Automation Skill Forge

Goal: turn recurring automation work and GitHub/source research into sharp, non-overlapping Codex skill candidates or a clear decision not to create any.

Naming Regex

Generated skills from this automation must use:

^auto-skill-(research|build|review|audit|ops|safety|trend)-[a-z0-9]+(?:-[a-z0-9]+){1,4}$

Format: auto-skill-<mode>-<domain>-<artifact>.

Examples:

• auto-skill-trend-agentic-disease-intake
• auto-skill-safety-mcp-secret-guard
• auto-skill-build-github-repo-radar

Workflow

1. Read automation memory and the current workspace index first.
2. Read references/maintenance-automation-scope.md for the full agency model.
3. Choose a primary aim: research, build, review, audit, ops, safety, or trend.
4. Gather compact evidence with rg, manifests, README/docs, and exact source files.
5. For multiple local source clones, use $auto-skill-build-source-pattern-synthesis to cluster workflow patterns before writing skill files.
6. Create as many useful non-overlapping skills as the evidence justifies.
7. Move broad catalogs or long examples into references.
8. Validate each new or changed skill with $skill-quality-gate.

Skill Candidate Contract

Name:
Regex match:
Generated by:
Aim:
Trigger:
Inputs:
Workflow:
Commands:
References:
Skip when:
Validation:
Useful end goal:

Command Defaults

• Search: rg, rg --files, exact symbols, package manifests.
• Repo scan: scripts/repo_snapshot.sh <repo> and scripts/token_budget.sh <repo> when available.
• Validation: scripts/skill_audit.sh plugins/codex-supercharge/skills and quick_validate.py.
• GitHub packaging: prepare files, docs, validation output, and commit-ready notes only; do not publish unless asked.

Trend Guardrail

Use trend signals to choose domains, not to invent claims. Current trend-backed skill ideas must cite source/date and include a safety or usefulness gate.

Read references/automation-skill-patterns.md and references/maintenance-automation-scope.md before creating generated skills or broad skill batches.

Validation

• Every generated skill name matches the automation regex unless promoted to a stable first-class name.
• Each candidate has a trigger, skip condition, workflow, references, and validation.
• Broad catalogs or examples are moved into references before installation.