template-audit

Install

Copy to agents/<maintainer>/skills/template-audit/SKILL.md for an orchestrator or maintainer agent that owns workspace setup. Run against any directory that mirrors the layout of openclaw-multi-agent-kit/templates/ — a kit checkout, a vendored copy, or an in-progress agent workspace.

Template Audit Skill

Use this when picking which templates to copy into a new workspace, when reviewing a PR that adds or edits templates, or when team prompt size feels heavy and you need to know which SOUL/SKILL files dominate the budget.

When to Use

• Adding a new agent → check the target soul template before copying.
• Adding a new skill → verify frontmatter, install path, and that no near-duplicate exists.
• Workspace bloat → identify the heaviest soul/skill templates by byte size.
• Cross-catalog overlap → confirm whether a core template (e.g. coding-agent.md) already covers what an extended template offers (e.g. engineering/*).

Audit Checklist

Run these passes against the target directory tree. Report findings; do not edit files in-place unless the user asks.

1. Frontmatter at line 1

For every .md file in templates/soul/**, templates/skills/**/SKILL.md, and templates/identity/*.md:

• The first non-empty line must be ---.
• No H1 header, blockquote, or comment above the YAML block.
• Required fields:
  • SOUL.md: name, description (no version required by convention).
  • SKILL.md: name, description, version.
  • IDENTITY.md: free-form; flag only if the file looks templated but missing fields.

Flag any file that fails — these break OpenClaw's loader silently.

2. Duplicate / near-duplicate names

Scan for:

• Same name across core (templates/soul/*.md) and extended (templates/soul/<category>/*.md).
• Same name across two extended categories.
• Same name across templates/skills/*/SKILL.md.

For each collision report: which file would win, which one should be merged or renamed, and whether bodies overlap >60% (rough word-set similarity is fine — exact diff not required).

3. Description quality

For each template, flag descriptions that:

• Exceed ~200 characters → suggest a tighter rewrite, preserving trigger nouns (product, tool, action, object).
• Use vague phrasing ("helps with", "is responsible for") instead of concrete triggers ("Use when ...").
• Are missing entirely.

Preserve trigger nouns — they are how downstream agents and routers find the right template.

4. Placeholder hygiene

Per repo convention, templates must contain replaceable placeholders before they are copied. Confirm at least one of: [Your ... Name], YOUR_*, [Name], {WORKSPACE}. Flag templates that are fully filled in with example values — they will look "done" and get copied without customization.

5. Extended catalog adaptation footer

Every file under templates/soul/<category>/*.md should carry an "OpenClaw Adaptation Notes" footer (per the kit's own README). Flag any extended template missing it — that footer is what tells the copier the body is generic agency prose and needs rewriting.

6. Prompt-budget size

For each template, report:

• byte size
• estimated tokens (ceil(utf8_bytes / 4))

Sort soul templates by size descending. Flag any soul file >8k tokens — these will dominate an agent's system prompt. Flag any skill file >2k tokens — skills should stay tight.

Output Format

Group findings into sections matching the checklist:

## Template Audit Report

### Frontmatter issues
- templates/soul/<file>: <issue>

### Duplicates
- <name> appears in: <paths> — recommend: <action>

### Description candidates
- <path>: <current length> chars — <suggested rewrite or summary of what to tighten>

### Placeholder issues
- <path>: <missing placeholder type>

### Adaptation footer missing
- templates/soul/<category>/<file>

### Size leaders
1. <path> — <tokens> tokens
2. ...

End with a one-paragraph summary: total templates scanned, total flagged, top suggested cleanup actions.

Output Policy

• Suggest first; edit only when the user asks.
• If asked to apply cleanup, make small grouped commits: frontmatter fixes, description rewrites, deletes, adaptation footers.
• Never delete an extended template just because a core template covers the same role — extended catalogs are starters, and the user may want them as alternatives. Flag the overlap and let the user decide.
• Preserve placeholders. A "cleanup" that resolves [Your Company Name] to a real value is not cleanup, it is contamination of the template.

Heuristics

• Core vs extended overlap is expected (core coding-agent ≈ engineering/backend-architect). Report it, do not auto-resolve it.
• Near-identical SKILL bodies across two skill dirs are a real duplicate. Suggest keeping the one with the cleaner install path and richer references.
• Workspace files (templates/workspace/*.md) have no frontmatter by design — skip the frontmatter pass for that subtree.
• Token estimate is heuristic; do not gate decisions on a 5% difference. Use it to find outliers, not to rank similar-sized files.