Create a New Skill from Topic

Purpose & When-To-Use

Use when a developer specifies a new Skill topic and needs a compliant, token-efficient SKILL.md with a tiny example, tests, and an index entry.

Pre-Checks

• Authoritative time: Set NOW_ET to ISO 8601 in America/New_York using NIST/time.gov semantics. Include NOW_ET in citations’ access dates.
• Scope: Extract intended users, inputs/outputs, success criteria, and token budgets.
• Sources: Identify primary standards/docs. Prefer authoritative sources; record title, URL, and access date (NOW_ET).
• Footprint: Default to Tier 1 unless ambiguity or regulation requires expansion.

Procedure

Tier 1 — Draft Minimal Skill (T1≤2000 tokens)

1. Write METADATA: concise name, slug, ≤160-char description, capabilities, inputs, outputs, keywords, version, owner, license, security, links.
2. Author compact sections: Purpose & When-To-Use, Pre-Checks, Procedure (Tiered), Decision Rules, Output Contract, Quality Gates, Resources.
3. Add one short runnable example (≤30 lines) or precise pseudo-code.
4. Produce index_entry JSON for index/skills-index.json.

Tier 2 — Source-Backed Enrichment (T2≤6000 tokens)

1. Gather 2–4 primary sources; add citations with titles, URLs, and NOW_ET access dates.
2. Add explicit Decision Rules (ambiguity thresholds, abort conditions).
3. Link small resources/ artifacts (schemas/templates) without pasting large texts.

Tier 3 — Deep Validation (T3≤12000 tokens)

1. Create tests/evals_<slug>.yaml with 3–5 scenarios (happy path, missing input, edge case).
2. Add security notes (secrets handling, PII posture, auditability).
3. Emit CHANGELOG.md (v1.0.0) and finalize index_entry.

Decision Rules

• If regulated domains (NIST/OMB/FedRAMP) are involved → require Tier 2.
• If schemas/versions are unknown → emit TODO and stop with needs-input.
• If example exceeds 30 lines → truncate and link to resources/.

Output Contract

• Create skills/<slug>/ containing:
  • SKILL.md (this format)
  • examples/<slug>-example.txt (≤30 lines)
  • resources/ (optional, linked)
  • CHANGELOG.md
• Emit index_entry JSON with: slug, name, summary (≤160 chars), keywords, owner, version, entry.

Examples

Minimal example (pseudo-code, ≤30 lines):

INPUT:
  topic: "OSCAL SSP validator for FedRAMP Moderate"
  constraints: { "strict": true }

OUTPUT (files):
  skills/compliance-oscal-validator/SKILL.md
  skills/compliance-oscal-validator/examples/compliance-oscal-validator-example.txt
  tests/evals_oscal-ssp-validate.yaml
  index_entry (JSON fragment)

INDEX_ENTRY:
  {
    "slug": "compliance-oscal-validator",
    "name": "OSCAL SSP Validator for FedRAMP Moderate",
    "summary": "Validates OSCAL SSP against FedRAMP Moderate baseline requirements.",
    "keywords": ["oscal", "ssp", "fedramp", "validation"],
    "owner": "cognitive-toolworks",
    "version": "1.0.0",
    "entry": "skills/compliance-oscal-validator/SKILL.md"
  }

Quality Gates

• Token budgets: T1≤2k, T2≤6k, T3≤12k — enforce via validate_skill.py.
• Example size: ≤30 lines runnable/pseudo-code; longer samples → resources/.
• Metadata: description ≤160 chars; all required front-matter keys present.
• Sources: T2+ requires 2–4 authoritative links with titles/URLs and access dates (NOW_ET).
• Safety: No secrets, PII, or long pasted texts; validate_skill.py checks for secret patterns.
• Determinism: Index entries must be unique slugs; build_index.py enforces no duplicates.
• Evals: 3–5 scenarios in tests/evals_<slug>.yaml; must pass basic sanity checks.

Resources

• Anthropic Skills Overview (accessed 2025-10-25): https://docs.claude.com/en/docs/agents-and-tools/agent-skills/overview
• Anthropic Skills Best Practices (accessed 2025-10-25): https://docs.claude.com/en/docs/agents-and-tools/agent-skills/best-practices
• Claude Code Skills Documentation (accessed 2025-10-25): https://docs.claude.com/en/docs/claude-code/skills
• NIST Time Services (accessed 2025-10-25): https://www.nist.gov/pml/time-and-frequency-division/time-services
• ISO 8601 Date/Time Format (accessed 2025-10-25): https://www.iso.org/iso-8601-date-and-time-format.html