# Template Maintenance

> Use when maintaining universal-template policy, prompts, skills, templates, MCP declarations, or publication tooling; review semantic coherence and select only relevant exact checks.

- Skill: `ryan-brosas/template-maintenance` (Agent Skill)
- Install (CLI): `npx skillmds@latest add ryan-brosas/template-maintenance`
- Raw SKILL.md: https://api.skillmd.com/api/skills/ryan-brosas/template-maintenance/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: ryan-brosas (https://skillmd.com/u/ryan-brosas)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/ryan-brosas/template-maintenance

---


# Template Maintenance

## Core Principle

The model owns meaning. Deterministic checks own only facts that follow exactly
from source bytes, filesystem state, Git state, or runtime output.

## When to Use / NOT

- **Use when:** changing this template's canonical content or maintainer tools.
- **NOT when:** ordinary work in another repository, or when a direct project
  test already owns the acceptance boundary.

## Workflow

1. Inspect the current diff, nearby owners, callers, and any unmanaged changes.
2. Review changed policy, skills, prompts, and docs together for semantic
   consistency. Judge relevance, overlap, stale artifacts, prose quality, and
   whether each instruction still earns its place.
3. Ground factual claims in current source, Git, host inventory, and runtime
   output. Treat session history as evidence, never current truth.
4. Run only checks that prove an affected exact contract, such as strict YAML
   and metadata types, names, references, tracked-file ownership, disjoint
   hot/cold context and its budget, path containment, generated parity, secret
   patterns, or safe atomic filesystem mutation.
5. Separate hard failures from judgment calls. Fix objective failures; explain
   semantic tradeoffs with evidence instead of inventing a regex proxy.
6. Preserve unrelated files and report the commands that actually ran and their
   real results.
7. Recommend deleting a check when it duplicates reliable model review and has
   no objective contract to prove.

Stop when the affected content is coherent, relevant hard contracts pass, and
remaining uncertainty is reported.

## Red Flags

- Encoding nuanced policy or prose judgment as phrase tables.
- Running every historical script because an old checklist names it.
- Treating generated catalogs, projections, or diagnostics as ground truth.
- Publishing vendor runtime installs, session logs, private paths, or secrets.
- Replacing retired Python policy machinery with another language framework.

## Verification

The report names changed ownership boundaries, relevant hard checks and output,
semantic decisions made by review, preserved unrelated changes, and unresolved
uncertainty. Never claim a pass that was not observed.

## References

- `../../CONTRIBUTING.md`, current publication contract and tool classes.
- `../../docs/maintainer-tooling.md`, retained and retired script ownership.