feat(writing-plans): add Evaluator section + self-review check

[suresh2824]

Summary

Adds a top-level ## Evaluator section to the writing-plans skill so every plan declares the deterministic check that gates moving the plan from in-progress to complete. Without one, "done" defaults to vibes — a maintainer reads the implementation, decides it looks right, moves the plan to completed/. Weeks later a regression surfaces with no record of what "passing" was supposed to mean.

This formalizes the Planner → Generator → Evaluator topology that the rest of the superpowers skillset already assumes: the writing-plans Planner produces the spec, executing-plans / subagent-driven-development Generators execute tasks, and the Evaluator returns pass/fail before lifecycle transition.

What's added

The new ## Evaluator section sits between ## Remember and ## Self-Review, enumerating acceptable Evaluator forms so plans in any tech stack can declare a fit-for-purpose check:

• Test command (pytest path::test_function, npx playwright test, cargo test <name>, go test ./pkg/...) that exits 0
• HTTP assertion (curl -fsS <url> with expected status + body shape)
• Database query asserting expected state
• verify-* skill invocation (project-local verification)
• Rubric pass-list for AI-system or design plans where a deterministic check isn't possible

Also includes a quadruple-backtick example to demonstrate the placement rule (## Evaluator heading above the terminal ## Rollback/sign-off, grep-discoverable across the plan corpus), and adds Self-Review item #4: Evaluator declared so authors verify the section is present before considering the plan finished.

Why submit this

The pattern has been running in a production codebase (mw-vastra retail-ops platform, ~7 weeks) where every non-trivial plan declares an Evaluator. It catches the silent-completion failure mode at the plan-folder lifecycle boundary (partial/ → completed/), which is exactly the place where "looks done" diverges from "is done." Promoting it upstream means every Claude Code user gets the gate without per-project overlay.

Test plan

• head skills/writing-plans/SKILL.md — frontmatter intact
• Section order verified — new ## Evaluator between ## Remember and ## Self-Review
• Markdown fences correct — quadruple-backtick outer + triple-backtick inner for the example (no nested-fence rendering bugs)
• No breaking changes — additive only; existing plan-writing workflow unchanged

Source

Pattern originally formalized in mw-vastra PR #488 (project-level rule), promoted upstream here so it's not Vastra-specific. Inspired by Anthropic's agentic-AI design guidance on full-stack Planner → Generator → Evaluator topologies.

🤖 Generated with Claude Code

[YOMXXX]

Review/triage note:

This is a behavior-shaping change to skills/writing-plans/SKILL.md, not just documentation. Based on recent maintainer feedback on similar skill-content PRs, I think this needs more evidence before line-by-line review is useful.

Specific gaps I would expect reviewers to ask about:

• The PR targets main; current contribution flow appears to expect dev.
• The PR body does not use the current PR template sections.
• The problem statement is general and project-derived. It does not include a concrete Superpowers session where a generated plan lacked a completion evaluator and that caused a real bad outcome.
• There is no RED/GREEN behavioral eval showing current writing-plans omits this gate on representative prompts, and that the new wording reliably adds useful evaluator sections without bloating plans or conflicting with the existing per-task verification steps.
• The new section adds a mandatory lifecycle rule (may NOT move a plan to completed) but this repo's active workflow docs do not currently define that partial/ -> completed/ lifecycle as a core Superpowers behavior, so this may be importing a project-local convention into core.

The idea may be worth discussing, but I would treat it as a design/eval proposal first, not a ready-to-merge skill prose change.