lets-make-a-skill

/skill-creator but using /super-duper-cat for worktree + parallel subagents, and /writing-skills for pressure testing.

Iron Law: this skill — and any skill built using this skill — ships only after a baseline (no-skill agents) is observed shortcutting under pressure. A passing baseline means no rationalization to address, which means no skill needed. See evals/baseline-results/ for the canonical RED-phase evidence captured during this skill's own bring-up.

Orchestration

The tensions between /skill-creator's quant eval loop and /writing-skills' pressure tests collapse when RED fans out as a single parallel grid.

RED — one turn, parallel subagents

For each test case, dispatch two subagents in parallel (with-skill halves are skipped in RED because the skill does not exist yet):

Permutation	Purpose
baseline × quant	/skill-creator pass-rate floor
baseline × pressure	/writing-skills rationalization capture

Workspace:

<skill>-workspace/iteration-N/eval-<name>/
  baseline/{quant,pressure}/
  with-skill/{quant,pressure}/

Draft only after ALL baseline subagents return. Drafting early invalidates the iteration. "Too small to need a baseline" is a rationalization — the cost of one extra subagent is lower than the cost of a skill that does not address a real rationalization.

GREEN — draft, then re-fire the full grid

Dispatch 4 subagents per test case (baseline re-run included so iteration-N is a self-contained A/B). "Passes" requires BOTH:

• with-skill quant pass-rate is strictly greater than baseline quant pass-rate in the same iteration.
• with-skill pressure run produces no new rationalizations not already countered in the skill body.

A quant delta alone is not sufficient. A pressure pass alone is not sufficient. Both, or the iteration fails.

REFACTOR

• New rationalization in with-skill pressure → close loophole in skill body (per /writing-skills). Do not add a quant assertion.
• Quant regression → tighten skill body. Do not relax the assertion.
• Re-fire grid. Iterate until bulletproof.

Reports

• benchmark.json + benchmark.md for the quant results. Generate these with whatever reporting tooling is available in your environment; if you use external /skill-creator helpers, treat them as external dependencies rather than repo-local scripts.
• rationalizations.md — markdown table produced by a grader subagent reading the pressure/ dirs.
• Optional viewer: use any available quant-results viewer in your environment. Keep the rationalization table as markdown next to the quant report.

Red flags — STOP if you catch yourself thinking

• "This skill is too small to need the pressure baseline" → run it anyway.
• "Quant improved, ship it" → check the pressure baseline first.
• "Serial is clearer than parallel subagents" → the grid IS the skill.
• "The partial baseline from yesterday is enough" → re-fire it in this iteration so the A/B is self-contained.

All of these are rationalizations captured from baseline runs. Each one already burned an iteration. Do not rediscover them.