[martin-conur]

Summary

Every worker now opens its pull request with a short ## Summary written for you, not for the reviewer: two or three plain sentences on what changed and why. That gives each PR a description you can read once its ticket number stops meaning anything, and it is what the task monitor will show for finished work. Reviewers flag a PR without one as a minor nit.

Closes #274

What changed

• Workers (all 8 prompts: claude gh/jira/notion/local, the dogfood copy, kiro gh/local/notion): step 8 now runs gh pr create … --fill --body-file - <<'BODY' with an example body that opens with ## Summary, then the tracker's own Closes line (Closes #<N>, the jira key, the notion page URL, or the local task file). --fill still sets the title. Workers are told to write the body flush left, since GitHub renders an indented ## Summary as code.
• Reviewers (all 6 prompts): a missing ## Summary is a nit that never blocks on its own.
• Docs: a ### PR Body section in .claude/gh-workflow.md and all seven steering templates; the README's PR step shows the new command.
• CHANGELOG: entry under Unreleased with Upgrading: re-run task-init <loadout>.

Not changed

task-done's printed gh pr create … --fill hint for humans is untouched. The spec covers the worker prompts only.

Tests

tests/worker_checklist.bats gains 5 tests: on every worker prompt, the example body's first line is ## Summary and the loadout's Closes line follows it; no prompt keeps a bare --fill; the rule is worded the same on every claude loadout; all reviewers call it a nit; every workflow doc and the README carry the section. I broke each one's target in turn, and each test failed. Full suite: 1388/1388 pass. tools/check-drift.sh is clean. No shell files changed.

🤖 Generated with Claude Code

[martin-conur]

Spec compliance (#274)

• Worker prompts on every loadout: delivered. All 8 (claude gh/jira/notion/local, the dogfood .claude/ copy, kiro gh/local/notion) now run gh pr create … --fill --body-file - <<'BODY'. The example body opens with ## Summary and then carries the tracker's own Closes line. The rule wording says "written for the user rather than a reviewer", as the spec asked.
• Reviewer prompts: delivered. All 6 (4 claude loadouts, the dogfood copy, kiro-gh) treat a missing ## Summary as a nit that never blocks on its own. kiro-local and kiro-notion have no reviewer agent, so nothing is missing there.
• Docs: delivered. The ### PR Body section is in .claude/gh-workflow.md and all 7 steering templates, and the README's PR step is updated.
• CHANGELOG: delivered, with the "Upgrading: re-run task-init <loadout>" note.
• Parity and doc suites pin it on every loadout: delivered. The 5 new bats tests cover the first line of the example body, the per-tracker Closes line, no bare --fill, byte-identical rule wording across the claude loadouts, the reviewer nit, and the docs and README.
• CI: a run exists for head dc20ac3, and bats (macOS and ubuntu), drift and ShellCheck all pass.
• PR body: opens with ## Summary, so no nit there.

Code-review findings

No correctness, security or edge-case issues found.

• Combining --fill with --body-file - is sound: --fill supplies the title and the explicit body wins. The old "opens an editor" line is gone, which fits a non-interactive worker.
• The quoted <<'BODY' heredoc matches the repo's no-backtick-substitution rule from reviewer PR comments are shell-interpolated: backticked code is command-substituted, and substituted-to-empty findings lose words silently #234.
• The worker is told to write the body and its terminator flush left. That is the right call, since the example sits indented inside a list and a quoted heredoc needs a flush-left terminator.
• The example Closes lines per tracker are correct. task-done's printed gh pr create … --fill hint is left untouched, as the PR body says, and the spec only covers worker prompts.

Verdict

clean — nothing to fix in this PR. Merge decision is PM's.