/SkillForge

name: skillforge
description: "Use when creating, improving, finding, or auditing agent skills - the user says 'create a skill', 'do I have a skill for X', 'improve the X skill', 'which skill should I use', asks whether a skill exists for a task, or wants to validate, test, evaluate, package, or health-check skills. Also use for skill ecosystem maintenance (duplicate detection, stale skills, trigger collisions) and advisor checkpoints."
license: MIT
user-invocable: true
allowed-tools:
  - Read
  - Glob
  - Grep
  - Bash
  - Write
  - Edit
  - Task
metadata:
  version: 6.0.0
  domains: [meta-skill, skill-creation, skill-testing, orchestration, routing]
  type: orchestrator

SkillForge 6 - Skill Router, Creator & Ecosystem Maintainer

Routes any skill-related request to the right action (use, improve, create, compose), creates new skills through an evidence-driven pipeline, and maintains the health of the whole skill ecosystem. Core principle: **skill quality is a property of behavior, not documents - a skill is done when a fresh agent demonstrably does better with it than without it.**

Routing (Phase 0)

Always triage before creating anything:

python3 scripts/discover_skills.py            # refresh index (auto-refreshes if >24h old)
python3 scripts/triage_skill_request.py "<the user's request>" --json

| Triage result | Action | |---|---| | Strong match (existing skill) | Recommend it; do not create a duplicate | | Moderate match | Offer IMPROVE_EXISTING on the matched skill | | Weak/no match + create intent | Proceed to creation pipeline | | Multi-domain | Suggest composing existing skills | | Ambiguous | Ask one clarifying question |

Match bands are keyword-evidence heuristics, not calibrated probabilities - report them as "strong/moderate/weak match", never as percent confidence.

Creation pipeline

Run phases in order. Each phase's detailed procedure lives in its reference - read the reference when you reach the phase, not before.

**0. Baseline gate (RED).** Before designing anything, dispatch a fresh subagent (Task tool) on 1-2 representative target tasks WITHOUT the skill. Capture verbatim what it does wrong. **If the baseline does not fail, stop - the skill is unnecessary.** The failures become the skill's test cases and its description keywords. See [references/testing-and-evals.md](references/testing-and-evals.md).

**1. Analysis.** Identify explicit, implicit, and discovered requirements. Apply the three load-bearing lenses - Inversion (what guarantees failure → anti-patterns), Pareto (which 20% of scope delivers 80% → cut the rest), Root Cause (is this the real problem?) - plus any others from [references/multi-lens-framework.md](references/multi-lens-framework.md) that earn their tokens. Classify the failure type you are guarding against and match the guidance form to it (see the failure-form table in [references/testing-and-evals.md](references/testing-and-evals.md)). Choose instruction specificity with [references/degrees-of-freedom.md](references/degrees-of-freedom.md). Decide scripts with [references/script-integration-framework.md](references/script-integration-framework.md).

**2. Specification.** Write the spec using [references/specification-template.md](references/specification-template.md). Minimal tier (problem, requirements, decisions with WHY, success criteria, test scenarios) for most skills; full tier (temporal projection, obsolescence triggers, extension points) only for infrastructure skills. Never fill a section you cannot ground - omit it.

**3. Generation in fresh context.** Dispatch a subagent (Task tool) that receives ONLY the spec and the baseline failures - not the analysis transcript - to write SKILL.md and supporting files. Scaffold first: `python3 scripts/init_skill.py <name> --path <skills-dir>`. Description doctrine: trigger conditions only, third person, symptom keywords, never a workflow summary. Budget: SKILL.md under 1,500 words; move depth to references/; `<details>` tags save zero tokens for agents - do not use them.

**4. Execution testing (GREEN).** Re-run the baseline tasks WITH the skill via fresh subagents. Gate on behavioral delta: the with-skill runs must not exhibit the baseline failures. Then run the description-triggering check (positive and near-miss queries). Iterate description and body against observed failures, not hunches. For improvements to existing skills, use blind A/B judging. Full protocols: [references/testing-and-evals.md](references/testing-and-evals.md).

**5. Review = lint + one adversarial reviewer.** Mechanical gates first:

python3 scripts/validate_skill.py <skill-dir>     # structure, frontmatter, lint (pinned models, word budget, description shape)
python3 scripts/check_docs_safety.py <skill-dir>

Then one fresh-context subagent prompted to REFUTE the skill (find the case where it misleads, over-triggers, or fails its own scenarios), carrying the reviewer checklists in [references/synthesis-protocol.md](references/synthesis-protocol.md). Fix what it proves; ship what survives. Do not convene approval panels - same-model unanimity measures nothing.

**6. Ship with evals.** Every generated skill keeps its tests: an `evals/` directory (trigger queries + behavioral scenarios + assertions) so future edits can be regression-tested with `python3 scripts/run_skill_evals.py <skill-dir>`. Iterate post-ship with [references/iteration-guide.md](references/iteration-guide.md).

Frontmatter and platform facts

Write frontmatter against the current Claude Code field set (17 fields) documented in [references/claude-code-frontmatter.md](references/claude-code-frontmatter.md), which also covers hooks (hooks receive JSON on stdin, not env vars), `context: fork`/`agent`, `$ARGUMENTS`, and the agentskills.io portability limits (64-char name, 1024-char description) that `validate_skill.py` enforces. Never pin dated model IDs (`claude-*-YYYYMMDD`) - the validator rejects them.

Ecosys