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
- Inspect the current diff, nearby owners, callers, and any unmanaged changes.
- 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.
- Ground factual claims in current source, Git, host inventory, and runtime
output. Treat session history as evidence, never current truth.
- 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.
- Separate hard failures from judgment calls. Fix objective failures; explain
semantic tradeoffs with evidence instead of inventing a regex proxy.
- Preserve unrelated files and report the commands that actually ran and their
real results.
- 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.
1---2name: template-maintenance3description: Use when maintaining universal-template policy, prompts, skills, templates, MCP declarations, or publication tooling; review semantic coherence and select only relevant exact checks.4---56# Template Maintenance78## Core Principle910The model owns meaning. Deterministic checks own only facts that follow exactly11from source bytes, filesystem state, Git state, or runtime output.1213## When to Use / NOT1415- **Use when:** changing this template's canonical content or maintainer tools.16- **NOT when:** ordinary work in another repository, or when a direct project17 test already owns the acceptance boundary.1819## Workflow20211. Inspect the current diff, nearby owners, callers, and any unmanaged changes.222. Review changed policy, skills, prompts, and docs together for semantic23 consistency. Judge relevance, overlap, stale artifacts, prose quality, and24 whether each instruction still earns its place.253. Ground factual claims in current source, Git, host inventory, and runtime26 output. Treat session history as evidence, never current truth.274. Run only checks that prove an affected exact contract, such as strict YAML28 and metadata types, names, references, tracked-file ownership, disjoint29 hot/cold context and its budget, path containment, generated parity, secret30 patterns, or safe atomic filesystem mutation.315. Separate hard failures from judgment calls. Fix objective failures; explain32 semantic tradeoffs with evidence instead of inventing a regex proxy.336. Preserve unrelated files and report the commands that actually ran and their34 real results.357. Recommend deleting a check when it duplicates reliable model review and has36 no objective contract to prove.3738Stop when the affected content is coherent, relevant hard contracts pass, and39remaining uncertainty is reported.4041## Red Flags4243- Encoding nuanced policy or prose judgment as phrase tables.44- Running every historical script because an old checklist names it.45- Treating generated catalogs, projections, or diagnostics as ground truth.46- Publishing vendor runtime installs, session logs, private paths, or secrets.47- Replacing retired Python policy machinery with another language framework.4849## Verification5051The report names changed ownership boundaries, relevant hard checks and output,52semantic decisions made by review, preserved unrelated changes, and unresolved53uncertainty. Never claim a pass that was not observed.5455## References5657- `../../CONTRIBUTING.md`, current publication contract and tool classes.58- `../../docs/maintainer-tooling.md`, retained and retired script ownership.
ryan-brosas