Blcaptain Meta Skill

BLCaptain Meta Skill

Contract

Productize only repeated high-value workflows worth becoming Agent Skills. Keep the entrypoint thin, move detail into resources, prove behavior, and keep audit evidence.

First Move

Inspect source or existing skill. Run the Non-Skill gate. For new Codex skills, prefer skill-creator. Track baseline, decisions, files, commands, and risks.

Workflow

Follow in order: Research queries/artifacts/failures → Analyze NABC/ROI/boundaries → Plan tier/files/evals/gates → Develop metadata, thin SKILL.md, resources, scripts, templates, examples, evals, manifest → Validate structure/links/context/scripts/security → Test baseline, positive, negative, near-neighbor, execution, pressure, regression → Audit blockers → Summarize proof and risks.

Resource Loading

• Stage 0 proof: references/00-real-delivery-first.md.
• Non-Skill gate: references/01-non-skill-decision.md.
• Structure/description/packaging/templates: references/02-skill-design-and-packaging.md.
• Evals/routes/baseline/near-neighbor/failures: references/03-eval-protocol.md.
• Permissions/scripts/supply chain: references/04-safety-and-supply-chain.md.
• Release/governance/versioning: references/05-governance-and-operations.md.
• Platform differences: references/06-platform-differences.md.
• Role handoffs: references/07-role-playbooks.md.
• Worked example: examples/three-scenario-worked-examples.md.
• Visual creator example: examples/visual-card-validation.md.
• Route fixtures: evals/route_cases.json.
• Scenario fixtures: evals/scenario_cases.json.
• Failure families: evals/failure_library.json.
• Forward-test evidence: evals/forward_test_results.json.
• Regression history: evals/regression_history.json.
• Platform schema/tokenizer check: evals/platform_schema_tokenizer_check.json.
• Templates: assets/templates/skill-brief.md, assets/templates/skill-design-spec.md, assets/templates/eval-case.md, assets/templates/gotcha.md, assets/templates/iteration-record.md, assets/templates/governance-manifest.json.

Build Rules

• description routes; it is not marketing.
• Keep SKILL.md small; move detail, checks, and skeletons to resources.
• Create only behavior/evidence-bearing files.
• Use real failures for gotchas.
• Require runnable scripts; never include secrets, destructive defaults, or policy-bypass text.

Useful Commands

python3 scripts/validate_meta_skill.py .
python3 scripts/context_budget.py SKILL.md
uv run --with tiktoken python scripts/context_budget.py SKILL.md --tokenizer openai --encoding o200k_base
python3 scripts/eval_routes.py evals/route_cases.json

For Codex structure, also run:

python3 "${CODEX_HOME:-$HOME/.codex}/skills/.system/skill-creator/scripts/quick_validate.py" .

context_budget.py defaults to an offline heuristic and supports OpenAI tiktoken when installed. agents/openai.yaml follows local skill-creator guidance; recheck target schema before public release.

Output Contract

Report: decision, files/reasons, evidence, blockers/risks, next iteration.

Gotchas

• Do not build a Skill just because the source material is long.
• Do not turn one-off explanation, summary, translation, or brainstorming into a Skill.
• Do not put the whole PRD into SKILL.md.
• Do not treat fixture eval as real model forward-test.
• Do not add governance files to hide missing real examples, evals, or gotchas.
• Do not call a package publishable if scripts cannot run or resources cannot be found.