meta-skill

The thesis: AI agents writing arbitrary code is risky and unbounded. AI agents filling a structured template under hard validation is bounded and shippable. This skill is the demonstration: a tightly-shaped pipeline where the LLM authors only the parts of a new skill that humans would hand-write anyway, while the bash skeleton, error handling, and gate calls are generated deterministically.

Procedure

bash skills/meta-skill/run.sh "I need a skill that watches X for Y and writes to Z"
bash skills/meta-skill/run.sh --dry-run "<problem>"  # parse + validate, no write

Pipeline

1. Parse the problem string via local Ollama qwen3.5:latest. Extract:

• skill_name (kebab-case, 3-6 words)
• owner_agent (one of: content, builder, ops, sales, research, accountant, trading, hermes)
• description (one sentence, voice-rule-compliant)
• pipeline_steps (3-7 numbered steps)
• trigger (cron / event / manual / on-skill-call)
• inputs (named parameters; empty list OK)
• outputs (paths or stdout shape)

1. Generate SOUL.md (intent + failure mode + voice anchor; 100-200 words).
2. Generate SKILL.md (frontmatter + procedure + status codes; static template

filled with extracted fields).

1. Generate run.sh (skeleton with proper error handling, voice-gate calls

if outputs are text-bound, smoke-test stub, TODO markers for the actual pipeline body).

1. Validate:

a. python3 tools/voice/score.py on the SKILL.md description → must score >= 60. b. python3 tools/voice/scrubber.py on the description → must pass. c. bash -n run.sh → must pass syntax check. d. Smoke-test: bash run.sh --dry-run → must exit non-error (rc != 2) and produce a STATUS= line.

1. Land:

• If all 4 validations pass → write to ~/.openclaw/skills/<name>/

(or ~/.hermes/skills/<name>/ for hermes-owned).

• If any fails → write to data/proposed_skills/cont-17-meta/<name>/

for Cooper review, with a _VALIDATION.md file detailing what failed.

1. Output JSON to stdout: {status, skill_path, smoke_test_result, gates_passed}.

Status codes

STATUS=ok skill=<name> path=<dir> (all gates passed, landed in runtime) STATUS=proposed skill=<name> path=<dir> failed=<comma-list> (some gate failed, landed in proposed/) STATUS=skip reason=<...> (Ollama down, etc.) STATUS=error reason=<...>

Gates

• voice-gate (advisory threshold 60 for SKILL.md description — descriptions

are short and don't need newsletter-tier polish).

• scrubber HARD on the description (it gets exposed in the skill catalog).
• bash -n on run.sh.
• smoke-test --dry-run must produce a STATUS= line and exit clean.

What this skill WILL NOT do

• It does not write the pipeline body of the new skill. The body is

TODO markers the human or a follow-up agent fills in. The meta-skill guarantees the shape and gates, not the substance.

• It does not bypass any of the existing voice/scrubber gates. A

generated skill is held to the same publishing standards as a hand-written one.

• It does not auto-bootstrap a launchd plist. The plist is a Cooper

decision (which schedule? which environment?). The generated skill is callable via bash run.sh from day one but does not run on a cron until Cooper writes the plist.

Why this is the demonstration

A solo founder using DAI OS can speak a problem ("I need a thing that does X") into the system and walk away with a scaffolded skill in 60 seconds. The compounding loop the methodology describes is not hand-waving: this skill makes the catalog grow autonomously while keeping every output under the same gate disciplines as Cooper's hand-written skills.