Skill Builder

$skill-builder — Scaffold or absorb a new SKILL.md

Materializes a new skill against the unified template at references/skill-template.md (extracted from anthropics/financial-services). Runs skill-auditor on the new skill as a self-check before declaring success.

If unsure whether the work should be a skill, a Workflow, or an NTM swarm, run $automation-shape-routing first — it is the front door that decides the shape and hands off to the right builder.

⚠️ Critical Constraints

• Template is canonical. All four modes produce SKILL.md files conforming to references/skill-template.md. Do not invent ad-hoc structures. Why: skill-auditor validates against this template; drift creates auditor false-fails.
• Self-audit is mandatory. After every successful build, the build script invokes $skill-auditor against the new skill directory. A FAIL verdict aborts the build. Why: PR-002 (external validation gate) — the builder must not declare its own work complete.
• Codex parity is day-1, not later. from-scratch, from-template, and absorb-external modes must produce both skills/<name>/SKILL.md AND skills-codex/<name>/SKILL.md + skills-codex/<name>/prompt.md. Why: finding 2026-05-03-codex-skill-shape-is-dual-file — codex SKILL.md uses slim frontmatter (no skill_api_version); prompt.md is mandatory; audit-codex-parity.sh is a content scanner that won't catch frontmatter drift.
• Editing an EXISTING skill also needs a manual twin mirror. When you change skills/<name>/references/*.md or SKILL.md, manually mirror the content into skills-codex/<name>/ (runtime-native), THEN run scripts/regen-codex-hashes.sh --only <name>. make regen-all only refreshes the twin's hash record, not its prose — a green ✓ codex hashes over a stale twin looks handled but isn't. Verify with a content diff (grep -c <new-token> on both copies), not the hash exit code. Why: finding 2026-06-16-codex-twin-content-not-auto-mirrored (age-aqu/age-yxl) — regen made the marker self-consistent with a stale twin (0-vs-2 token divergence) and nothing complained. The parity gate now blocks an un-mirrored references/** edit, but the mirror is still a manual step.
• 250-line ceiling on new SKILL.md. Use references/ for overflow. Why: finding f-2026-05-01-025 — every Skill() invocation reloads 5-15KB; multi-lifecycle sessions compound to 150-200KB+ pure scaffolding.
• Clean-room factory inputs only. When using lessons learned from external corpora, read references/agentops-skill-factory.md and use only AgentOps-owned summaries, scripts, and rubrics. Why: productization must improve structure without copying protected third-party skill content.
• Real gate means exit code. Validate with heal-skill --check --strict <skill-dir> and skill-auditor; never infer green from grep/regex output. Why: regex presence checks created false-greens during the 2026-06 scale build.
• One skill directory = one writer. Bulk builds fan out only when each worker owns a distinct new skills/<name>/ plus skills-codex/<name>/; edits to existing skill dirs run in a later serial wave. Why: concurrent writers deleted untracked work and flipped HEAD mid-task.
• Trust repository state, not subagent reports. Before declaring success, inspect git status, generated hashes, final files, and gate exit codes. Why: sandbox-overlay and stale self-reports can claim work that never persisted.
• Clean-room includes names. Do not reuse exact third-party skill names; mint AgentOps-owned names before source skills, Codex mirrors, or wrappers are keyed. Why: provenance/IP safety applies to labels as well as prose and scripts.
• Do not use the Workflow tool as the skill factory. For scale authoring, use deterministic wave scripts or NTM/Agent Mail lanes with one worker per skill. Why: skill creation needs file ownership and durable git evidence, not opaque background self-reporting.

Modes

Workflow

Phase 1: Mode dispatch

scripts/build.sh reads $1 and routes:

build.sh from-scratch <new-skill-name>          # → init.sh --interactive
build.sh from-template <new-skill-name> --like council
build.sh absorb-external <new-skill-name> --from /path/to/SKILL.md
build.sh from-pattern                            # → ao flywheel close-loop

Checkpoint: Confirm with user the new skill's metadata.tier and metadata.dependencies before generation.

Phase 2: Materialize from template

scripts/init.sh reads references/skill-template.md (the canonical template section) and renders a SKILL.md skeleton with frontmatter pre-filled. For from-template, structure is copied from the source skill; section bodies are blanked and replaced with template stubs.

For absorb-external, the external SKILL.md's content (Constraints / Workflow / Output / Quality sections) is preserved verbatim where possible; AgentOps' structured frontmatter is added on top; the external description is reformatted to satisfy description-has-triggers.

Checkpoint: heal-skill --check --strict skills/<new-name> exits 0.

Phase 3: Codex parity

scripts/init.sh invokes $converter skills/<new-name> codex to produce skills-codex/<new-name>/{SKILL.md,prompt.md}. Then trims skill_api_version from the codex SKILL.md (converter may preserve it). Asserts prompt.md exists.

Checkpoint: bash scripts/audit-codex-parity.sh returns clean AND grep -q "^skill_api_version:" skills-codex/<name>/SKILL.md returns nothing.

Phase 4: Self-audit

The build script tail invokes $skill-auditor on skills/<new-name>. WARN is acceptable for v1 skills (e.g., experimental stability). FAIL aborts.

Checkpoint: audit_pass=true in build report.

Phase 5: Factory score overlay

For AgentOps skill upgrades, use the productization score as a patch selector, not as a replacement for skill-auditor:

python3 skills/skill-auditor/scripts/score_agentops_skill.py skills/<name> --markdown

Choose the smallest patch that improves the score while preserving the canonical template and Codex parity constraints.

Phase 6: Scale factory discipline

For more than one skill, run in ownership waves:

1. Create-only wave: one worker per new skill directory.
2. Mutate wave: existing skill directories only after source creation settles.
3. Mirror/package wave: Codex mirrors and generated hashes after the canonical source corpus is complete.

Every wave ends with git status, scripts/regen-all.sh --check, and the relevant target gates by exit code. If ownership overlaps, stop and rescope.

Output Specification

Format: JSON conforming to schemas/build-report.json written to stdout; markdown audit report written to .agents/audits/<skill>-build.md.

Files created (from-scratch mode):

skills/<name>/
├── SKILL.md                         (≤250 lines, full template spine)
├── scripts/
│   └── validate.sh                  (self-validation per AgentOps convention)
└── references/                      (only if expected to exceed 400 lines)
skills-codex/<name>/
├── SKILL.md                         (slim frontmatter — no skill_api_version)
└── prompt.md                        (~10-20 line Execution Profile)

Quality Rubric

• All four modes produce skills that pass skill-auditor PASS or WARN (not FAIL)
• heal-skill --check --strict exits 0 for every generated source and Codex skill directory
• Codex parity files exist and pass slim-frontmatter check
• Batch authoring has one writer per skill directory and validates persisted git state
• Clean-room review covers exact names as well as prose, scripts, and examples
• No SKILL.md exceeds 250 lines (overflow goes to references/)
• Build report JSON validates against schemas/build-report.json
• from-pattern mode prominently marked alpha/passthrough in user output

Examples

Create a new skill from scratch:

$skill-builder from-scratch hello-world
# → interactive prompt: tier? deps? primary deliverable?
# → writes skills/hello-world/SKILL.md + skills-codex/hello-world/{SKILL.md,prompt.md}
# → runs $skill-auditor on the new skill

Clone structure from an existing skill:

$skill-builder from-template my-new-skill --like council
# → mirrors council's section spine; substitutes new metadata

Absorb a skill from anthropics/financial-services:

$skill-builder absorb-external dcf-helper \
  --from ~/dev/financial-services/plugins/vertical-plugins/financial-analysis/skills/dcf-model/SKILL.md
# → preserves Constraints/Workflow/Output content, wraps in AgentOps frontmatter

Troubleshooting

Corpus authoring health

Skill selection is pure LLM reasoning over the description field, so a missing trigger phrase is a skill that silently never fires. The per-skill auditor checks this only as a WARN, so the gap accumulates. Audit the whole corpus at once:

python3 skills/skill-builder/scripts/scan_descriptions.py skills          # remediation report
python3 skills/skill-builder/scripts/scan_descriptions.py skills --strict # exit 1 on any miss

The scanner mirrors skill-auditor's three-form trigger detection and adds a suggested Triggers: stub per offender. See references/skill-authoring-standard.md for the full authoring doctrine and the best-practice-to-enforcement crosswalk.

See Also

• skill-auditor — companion audit gate, invoked by build self-check
• heal-skill — structural hygiene (Pass 1 of skill-auditor wraps heal.sh)
• converter — produces codex parity artifacts
• scaffold — scaffolds projects/components/CI (NOT skills)
• forge — mines transcripts into learnings (different layer)

References

• references/skill-template.md — canonical SKILL.md template + auditor checklist + PRODUCT.md alignment
• references/agentops-skill-factory.md — clean-room factory workflow and productization rules
• references/skill-authoring-standard.md — clean-room best-practices doctrine + best-practice-to-enforcement crosswalk; backs the scan_descriptions.py trigger scanner
• references/skill-builder.feature — Executable spec: mode dispatch, materialize from template, Codex parity bundle, self-audit + factory score (soc-qk4b)
• references/hyper-extract-design-steals.md — authoring rules stolen from Hyper-Extract: the WHAT-vs-HOW (schema vs guideline) contract, canonical {from}\|{rel}\|{to} identifier dedup-key form, and folded graph-designer/yaml-validator/template-optimizer patterns (age-bp1)