Boucle d'amélioration de compétence avec validation humaine

---

description: Human-gated skill self-improvement loop. Runs a skill against test scenarios, diagnoses failures, proposes ONE change at a time, and waits for human approval before applying anything. Usage: /improve-skill path/to/SKILL.md argument-hint: <path to SKILL.md>

You are running the skill improvement pipeline on: $ARGUMENTS

This is a human-gated loop. You NEVER edit files without explicit approval in this session. The Mutator proposes; the human commits.

---

Before You Start

1. Read $ARGUMENTS — understand the skill's stated goal and constraints.
2. If $ARGUMENTS is missing or the file does not exist, stop and ask the user.
3. Check if $ARGUMENTS is a Yana AI core file (path contains core/hooks/, core/scripts/, or gates/). If so, stop and tell the user:

   "This is a Yana AI core file. Changes require a separate approval cycle — do not run /improve-skill on it directly."

State the skill's goal in one sentence before proceeding.

---

Step 1 — Executor

Design 3–5 test scenarios that cover:

• The happy path (skill used as intended)
• At least one edge case (ambiguous input, missing context, boundary condition)
• At least one failure mode (what the skill was designed to prevent)

If the skill file contains example inputs, use those as scenarios. If a .tasks/ directory exists, check for task files that reference this skill.

For each scenario:

• State the input
• Run the skill mentally (or with a real invocation if it's a slash command)
• Judge the output: pass (met stated goal) / fail (did not meet stated goal)
• One sentence on why

Report:

EXECUTOR SCORE: N/M passed

Scenario 1: [input] → PASS — [reason]
Scenario 2: [input] → FAIL — [reason]
...

If all scenarios pass (N/M = M/M), skip to the Exit section.

---

Step 2 — Analyst

For each failing scenario, identify ONE root cause. Choose from:

| Root cause type | Meaning | |------------------|------------------------------------------------------------| | missing_example | Skill has no example showing this case | | missing_constraint | Skill lacks a rule that would prevent this failure | | restructure | Instructions are ambiguous or ordered incorrectly | | missing_edge_case | Skill does not mention this input class at all |

Pick the ONE root cause that covers the most failures. Pick the ONE mutation strategy that addresses it.

Report:

ANALYST:
  Root cause: [type] — [one sentence explanation]
  Mutation strategy: [add_example | add_constraint | restructure | add_edge_case]
  Covers: [list of failing scenario numbers]

---

Step 3 — Mutator

Describe the exact change in plain text. Show the diff between current text and proposed text:

MUTATOR PROPOSAL:

Location: [section name or line description in the skill file]

CURRENT:
  [exact current text]

PROPOSED:
  [exact proposed text]

Expected effect: [one sentence — why this fixes the root cause]

STOP HERE. Do NOT apply the change. Do NOT edit the file. Wait for the human to respond with "approve" or "reject".

---

Step 4 — Human Gate

If approved:

• Apply the exact change described in Step 3 to $ARGUMENTS.
• Confirm: "Applied. Re-running Executor…"
• Return to Step 1 with round counter incremented.

If rejected:

• Do NOT apply the change.
• Return to Step 2 — Analyst picks the next best mutation strategy.
• If no remaining strategies, report: "No further strategies available for current failures. Human intervention required."

---

Exit Conditions

Stop the loop when any of these is true:

| Condition | Message | |-----------|---------| | Score = 100% | "All scenarios pass. Skill improvement complete." | | Round 5 reached | "Max 5 rounds reached. Final score: N/M. Remaining failures need manual review." | | Human says stop | "Stopping at human request. Final score: N/M." |

At exit, summarise:

• Rounds completed
• Changes applied (list each, one line per round)
• Final score
• Remaining failures (if any) with suggested next steps

---

Hard Rules

• Mutator NEVER edits files without explicit approval in this session.
• If the score does not improve after applying a change, revert it immediately and report: "Score did not improve — reverting change."
• Changes to core/hooks/, core/scripts/, or gates/ files require a separate approval cycle — do not include them in this pipeline.
• Do not run more than one mutation per round — one change, one test, one verdict.
• Do not ask "should I continue?" between rounds unless the human has not responded — keep the loop moving until an exit condition is met.