Port provenance: adapted from the pinned upstream source at 984023ddac0d5e27624f2baacde6105e45de375f under MIT; see the repository parity ledger for the exact path and blob.
Improve a Codex skill safely
Improve an existing skill from evidence, not preference. Preserve its public name
and intended trigger unless the user explicitly approves a breaking change.
- Run
$ai-game-studio:skill-test and record the baseline failures.
- Read the official skill authoring constraints available in the current Codex
installation. Inspect all local references before changing instructions.
- Diagnose each failure as routing ambiguity, missing workflow detail, unsafe
mutation, broken reference, unsupported metadata, platform coupling, excessive
context, or unverifiable output.
- Propose the smallest patch, showing affected files, behavior change, possible
regressions, and rollback. Wait for approval before editing.
- Preserve frontmatter with only
name and description. Keep the starter
prompt namespaced and realistic. Disable implicit invocation for sensitive
setup, install, control, refresh, or destructive workflows.
- After approval, patch with the repository editing mechanism. Do not rewrite
unrelated prose or generated assets.
- Re-run static validation and all direct, implicit, near-miss, and negative
forward tests. Compare results with the baseline.
Return the exact diff summary, validation evidence, remaining limitations, and a
one-command rollback. Never claim improvement solely because wording changed.
Codex portability
Use the search, file-editing, shell, user-input, and subagent capabilities available in the active Codex surface. Use PowerShell syntax on Windows and POSIX syntax on macOS/Linux; do not require a Unix compatibility layer on Windows. Inherit the active model and permission mode, and do not weaken approval or sandbox boundaries.
1---2name: skill-improve3description: Improve a skill using a test-fix-retest loop. Runs static checks, proposes targeted fixes, rewrites the skill, re-tests, and keeps or reverts based on score change.4---56> Port provenance: adapted from the pinned upstream source at `984023ddac0d5e27624f2baacde6105e45de375f` under MIT; see the repository parity ledger for the exact path and blob.78# Improve a Codex skill safely910Improve an existing skill from evidence, not preference. Preserve its public name11and intended trigger unless the user explicitly approves a breaking change.12131. Run `$ai-game-studio:skill-test` and record the baseline failures.142. Read the official skill authoring constraints available in the current Codex15 installation. Inspect all local references before changing instructions.163. Diagnose each failure as routing ambiguity, missing workflow detail, unsafe17 mutation, broken reference, unsupported metadata, platform coupling, excessive18 context, or unverifiable output.194. Propose the smallest patch, showing affected files, behavior change, possible20 regressions, and rollback. Wait for approval before editing.215. Preserve frontmatter with only `name` and `description`. Keep the starter22 prompt namespaced and realistic. Disable implicit invocation for sensitive23 setup, install, control, refresh, or destructive workflows.246. After approval, patch with the repository editing mechanism. Do not rewrite25 unrelated prose or generated assets.267. Re-run static validation and all direct, implicit, near-miss, and negative27 forward tests. Compare results with the baseline.2829Return the exact diff summary, validation evidence, remaining limitations, and a30one-command rollback. Never claim improvement solely because wording changed.3132## Codex portability3334Use the search, file-editing, shell, user-input, and subagent capabilities available in the active Codex surface. Use PowerShell syntax on Windows and POSIX syntax on macOS/Linux; do not require a Unix compatibility layer on Windows. Inherit the active model and permission mode, and do not weaken approval or sandbox boundaries.
frabcd