/pr — Create or Update a Pull Request
Description
Creates a PR with a structured description that lets a reader understand the change without reading the diff. Can also update an existing PR's description.
The canonical example for this project is PR #364 (PlonkStorage slice 5 — PlonkNats CRD + conditions-based reconciler/worker). When in doubt about structure, tone, or how to weave code blocks and mermaid diagrams into the body, open that PR and match it.
Instructions
When the user runs /pr or /pr <pr-number>, follow this workflow:
Phase 1: Gather context
-
Determine mode:
/prwith no number: create a new PR/pr <number>: update the description of an existing PR
-
Read the branch state:
git log --oneline origin/main..HEAD git diff --stat origin/main..HEAD -
Read the full diff to understand all changes:
git diff origin/main..HEAD -
Check for linked issues in commit messages (look for
#<number>,Part of #<number>,Closes #<number>).
Phase 2: Draft the description
The description has a fixed skeleton plus change-shape sections that you include only when the change actually has them. The skeleton always present:
Summary
2–4 sentences. State what the PR does and the headline test signal (e.g. make e2e-full → 39/39 green in 9m). Link the epic it's part of. If it lands a methodology shift (new architectural pattern, refactor target), name it here and link the doc that codifies it.
Test plan
Markdown checklist. Mark completed items with [x]. List every verification command actually run, with results:
cargo fmt --allcargo clippy --workspace --all-targets --all-features -- -D warningscargo test -p <package>with pass countmake e2e-fullwith pass count + wall-clock when relevantmake check-docswhen docs were touched- Manual verifications and the commands that produced them
Change-shape sections (include the ones that apply, in this order):
Wire-format diff
Whenever the PR changes a CRD spec/status, a FlatBuffer message, an env var contract, or any other persisted interface, show the before → after as YAML/Rust/proto snippets in a fenced code block. The reader should see the shape change without inferring it from the diff.
Architecture
Explain how the new pieces fit together. For reconciler/worker work this means: which fieldManager owns which condition, how the watch fires, what the worker's exact responsibility is. Include a mermaid sequence diagram when the flow crosses ≥3 actors (CR → reconciler → NATS → worker → apiserver → kube-controller). Validate every mermaid block with make check-docs (or mmdc) before pushing — broken diagrams render as raw text on GitHub.
When this section is non-trivial, put the full version in docs/<feature>.md and link it from the PR; the PR section is the precis.
CLI flow
If plonk install / plonk uninstall ordering changed, list the new steps with the prior order alongside (numbered list or table). Call out steps that exist purely to dodge a chicken-and-egg (e.g. bootstrapping NATS with the operator's fieldManager so the first operator reconcile is a no-op SSA).
Storage layout / mount paths
If persistent storage shape changed (new PVC, new mount path, new tier consumer), state the path convention applied (<tier-base>/<workload>/<n> for StatefulSets, <tier-base>/<workload> for Deployments) and how the ordinal subdir is realised (downward API → shell wrapper).
HA semantics
For workloads that scale: explain the replicas: 1 behaviour and the replicas: N behaviour. State the failure mode tolerated (e.g. "JetStream streams configured replicas: 3 tolerate single-node failure"). Name any production-only knobs the change unlocks.
What this slice does NOT ship
Bullet list of related work explicitly out of scope, each pointing at the issue/epic that tracks it. The point is to head off "why didn't you also fix X?" review comments by showing that X is known and queued.
Spec-change re-dispatch (reconciler PRs only)
When the reconciler decides whether to re-publish a worker task, name the helper (e.g. dispatched_for_generation) and the comparison it does (Dispatched.observedGeneration vs metadata.generation). Link the unit test that pins boundary behaviour.
Phase 3: Create or update the PR
For new PRs:
Write the body to a tempfile and use --body-file. Heredoc + inline backticks in the body breaks GitHub's markdown rendering — backticks inside heredocs get escaped or mangled. Tempfile + --body-file is the only reliable form.
BODY_FILE=$(mktemp)
cat > "$BODY_FILE" <<'EOF'
## Summary
...
EOF
git push -u origin <branch-name>
gh pr create --draft --title "<title>" --body-file "$BODY_FILE"
rm "$BODY_FILE"
PRs are always created as draft ([[feedback_draft_prs]]). The user marks ready for review; never do it automatically.
For existing PRs:
BODY_FILE=$(mktemp)
cat > "$BODY_FILE" <<'EOF'
...
EOF
gh pr edit <number> --title "<title>" --body-file "$BODY_FILE"
rm "$BODY_FILE"
Rules
- Title: Under 70 characters. Lead with a verb or the epic/slice name (
PlonkStorage slice 5: ...). No period. - Body: Use a tempfile +
--body-file, never inline heredoc with backticks. - Tone: Concrete and specific. Code blocks beat prose summaries — show the actual YAML/Rust diff instead of describing it ([[feedback_human_communication]]).
- Mermaid diagrams: Validate with
make check-docsbefore pushing. Avoid semicolons insideNote overtext — the parser breaks. - No AI references: Never mention AI tools, Claude, or add Co-Authored-By tags.
- No fabricated estimates: Don't add "rollout takes ~2h" or similar timings unless they were actually measured ([[feedback_no_fabricated_estimates]]).
- No forward-looking phrasing in inline comments the PR introduces: keep PR-anchored TODOs out of source per the CLAUDE.md comment-hygiene rules.
- Always show the user the drafted title and description before creating/updating.
- Return the PR URL when done.
Expert Next.js App Router
Developpement
Un skill qui transforme Claude en expert Next.js App Router.
Générateur de README
Developpement
Crée des README.md professionnels et complets pour vos projets.
Rédacteur de Documentation API
Developpement
Génère de la documentation API complète au format OpenAPI/Swagger.