Track A: Supplies System

Water, food, power. With rotation schedules and storage locations.

SUPPLIES_SYSTEM.md
Track B: Evacuation Plan

Routes, rally points, go-bag contents, 15-minute drill protocol.

EVAC_SYSTEM.md
Track C: Communications

Family check-in protocol, out-of-area contact, radio backup.

COMMS_PLAN.md
Track D: Medical Logistics

Prescription backup, first aid, special needs, medical info sheet.

MEDS_LOGISTICS.md
Track E: Digital Go-Bag

Document copies, account recovery, encrypted USB backup.

DIGITAL_GOBAG.md
Track F: Home Hardening

Shutoff locations, furniture securing, defensive positioning.

HARDENING_PLAN.md
Track G: Maintenance Schedule

Monthly checks, quarterly drills, annual refresh protocol.

MAINTENANCE_SCHEDULE.md
Context Economics

Each module has its own canonical artifact. No giant "emergency prep" thread that rots.

A plan you can't execute in 15 minutes under stress is vibes. You need deterministic verification.

15-Minute Drill Prompt (Supervisor → AI)

You are my Drill Supervisor.

Using EVAC_SYSTEM.md and HOUSEHOLD_PROFILE.md, design a 15-minute evacuation drill that we can run quarterly. The drill must be timed, role-assigned, and produce a retro report.

Output MUST include:

  • • Pre-drill staging checklist (what must already be true)
  • • Roles (who grabs what, who handles pets, who sweeps rooms)
  • • Timer checkpoints (0–5 / 5–10 / 10–15) with pass/fail criteria
  • • Variants (nighttime, one adult absent, pet panics, phone dead)
  • • After-action template (what failed, what to change, what to restock)

Write:

EVAC_DRILL_PROTOCOL.mdAFTER_ACTION_REPORT_TEMPLATE.md
Drill Artifacts
EVAC_DRILL_PROTOCOL.md
AFTER_ACTION_REPORT_TEMPLATE.md

0-5 minutes

  • • Grab go-bags (pre-packed, known locations)
  • • Secure pet + carrier
  • • Grab keys, wallet, phone

5-10 minutes

  • • Load car (staged order)
  • • Final sweep (checklist, not memory)
  • • Check utilities (shutoffs if needed)

10-15 minutes

  • • Depart
  • • Confirm route (primary or alternate)
  • • Send status message to out-of-area contact

Quarterly drill: actually run this. Time it. Fix what doesn't work.

Stochasticity Management

One imagined plan is one dice roll. Drills reveal reality. Sample multiple scenarios.

The 7 modules are not the thing you use during an emergency. They are the source-of-truth inputs. Reconciliation turns them into a small number of artifacts that can actually be executed when you're stressed.

Reconciliation Prompt (Supervisor → AI)

You are the Reconciler.

Inputs: HOUSEHOLD_PROFILE, HOUSEHOLD_CONSTRAINTS, THREAT_MODEL, DECISION_TRIGGERS, SUPPLIES_SYSTEM, EVAC_SYSTEM, COMMS_PLAN, MEDS_LOGISTICS, DIGITAL_GOBAG, HARDENING_PLAN, MAINTENANCE_SCHEDULE.

Your job:

  • • Merge the 7 modules into FOUR final artifacts
  • • Resolve conflicts (e.g., evac route vs mobility constraints) explicitly
  • • Produce quality gates that prevent unusable outputs

Final artifacts MUST be structured as follows:

  • EMERGENCY_BINDER.md — Canonical master doc with sections per threat + annexes (medical sheet, shutoffs, inventory, drill logs)
  • FRIDGE_SHEET.md — One page, printable. MUST include: decision triggers, meetup points, shutoff locations, comms check-in text
  • WALLET_CARD.txt — Ultra-short. MUST include: out-of-area contact, meetup points, QR/shortlink to binder
  • FAMILY_ROLES.md — Role assignments by scenario + by time window (0–5 / 5–15 / 15–60 minutes)

Additionally include a reconciliation appendix in EMERGENCY_BINDER.md:

  • • A "Conflict Ledger" of every contradiction discovered across modules + resolution
  • • A "Missingness List" (unknown shutoff, missing documents, unknown rally point) with assignments
  • • A "Change Log" showing what you edited when merging modules

Quality Gates (Non-negotiable)

  • Every action step must have an owner (person/role), not "someone"
  • Every critical resource must have a storage location (e.g., water in garage shelf 2)
  • Every scenario must define a default posture (shelter/evacuate) + triggers that change it
  • If two modules conflict, it must appear in the Conflict Ledger with a resolution

Artifact Structure: FRIDGE_SHEET

Treat it like a schema:

decision_triggersmeetup_pointsout_of_area_contactutility_shutoffsfirst_60_minutespet_plan

Artifact Structure: WALLET_CARD

Strict minimal fields:

name(s)out_of_area_contactmeetup_point_primarymeetup_point_altbinder_link
Final Output Bundle
EMERGENCY_BINDER.md (complete master document)
FRIDGE_SHEET.md (one-page printable summary)
WALLET_CARD.txt (meetup points + contact protocol)
FAMILY_ROLES.md (who does what in each scenario)
Parallel Tracks Must Reconcile

7 modules → 4 usable artifacts. The binder is comprehensive; the fridge sheet is actionable; the wallet card travels with you.

Supplies expire. Plans become stale. Numbers change. Orchestrators accept this and schedule it.

Maintenance Cadence

  • Monthly: Water rotation, battery check, medication expiry review
  • Quarterly: Evacuation drill (timed), communications test, go-bag audit
  • Annually: Full document refresh, threat model update, equipment replacement
Garbage Collection

This is "threads rot" applied to your safety. Schedule garbage collection or pay when it matters most.

Preparedness is an Operating System, Not a Purchase

Wheel spinners buy gear and feel safe. Orchestrators build systems: modular preparation, deterministic checklists, scheduled drills, and maintenance cadences. When reality arrives, they don't freeze. They execute a plan they've verified.

Principles demonstrated: Power ≠ Autonomy (having stuff ≠ having capability), Tools vs Vibes (drills = deterministic verification), Parallel Tracks (7 modules), Garbage Collection (scheduled maintenance).

Your Turn

Create Your Own Orchestration Plan

You've seen the examples. Now generate a personalized plan for any goal—with parallel tracks, quality gates, and reconciliation steps built in.

Parallel TracksQuality GatesReconciliation StepsShareable Links
Quick Reference

Orchestrator Cheat Sheet

Context Economics

"What does the model need right now? What's noise?"

Compaction

"Is this thread getting long? Time to summarize and start fresh?"

Tools vs. Vibes

"Should this be generated text or a deterministic tool call?"

Stochasticity

"Am I treating one output as destiny? Should I sample multiple?"

Leaky Abstractions

"Did I provide docs, examples, and acceptance criteria?"

Workflow Patterns

"Which pattern fits: Map/Reduce, Evaluator, Router, Planner/Executor?"

Delegation Paradox

"Did I define the skeleton, or am I expecting AI to be architect and builder?"

Before You Submit Any Prompt

  • Have I defined what 'good' looks like?
  • Are my constraints explicit, not implied?
  • Do I have quality gates before accepting output?
  • Am I steering, or have I let go of the handlebars?

The bicycle is electric. You still have to steer.

The Pattern is Clear

Across investigations, job hunts, travel, moving, personal growth, major purchases, events, and emergencies, the orchestration pattern remains the same. Define success, run parallel specialists, reconcile into single sources of truth, add quality gates, and steer continuously.

Power ≠ Autonomy

AI can generate; it cannot self-drive your constraints

Never Let Go

You enforce gates, standards, and verification

Context Economics

Canonical artifacts prevent drift and overwhelm

Tools vs Vibes

Deterministic verification beats confident text

Stochasticity Managed

Multiple plans + evaluator prevents fantasy lock-in

Garbage Collection

Compaction prevents permanent chaos debt

The wheel spinner asks: "What should I prompt?"

The orchestrator asks: "What is the structure of this problem, and how do I assign specialized agents to attack each part while maintaining coherence?"

Stop Oscillating. Start Orchestrating.

The people who win with AI in 2026 aren't the ones who ask for magic. They're the ones who design factory lines, run parallel specialists, and never let go of the handlebars.

The bicycle is electric. You still have to steer.

Read Part I: The Theory