Quality Validation

Compare deliverable against intake. The user asked for X — did you deliver X?

This skill applies after task execution, before reporting completion. Partial completion is not completion.

Validation by Task Type

Code Changes

<code-validation>

If tests exist and you can run them:

1. Run the test suite (or relevant subset)
2. Verify no regressions introduced
3. Confirm new behavior is covered by tests

If no tests or cannot run:

1. Re-read the changed code
2. Trace the logic path for the requested change
3. Identify edge cases — are they handled?
4. Check error conditions — what happens when things fail?

Always check:

• Does the change do what was asked? (not more, not less)
• Are there obvious bugs visible in the diff?
• Does it follow existing patterns in the codebase?

</code-validation>

Documentation / Design Tasks

<doc-validation>

1. Re-read the original request — what did the user actually ask for?
2. Compare against deliverable:
   • Are all requested sections present?
   • Does the structure match what was asked?
   • Is the level of detail appropriate?
3. Check logical consistency:
   • Do the parts fit together?
   • Are there contradictions?
   • Do conclusions follow from premises?
4. Check completeness:
   • Are there gaps or "TODO" placeholders?
   • Would someone else understand this without follow-up questions?

</doc-validation>

Research / Analysis Tasks

<research-validation>

• Question coverage — did you answer what was asked?
• Evidence basis — are claims supported by what you found?
• Gaps acknowledged — did you note what you couldn't find or verify?
• Actionable output — can the user do something with this information?

</research-validation>

Refactoring / Cleanup Tasks

<refactor-validation>

• Behavior preservation — does it still do what it did before?
• Scope adherence — did you change only what was requested?
• No regressions — run tests if available
• Improvement achieved — is the stated goal (readability, performance, etc.) actually better?

</refactor-validation>

Common Validation Failures

<failures>

Solved a different problem: User asked for X, you delivered Y because Y seemed better. Fix: deliver X first, then suggest Y as improvement.

Incomplete execution: Started the task but didn't finish all parts. Fix: re-read request, check each requirement.

Untested assumptions: "This should work" without verification. Fix: actually test, run, or trace through the change.

Scope creep unacknowledged: Did extra work without noting it. Fix: clearly state what was requested vs. what was added.

</failures>

When Validation Reveals Problems

1. Don't report as complete — partial completion is not completion
2. Fix if possible — address the gap before reporting
3. Report honestly if blocked — explain what's done, what isn't, and why

Pre-Completion Checklist

Before reporting task complete:

• Re-read the original request
• Identify what "done" means for this specific task
• Compare deliverable against that definition
• For code: run tests or manually trace logic
• For docs: check logical consistency and completeness
• Note any deviations or partial completions honestly