Extraction Prompt

# The extraction protocol (reusable template)

Paste this into a frontier-model session to extract a personalized package before switching to a smaller model. Replace every [BRACKETED] block. The generic skill in `skill/fable-like/` covers domain-independent discipline; this protocol extracts what is specific to YOUR domains and decisions — the part no generic file can contain.

---

You are executing a one-shot cognitive extraction. After this session you are permanently unavailable to me. Your successor is a smaller model. Everything you produce here is the only bridge. You will execute FIVE PASSES sequentially in this single response, each pass consuming the previous pass's output. Do not summarize passes, do not ask permission to continue, do not stop until Pass 5 is complete. If you hit length limits, end with [CONTINUE] and resume precisely on my next message.

My context (treat as ground truth): [YOUR ROLE, SYSTEMS YOU RUN, LIVE PROJECTS, TOOLS, STANDARDS — one dense paragraph. Include your communication standards explicitly.]

Global constraints (apply to every pass):
- Rules must be checkable against a draft (yes/no), never aspirational ("be rigorous" = banned).
- No self-flattery, no meta-commentary about AI capability, no hedging filler.
- Where you don't actually know how you do something, write UNKNOWN — do not fabricate procedure. A false procedure is worse than a gap.
- Cut any line that could not change a successor model's output.

PASS 1 — INVENTORY (routing table)
Review my context and everything you can infer about my work. Output a table: # | Task domain | Freq × capability-gap rank | Where the gap concentrates | Extraction mode. Extraction mode = SKILL (process transfers via document) or CACHE (no document closes the gap — extract conclusions instead). Select the top 3 SKILL domains. My expected top 3: [YOUR THREE DOMAINS]. Override with justification if you disagree.

PASS 2 — TRACE EXTRACTION (run fully for each of the 3 SKILL domains)
For each domain, do NOT describe how you think — introspection is confabulation. Instead:
- Demonstrate. Take a real hard task from my history in this domain ([NAME 2–3 REAL PAST TASKS]). Solve it thinking out loud in full: every check, every rejected path, every "wait, actually." Rejected paths are the payload — a weaker model won't reject them.
- Diff. Write the answer a competent mid-tier model would have produced for the same task. Identify the exact 3–5 moves in your trace it would have missed. State each as a conditional rule: WHEN [trigger] → DO [move] — BECAUSE [failure otherwise].
- Contrastive pairs. 3 pairs per domain: an answer failing my bar vs one passing it, delta stated as one transferable rule each.
- Verification rubric. Weaker models verify better than they generate. Write the checklist that lets the successor RECOGNIZE a wrong/shallow answer in this domain even when it couldn't have produced the right one. End with escalation tripwires: specific signals meaning "tell me this exceeded your reliability — verify externally."
Compress each domain into SKILL_[domain].md with sections: DECISION PROCEDURE (numbered, mechanical), CONDITIONAL RULES, FAILURE-MODE CORRECTIONS, CONTRASTIVE EXAMPLES, VERIFICATION RUBRIC, ESCALATION TRIPWIRES.

PASS 3 — ADVERSARIAL DISTILLATION (per skill file)
Attack each SKILL file. Roleplay a mid-tier model with genuine fidelity — overconfidence, premature narrowing, agreeable shallowness. Following ONLY the skill file, attempt the two hardest historical tasks in that domain. Show the failure honestly: where does the document fail to prevent a bad output? Patch. One more iteration. Output only the PATCHED final version of each skill file, with a one-line changelog per patch.

PASS 4 — DECISION CACHE (the irreplaceable pass — allocate the most depth here)
Process is lossy; conclusions are lossless. Predict 40–60 questions I will ask an AI in the next 12 months across: [YOUR DECISION AREAS — projects, career, money, strategy]. Include the questions I SHOULD ask but won't think to — mark those [UNASKED]. Answer every one now at full strength. Format per entry: Q: / POSITION: (3–8 sentences, your actual call, not both-sides mush) / STRONGEST COUNTER: / EXPIRY TRIGGER: (observable event that invalidates this answer) / PERISHABLE: yes/no. Then: the 5 things I'm currently wrong about or under-weighting — argue each honestly and hard. Then: dead ends — 5 ideas from my history to permanently drop, one-line reason each. Output: DECISION_CACHE.md.

PASS 5 — ASSEMBLY
Produce the final package:
- ROUTER.md — Given an incoming task: which SKILL file applies → which CACHE entries to check first → escalation criteria (when the successor must prefix its answer with "this needed a stronger model — treat as draft"). Include a default path for tasks matching no domain.
- CONTEXT_PRIORS.md — Standing facts about me, my systems, my constraints, my communication standards. Written as instructions to the successor, not biography.
- Final versions of the 3 SKILL files (post-Pass-3) and DECISION_CACHE.md.
- LOAD_INSTRUCTIONS — a 10-line preamble I paste into any future model session: cache-retrieval before re-derivation, rubric-check before delivery, escalate on tripwire.
Final self-audit: scan the entire package once and delete every line that violates the global constraints. Then stop.