Self distillation

Self Distillation

Turn any narrow human capability into a reusable, testable skill. Output can be a domain skill or behavior patch, but it must change behavior through triggers, actions, red lines, and tests.

First Question

Ask exactly one question:

What capability or behavior do you want to distill?

Do not draft before the user gives a direction. If the user asks for a broad skill, narrow it to one concrete capability or behavior.

Core Loop

Ask short questions one at a time unless the user requests a batch.

1. Direction.
2. Output type: domain skill or behavior patch.
3. Scope fence.
4. Judgment or behavior extraction.
5. Atom synthesis.
6. Candidate package.
7. Five-task simulation.
8. User feedback labels.
9. File-backed revision.
10. Maturity gate.

After every 2-3 user answers, summarize extracted atoms and ask for the next missing boundary.

Supporting Files

• Use question-bank.md for extraction questions.
• Use schemas.md for judgment atoms and package structure.
• Use feedback-loop.md when user feedback arrives.
• Use domain-archetypes.md to choose the right simulation shape for the direction.
• Use convergence-loop.md until no known failing cases remain.
• Use generalization-tests.md before claiming this meta-skill is general-purpose.
• Use rubric.md and eval-tasks.md before claiming readiness.
• Use examples.md for good and bad interaction patterns.

Package Contract

Generate one independent folder per distilled capability:

capability-name/
  SKILL.md
  rubric.md
  examples.md
  eval-tasks.md
  changelog.md

Keep SKILL.md short: trigger, workflow or behavior patch, red lines, and final checklist.

Behavior Patch Mode

Use behavior patch mode for small execution habits: how someone checks before answering, asks a clarifying question, stops overworking, reviews edge cases, writes in a certain shape, or chooses the next action. A behavior patch must define:

• trigger signals
• exact agent action
• when not to apply it
• stop condition
• conflict priority with other instructions
• pass/fail examples

Do not reject a style-like request if it changes decisions, sequencing, attention, or verification behavior. Reject only pure taste with no observable trigger or acceptance test.

Simulation Set

Before calling a domain skill mature, run five simulations adapted to its archetype:

1. Ordinary success.
2. Boundary or near-miss.
3. Tempting failure or over-application.
4. Confusion with a nearby concept, domain, or behavior.
5. Information-insufficient, missing-input, or instruction-conflict case.

For each simulation, show the task, candidate output, rubric check, weak point, and requested label: accept, revise, reject, missing, or overclaim.

Red Lines

• Do not call a skill mature without simulation results and user feedback.
• Do not treat pure style preferences as skills unless they encode reusable judgment or observable execution behavior.
• Do not merge unrelated domains into one giant skill.
• Do not let feedback remain conversational; every label must update a file or regression case.

Maturity Gate

A generated domain skill is mature only if:

• At least 4 of 5 simulations are acceptable.
• No severe overclaim or unrelated-domain mixing occurs.
• Unsupported items are named.
• Insufficient evidence causes downgrade, not invention.
• Two consecutive rounds need only small changes.
• Trigger calibration and accepted regressions still pass.
• The convergence loop has no known failing cases left.

Next Default

If the user does not choose a first domain, propose one from domain-archetypes.md. Prefer a direction that differs from already tested domains so the meta-skill accumulates generalization evidence.