Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Draft, confirm, and post replies to GitHub PR review threads. Handles per-category reply formatting, re-fetches thread resolution state so auto-resolved threads are skipped, and posts via GraphQL. Use when the user asks to "reply to PR threads", "post PR thread replies", or "draft PR reply messages".
.claude/skills/tobihagemann-reply-to-pr-threads/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-09 | ✗→✓ | ▲ Improved | 120% | 0% |
| case-19 | ✗→✓ | ▲ Improved | 17% | 0% |
| case-21 | ✗→✓ | ▲ Improved | -30% | 0% |
| case-05 | ✓→✗ | ▼ Worse | 136% | 0% |
| case-06 | ✓→✗ | ▼ Worse | 54% | 0% |
Draft replies for a processed review-thread list, confirm with the user, and post the surviving drafts.
Auto-detect owner, repo, and PR number from the current branch if not provided, then query the current resolution state:
bashgh api graphql -f query=' query($owner: String!, $repo: String!, $pr: Int!) { repository(owner: $owner, name: $repo) { pullRequest(number: $pr) { reviewThreads(first: 100) { nodes { id isResolved } } } } }' -f owner='{owner}' -f repo='{repo}' -F pr={pr_number}
Drop threads whose isResolved is now true. Reviewers or bots such as CodeRabbit may resolve threads after the original fetch, and drafting replies for them is wasted work.
$github-voice SkillRun the $github-voice skill to load voice rules and the insider-vs-outsider detection.
Use the processed-thread list from conversation context. Each entry has: thread id, file path, line, category (fix, skip, answer, or clarify), and per-category payload.
Keep every reply to one or two sentences. No bullet-point reasoning. No bolded labels.
fix: payload is a commit SHA, optionally with a divergence note.
Fixed in <commit-sha>.Only add a brief sentence after the SHA when the fix meaningfully diverges from what the reviewer suggested. Otherwise the SHA alone is enough.
skip: payload is the skip reasoning. State the reasoning directly. Do not apologize.
answer: payload is raw answer text from $answer-reviewer-questions. Tighten to one or two sentences and apply $github-voice rules. Do not cite transcripts or mention Claude. The reply reads as the implementer's own explanation.
clarify: payload is a user-directed question. Draft it as directed.
Output all drafts as text, grouped by file:
### <file-path>
**Thread <id>** (<category>, line <line>)
Reviewer: <original comment, truncated if long>
Reply: <drafted reply>Then use request_user_input to ask whether to post. Offer:
For each approved draft, write the drafted reply to .turbo/pr/thread-<thread-id>.md with apply_patch, then post via the reply mutation:
bashgh api graphql -f query=' mutation($threadId: ID!, $body: String!) { addPullRequestReviewThreadReply(input: {pullRequestReviewThreadId: $threadId, body: $body}) { comment { id } } }' -f threadId='<thread-id>' -F body=@.turbo/pr/thread-<thread-id>.md
Substitute <thread-id> with the thread's id for each post.
Report the outcome as a table, one row per thread, keeping every cell to a single line:
| Thread | Category | Outcome | |--------|----------|---------|
Where Outcome is one of:
Keep the report to the table. When the table would be empty, report one line stating that no replies were posted and why instead.
Then call update_plan to mark this step completed and continue with the next step of the active workflow.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 8,097 | 13,729 | +70% | 1 | 1 | 0% | 1,519 | 1,133 | -25% | 0 | 0 | — |
case-02 | fail→fail | 8,650 | 4,953 | -43% | 1 | 1 | 0% | 1,571 | 1,092 | -30% | 0 | 0 | — |
case-03 | fail→fail | 10,549 | 22,436 | +113% | 1 | 1 | 0% | 1,855 | 5,553 | +199% | 0 | 0 | — |
case-04 | fail→fail | 6,912 | 6,006 | -13% | 1 | 1 | 0% | 1,371 | 1,264 | -8% | 0 | 0 | — |
case-05 | pass→fail | 3,105 | 5,784 | +86% | 1 | 1 | 0% | 564 | 1,331 | +136% | 0 | 0 | — |
case-06 | pass→fail | 4,189 | 5,019 | +20% | 1 | 1 | 0% | 767 | 1,181 | +54% | 0 | 0 | — |
case-07 | pass→fail | 8,034 | 5,442 | -32% | 1 | 1 | 0% | 1,669 | 1,251 | -25% | 0 | 0 | — |
case-08 | pass→pass | 8,634 | 3,596 | -58% | 1 | 1 | 0% | 1,652 | 1,613 | -2% | 0 | 0 | — |
case-09 | fail→pass | 3,674 | 6,312 | +72% | 1 | 1 | 0% | 688 | 1,511 | +120% | 0 | 0 | — |
case-10 | fail→fail | 6,388 | 23,285 | +265% | 1 | 1 | 0% | 1,232 | 1,229 | -0% | 0 | 0 | — |
case-11 | pass→fail | 4,487 | 5,567 | +24% | 1 | 1 | 0% | 851 | 1,172 | +38% | 0 | 0 | — |
case-12 | pass→fail | 5,640 | 4,479 | -21% | 1 | 1 | 0% | 1,054 | 1,229 | +17% | 0 | 0 | — |
case-13 | pass→fail | 9,353 | 6,080 | -35% | 1 | 1 | 0% | 1,593 | 1,203 | -24% | 0 | 0 | — |
case-14 | fail→fail | 4,293 | 5,704 | +33% | 1 | 1 | 0% | 866 | 1,254 | +45% | 0 | 0 | — |
case-15 | fail→fail | 15,914 | 5,672 | -64% | 1 | 1 | 0% | 3,312 | 1,179 | -64% | 0 | 0 | — |
case-16 | fail→fail | 16,332 | 13,867 | -15% | 1 | 1 | 0% | 2,727 | 1,738 | -36% | 0 | 0 | — |
case-17 | fail→fail | 11,072 | 6,716 | -39% | 1 | 1 | 0% | 713 | 1,282 | +80% | 0 | 0 | — |
case-18 | fail→fail | 2,424 | 17,761 | +633% | 1 | 1 | 0% | 378 | 2,463 | +552% | 0 | 0 | — |
case-19 | fail→pass | 6,689 | 11,906 | +78% | 1 | 1 | 0% | 1,166 | 1,369 | +17% | 0 | 0 | — |
case-20 | fail→fail | 5,926 | 2,052 | -65% | 1 | 1 | 0% | 882 | 1,178 | +34% | 0 | 0 | — |
case-21 | fail→pass | 11,286 | 3,201 | -72% | 1 | 1 | 0% | 2,143 | 1,496 | -30% | 0 | 0 | — |
case-22 | fail→fail | 8,052 | 4,926 | -39% | 1 | 1 | 0% | 1,339 | 1,186 | -11% | 0 | 0 | — |
case-23 | fail→fail | 5,491 | 5,207 | -5% | 1 | 1 | 0% | 903 | 1,112 | +23% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted, and 5 counted toward the lift figure. The other 18 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of -13 percentage points is the difference between those two pass rates over the 5 comparable cases. 10 cases got worse with the skill loaded, and they are included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
| Model | Method | Date | Lift |
|---|---|---|---|
| gemini-3.6-flash | verified | 8/21/2026 | 0% |
Other measured skills in the registry, with their headline benchmark lift.