maintain-verification-skill
Use only when the user explicitly invokes maintain-verification-skill to update an existing project verification skill and feature map against source and observed behavior.
SKILL.mdskills/local/maintain-verification-skill/SKILL.md
GitHubMaintain a Verification Skill
Run only when explicitly invoked, for example $maintain-verification-skill in
Codex or /maintain-verification-skill in Claude Code. A generic verification
request or a reference from another skill does not activate this workflow.
Keep an existing verification procedure usable as the application changes. Follow the same procedure on either agent; use the tools available in that session without requiring a particular model, subagent type, or CLI.
Establish the target
Locate the project-owned verification skill and its feature map. Check both
.agents/skills/verify-*/ and .claude/skills/verify-*/, resolving links to
identify the canonical source. Follow the project's established layout when
different. If several targets fit and the request does not identify one, ask
which to maintain. If none exists, report that prerequisite and recommend
create-verification-skill; do not invoke it automatically.
Inspect repository instructions, working changes, execution wrappers, and the verification skill's launch, readiness, driving, and cleanup procedures. Resolve links before writing: a project skill may point into the user's global skill view. Update only a project-owned target, preserving unrelated changes.
The editable scope is the verification skill's own instructions, feature map, and helpers. Product code, dependencies, shared configuration, and other skills require separate authorization. Treat maintained procedures as durable project operational documentation; keep run evidence in ignored temporary storage.
Reconcile the map with the source
Read the index and each feature's instructions. Check for missing, duplicate, obsolete, and unlinked entries. Trace each feature's user-facing entry point in the current source and note the behavior, prerequisites, and one observable verification recipe. Add a missing feature only with a concrete source path and a discoverable user contract.
Use the verification contract to resolve gaps in readiness, evidence, or cleanup. Source inspection can identify likely drift; it does not establish that a procedure works.
Work directly for a small map. Delegate bounded, read-only source inspection only when authorized and supported; the coordinator owns edits and live driving. Independent app instances must not compete for the same port or data directory.
Exercise the procedures
Run every mapped feature at least once using the documented execution boundary. Combine compatible recipes into a small number of app states. For a requested subset, name that subset and leave the rest explicitly unverified.
- Establish the artifact/revision, readiness, and ownership before driving. After an unexpected failure, restore a known state or recheck readiness before proceeding. A healthy process alone does not establish usable UI state.
- Use an owned long-lived instance for a service when appropriate, or a fresh isolated invocation for a short-lived CLI. Respect shared instances and data.
- Capture the action, expected and actual result, and retained evidence path for each feature. Record missing access or configuration and the attempted route; do not invent access, install dependencies, or claim the feature passed.
- Clean up owned processes and temporary state after failures as well as success. Confirm that logs, screenshots, and other evidence survive cleanup.
Classify and correct
Separate three findings:
- Procedure drift: the documented command, selector, prerequisite, or map is outdated. Correct it within the target skill and run the corrected path again.
- Helper gap: the application behaves as intended but the skill's helper cannot drive it. Fix that helper within scope and rerun the affected path.
- Product regression or unclear contract: implementation disagrees with the established behavior, or the intended behavior cannot be established. Report evidence; do not weaken an assertion, delete coverage, or rewrite the expected result simply to make the procedure pass. Do not repair product code here.
An unexpected result alone is not proof of an intentional behavior change. Ground changed expectations in an explicit requirement, specification, or recorded decision. Preserve unresolved coverage in the map.
Use prove-it-works before reporting. Review the diff and report one outcome:
- clean: all mapped features ran successfully and no correction was needed.
- changed: the scoped corrections were exercised and all mapped features ran successfully.
- blocked: any mapped feature remains unverified, inaccessible, failing, or ambiguous. List successful paths, any verified corrections, and the concrete blocker separately. A subset check is not a clean result for the whole map.
Include the target, revision, changed files, per-feature results, and evidence locations. Commit, publication, and review requests follow the user's existing authorization; this skill does not initiate them automatically.
Source
Adapted from Lauren Tan's pstack maintain-verification-skill. Local changes: shared Codex/Claude instructions, explicit invocation, canonical project ownership, bounded optional delegation, and incomplete-coverage outcomes. See LICENSE.