Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Nested lingtai-dev-guide reference for getting a PR review-ready: PR readiness gates, multi-model/daemon/Claude read-only review passes, self-contained local HTML explainers, PR body hygiene and gh pr edit troubleshooting, source-labeled deliverables with syntax/validation checks, and maintainer authorization boundaries for opening, editing, and merging PRs.
.claude/skills/lingtai-ai-dev-guide-pr-review-deliverables/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-09 | ✗→✓ | ▲ Improved | 65% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 57% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 22% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 95% | 0% |
| case-10 | ✗→✓ | ▲ Improved | 57% | 0% |
Nested lingtai-dev-guide reference. Read this after the top-level router sends you here to prepare a PR for review, build its human-facing explainer, and run review gates before asking for a merge.
It consolidates the PR-review-gate, HTML-explainer, deliverable-template, and release-style-audit practices, covering preparing and reviewing PRs and their human deliverables. Release publication is outside this bundled guide — use the dedicated release owner available in the maintainer's environment.
A PR is review-ready only when (1) the diff is clean and validated, (2) a self-contained HTML explainer tells the human the story, and (3) the change has passed at least one independent review pass. Opening, editing, and merging PRs are external side effects: do them only after an explicit maintainer authorization for that side effect.
git diff --check (no trailing whitespace / conflict markers).root router catalog/routing table were updated in the same commit.
ANATOMY.md is updated in the same commit.
Known repo pitfalls: docs/stars/stars.csv may use CRLF intentionally — do not "normalize" it; and portal tests may need frontend assets built first (npm --prefix portal/web run build).
Run at least one independent, read-only review pass before merge. Strongest first:
daemon (Claude Code / Codex) or a second model review correctness, scope, and regressions. Save the review as an artifact so it is auditable. Run API-backed reviewers with keys from the environment (<API_KEY_ENV>), never inline.
correctness bugs and reuse/simplification cleanups.
Treat each review as input, not a verdict: read the diff yourself (it is ground truth, not the daemon's summary), classify each finding (confirmed defect, edge risk, future refactor, doc/process note), and fix the smallest thing that resolves a real defect. Do not expand into a broad refactor because nearby code is related.
These reviews are gates, not authorization to merge — see §5.
Every non-trivial PR ships one self-contained .html explainer for the human. This is the canonical statement of the requirement; other dev-guide references point here rather than restating it.
Local-only by default. Routine PR explainers are local output, not repo artifacts. Write to an ignored location — artifacts/pr<NUMBER>-<slug>-explainer.html, reports/pr<NUMBER>-<slug>-explainer.html, tmp/<topic>-<date>.html, or the agent/worktree report workspace — and hand the human the absolute file://-openable path in the short pointer message. Do not commit routine explainers. Commit one only when the human explicitly asks, when it is a release/long-term reference artifact, or when repo documentation links to it deliberately; use git add -f for that exception and explain the reason in the PR.
Naming and timing. Name it pr<NUMBER>-<slug>-explainer.html once the PR exists; pre-push use <topic>-<date>.html and rename locally after. Write it before asking for review or merge, and refresh it when blockers or fixes materially change the story.
Required sections, in order: TL;DR / conclusion-first, baseline, what-was-done (with diff snippets), validation (commands + results), risks/decisions, next steps, and a source index. Plain text/Markdown is reserved for the short pointer message and conversational replies.
The only exception is a strictly one-line docs/chore PR where the human has explicitly waived the report. Absent that waiver, write the HTML even for a small fix; if a change is too small for a useful explainer, that is a signal to bundle it with related work.
source (file path, command, commit SHA, PR/issue number).
(<your-lingtai-checkout>); redact tokens/keys/chat IDs as <REDACTED>.
step. Sanity-check the markup before handing it over so a malformed tag does not silently swallow content. For the HTML skeleton, MathJax setup, and the full validation checklist, read ~/.lingtai-tui/utilities/swiss-knife/reference/html-report/SKILL.md (or load the swiss-knife skill and route to html-report) — it owns those mechanics. The PR-specific path, naming, and section rules above still govern.
gh pr editThe PR body cites the issue, summarizes the change, and lists validation steps, mirroring the explainer's TL;DR and validation without duplicating the whole HTML. Keep body and explainer in sync when fixes land.
gh pr edit can emit GraphQL deprecation warnings on some gh versions — those are warnings, not failures. If gh pr edit --body-file is noisy, re-run or set the body via the web UI; treat it as a version-dependent troubleshooting note, not a blocker.
Do not, without an explicit imperative authorization for that specific side effect:
Generating diffs, running read-only reviews, and writing local explainers are always allowed. When approval wording is ambiguous, show the maintainer the final candidate refs/diff and ask before mutating GitHub state. For consequential releases (tagging, publishing to PyPI/Homebrew, website release archive/blog), use the maintainer environment's dedicated release owner; this reference grants no publication authority.
reference/contributing/SKILL.md — issue → worktree → PR → merge loop anddaemon decomposition.
reference/skill-stewardship/SKILL.md — PR-ready cleanup when the change addsor edits a skill.
skills-manual — generic skill authoring (when the deliverable is a skill).Other measured skills in the registry, with their headline benchmark lift.