Skill Scaffold

Emit seed files for new additions to auto-claude-skills. Ensures agents follow existing patterns from the first file.

When to Use

During DESIGN phase when creating new skills, commands, plugins, hooks, or modules. Co-selects with writing-skills.

Step 1: Identify Addition Type

Ask: What are we creating?

Step 2: Emit SKILL.md Skeleton

Generate based on type. Include frontmatter, tiered detection where applicable, and output contract section.

Description rule: the frontmatter description states what the skill is for and when to use it. Do not summarize the workflow steps — an agent may follow the summary instead of reading the full skill.

Domain skill SKILL.md skeleton:

---
name: <skill-name>
description: <one-line description>
---

# <Skill Name>

<One paragraph purpose statement.>

## When to Use

<Phase and activation context.>

## Step 1: Detect Available Tools

<Tiered detection pattern if the skill depends on external tools.>

## Step 2: <Primary Action>

<Core behavior.>

## Output Contract

<What the skill produces -- artifacts, reports, structured data.>

Optional anatomy sections (discipline / readiness-claim skills)

Add these ONLY where they earn their place. A skill that enforces a discipline, is commonly skipped or shortcut under pressure, or makes a completion/readiness claim should carry the matching section. Keep a section only if it earns its place -- delete the rest. Filler anatomy on tool/domain skills is the anti-goal.

Template (paste, then keep-or-delete per the table above):

## Rationalizations

| Excuse | Reality |
|--------|---------|
| "<excuse the agent makes to skip this>" | "<why that's wrong>" |

## Red Flags

- **<anti-pattern>** -- <why it means stop>

## Verification

Before claiming <the skill's specific completion claim>, confirm -- with evidence, not inference:

- <observable check tied to the claim>

The qualifier after confirm -- is the canonical default; adapt it to the claim's failure mode (e.g. do not infer for state checks, from observed output for execution checks).

Step 3: Emit Routing Entry Snippet

Generate a JSON snippet for config/default-triggers.json:

{
  "name": "<skill-name>",
  "role": "domain",
  "phase": "<PHASE>",
  "triggers": [
    "<regex-pattern>"
  ],
  "keywords": ["<keyword1>", "<keyword2>"],
  "trigger_mode": "regex",
  "priority": 15,
  "precedes": [],
  "requires": [],
  "description": "<one-line description>",
  "invoke": "Skill(auto-claude-skills:<skill-name>)"
}

The routing entry MUST be added to both config/default-triggers.json and config/fallback-registry.json.

The routing entry description field follows the same description rule as the frontmatter: purpose and when-to-use, never workflow steps.

Step 4: Emit Test Snippets

Routing fixture (tests/fixtures/routing/<skill-name>.txt) — the deterministic done-gate. REQUIRED for every owned skill that has a trigger regex:

# <skill-name> routing fixtures
MATCH: <realistic prompt the trigger regex matches>
MATCH: <a second, differently-phrased positive>
NO_MATCH: <a decoy borrowed from another skill's fixture — must NOT match>

Borrow the NO_MATCH decoy from an existing tests/fixtures/routing/*.txt so an over-broad regex fails the gate. tests/test-fixture-coverage.sh fails the build if this file is missing or lacks a MATCH or NO_MATCH line.

Trigger-accuracy eval (skills/<skill-name>/evals/evals.json) — optional but recommended; the held-out LLM trigger eval (skill-creator / skill-eval.yml) reads it. See docs/eval-pack-schema.md.

Content/behavior assertion (for a tests/test-<skill>-content.sh): keep the SKILL.md non-empty + frontmatter name: check.

test_<skill_name>_content_contract() {
    echo "-- test: <skill-name> SKILL.md has required sections --"
    local skill_file="${PROJECT_ROOT}/skills/<skill-name>/SKILL.md"

    # Bash 3.2: declare `local` and assign on SEPARATE lines. `local x=$(...)`
    # swallows the command's exit code (local always returns 0), masking failures.
    local content
    content="$(cat "${skill_file}" 2>/dev/null || echo "")"
    assert_not_empty "<skill-name> SKILL.md exists and is non-empty" "${content}"
    assert_contains "<skill-name> has frontmatter name field" "name:" "${content}"
}

Constraints

• Produces snippets, not complete files. The user or agent integrates them.
• Does not auto-register skills -- registration is a deliberate step during REVIEW.
• Adapts skeleton to skill type (domain, workflow, edge-overlay process).
• All output follows Bash 3.2 and repo JSON conventions.