chunk-pr

name	chunk-pr
description	>-

Chunk PR

Break one big PR into a sequence of small, logical, merge-safe PRs. Suggest, don't decide — the user approves every chunk and every destructive action.

Preferences

Read ~/.claude/skills/chunk-pr/preferences.md using the Read tool. If not found, no preferences are set.

Context

On startup, use Bash to detect: current branch (git branch --show-current), base upstream, commits ahead, total lines changed (git diff {base}...HEAD --shortstat), and repo (gh repo view --json nameWithOwner -q .nameWithOwner). Skip any that fail. Do NOT run destructive commands at this stage.

Command routing

Check $ARGUMENTS:

• help → display help then stop
• config → interactive setup then stop
• reset → delete ~/.claude/skills/chunk-pr/preferences.md, confirm, stop
• anything else (including empty) → run the chunking workflow

Flags (parsed from $ARGUMENTS)

• --base <branch> — override target base (default: auto-detect or preference)
• --max-lines <N> — target max lines per chunk (default: 400 or preference)
• --strategy <kind> — conservative (more/smaller), balanced (default), aggressive (fewer/larger)
• --dry-run / --plan-only — stop after presenting the plan; don't create branches or PRs
• --draft — open chunks as draft PRs (default)
• --ready — open chunks as ready-for-review (overrides --draft)
• --no-push — create local branches only, skip push/PR creation
• --no-linear — skip Linear linking
• Positional — PR number (e.g. 234), PR URL, branch name, or commit range (main..HEAD). If omitted, use current branch.

Help

Chunk PR — Split a big PR into a sequence of smaller, merge-safe ones

Usage:
  /chunk-pr                                Analyze current branch
  /chunk-pr <pr-number>                    Analyze a specific PR
  /chunk-pr <pr-url>                       Analyze a PR by URL (same or other repo)
  /chunk-pr <branch>                       Analyze a local branch
  /chunk-pr main..HEAD                     Analyze a commit range
  /chunk-pr --max-lines 300                Prefer chunks under 300 lines
  /chunk-pr --strategy conservative        More, smaller chunks
  /chunk-pr --dry-run                      Plan only — don't touch git
  /chunk-pr config                         Set preferences
  /chunk-pr reset                          Clear preferences
  /chunk-pr help                           This help

What it does:
  1. Identify the subject PR / branch / range
  2. Classify commits and files into logical groups
  3. Propose an ordered, dependency-aware split plan
  4. Let you approve, reorder, merge, or re-split chunks
  5. On approval, create branches + draft PRs + Linear links
  6. Leave the parent branch untouched

Current preferences:
  (shown above under Preferences)

Config

Use AskUserQuestion:

Q1 — "Target max lines per chunk?" (200, 400 (default), 600, 1000, no limit)

Q2 — "Default split strategy?" (conservative, balanced (default), aggressive)

Q3 — "Default action after plan approval?" (per-chunk confirm (default), execute all, branches only no PRs, plan only)

Q4 — "Branch naming convention for chunks?"

• {parent}/chunk-{N}-{slug} (default — e.g. user/ais-810-refactor/chunk-1-schema)
• {parent}-{N} (short — e.g. user/ais-810-refactor-1)
• chunk/{slug} (flat — e.g. chunk/schema)
• custom (user provides pattern with {parent}, {N}, {slug} placeholders)

Q5 — "Link chunks to parent Linear issue?" (Yes as sub-items (default), Yes as relates-to, No, Ask each time)

Q6 — "Default PR type for chunks?" (draft (default), ready for review)

Q7 — "Classification weighting?" (by file type (default), by commit topic, hybrid)

Save to ~/.claude/skills/chunk-pr/preferences.md.

Reset

Delete ~/.claude/skills/chunk-pr/preferences.md and confirm: "Preferences cleared. Using defaults."

First-time detection

If no preferences file exists, show: "First time using /chunk-pr? Run /chunk-pr config to set defaults, or just continue — I'll pick sensible defaults."

Then proceed.

Steps

1. Identify the subject

Resolve in order:

1. $1 looks like a PR URL (https://github.com/*/pull/N) → use gh pr view <url> --json ... or WebFetch for external repos.
2. $1 is a number → gh pr view <N> on current repo.
3. $1 is a commit range (contains ..) → use it directly with git log/git diff.
4. $1 is a branch name → treat as {base}..{$1}.
5. No $1 → use current branch vs. its upstream/default-base.

Early bail-outs:

• If total changed lines < max-lines preference × 1.5 → tell the user this probably doesn't need chunking, ask if they still want to proceed.
• If there's 1 commit touching 1 file → refuse politely; this isn't chunkable.
• If uncommitted changes exist on current branch → stop and ask (per global git-safety rule; never proceed silently).

2. Gather context

From git / gh:

• All commits in range: git log {base}..{head} --pretty=format:"%H|%s|%an|%ai" --numstat
• Per-file stats: git diff {base}...{head} --numstat and --stat
• Diff body (for classification, not pasted back): git diff {base}...{head} — read in chunks if huge
• PR metadata if applicable: title, body, labels, reviewers
• Existing review comments if PR is open: gh api repos/{owner}/{repo}/pulls/{N}/comments

From Linear (unless --no-linear):

• Extract issue ID from branch name or PR body
• get_issue for title, description, project, sub-issues, related issues

Base branch detection (same as /create-pr):

1. --base flag → preference → upstream tracking → repo default branch

3. Classify commits and files

File-based classification (default weighting):

Group	Patterns (examples)
schema	**/migrations/**, **/schemas/**, drizzle/**, prisma/**, *.sql
api	src/router/**, src/api/**, src/handlers/**, src/server/**
ui	src/components/**, src/app/**, **/*.tsx, **/*.css
tests	**/__tests__/**, **/*.test.*, **/*.spec.*, e2e/**
docs	**/*.md, docs/**, README*, CHANGELOG*
config	package.json, tsconfig*.json, .env.example, vite.config.*, next.config.*, CI workflows
deps	pnpm-lock.yaml, package-lock.json, yarn.lock — bundle with config
refactor	commits whose message starts with refactor:/chore: AND whose diff is behavior-preserving (renames, moves, type-only)
other	anything unmatched

Commit-based signals (boost classification):

• Conventional prefix (feat:/fix:/refactor:/chore:/test:/docs:)
• Scope in parens (feat(meetings): ...) — use scope as a chunk hint

Cross-cutting flag — if a single file appears in multiple groups' commits, flag it. These are the hardest to split cleanly and usually belong in the earliest chunk.

Classification presets:

• by file type (default) — group by the table above
• by commit topic — group by conventional scope (meetings, knowledge, auth)
• hybrid — topic first, then file-type within a topic

Read the diff itself (via Read/Grep where needed) when the filename isn't enough to classify — e.g. a .ts file could be API or UI.

4. Propose the split plan

Build an ordered list. Ordering rules:

1. config/deps first if they unblock later chunks
2. schema before api before ui
3. refactor before feature commits that depend on the refactor
4. tests bundle with the behavior they cover (don't create a tests-only chunk unless the tests pre-existed and were just moved)
5. docs last

For each chunk, produce:

Chunk {N}: {title with conventional prefix}
  Scope:        {1-2 sentences}
  Commits:      {N} ({oldest-SHA}..{newest-SHA})
  Files:        {N} files, +{added}/-{removed} lines
  Depends on:   Chunk {N-1} or "none"
  Safe to ship: yes / yes-behind-flag / only-after-{N}
  Rationale:    {why this grouping}
  Risks:        {conflicts, cross-cutting files, behavior change}

Size policing — if a chunk exceeds max-lines:

• Try splitting by file-subgroup (UI → forms vs. views)
• Or split the chunk's commits into "refactor" + "feature" halves
• If neither helps, flag the chunk as "oversized — consider landing behind a flag"

Conflict forecasting — if two chunks both touch the same file at overlapping lines, call it out; the later chunk will need a rebase after the earlier one lands.

Present the full plan as a compact, numbered list the user can react to.

5. Review with the user

Use AskUserQuestion to get the call. Options:

• Approve plan as-is — proceed to execute (step 6)
• Merge chunks {A} and {B} — combine two chunks into one, re-present plan
• Split chunk {N} further — break one chunk into smaller ones, re-present plan
• Reorder — swap or reshuffle; re-present plan
• Drop chunk {N} — leave those commits on the parent branch, don't chunk them
• Change strategy — re-run classification with aggressive / conservative
• Dry-run stop — save the plan to a file, don't touch git

Loop until the user approves or cancels. Keep iterations cheap — don't re-fetch git/Linear state between iterations unless the user changes scope.

Optional: if Chrome MCP is set up and the user asks, open the original PR in a tab via tabs_create_mcp for visual cross-reference while reviewing the plan. Do not open Chrome by default — it's noisy.

6. Execute on approval

Pre-flight (non-negotiable):

• git status must be clean — else stop and ask (per global rule)
• Remember the current HEAD; print it so the user can roll back
• Never modify the parent branch or its ref. Only create new branches.

Per chunk, with per-chunk confirmation (unless preference says "execute all"):

1. Check out base: git checkout {base} && git pull --ff-only origin {base}
2. Create chunk branch: git checkout -b {chunk-branch-name} (using preferred naming convention)
3. Apply changes. Two modes — pick the one that fits the classification:
   • Commit cherry-pick (preferred when the chunk maps to contiguous commits): git cherry-pick {sha1} {sha2} ... Resolve conflicts if they appear — ask the user, never force.
   • File subset (when commits are cross-cutting and only a subset of the diff belongs): For each file: git checkout {head} -- {path} (or for partial-file: git checkout -p {head} -- {path}) Then a single git commit -m "{chunk-title}" with an accurate, non-generic message.
4. Run local sanity — read-only: git diff {base}...HEAD --shortstat, git log {base}..HEAD --oneline. Never run pnpm build or pnpm dev (global rule).
5. Push (unless --no-push): git push -u origin {chunk-branch}
6. Open draft PR (unless --no-push or --no-pr): gh pr create --draft --base {base} --title "{title}" --body "{body}" PR body includes:
   • 1-2 line scope
   • Part {N}/{total} of splitting #{original-pr} (or the parent branch)
   • Depends on: #{previous-chunk-pr-number} if applicable
   • Linear link if available
   • Test steps (brief, per-chunk)
7. Link to Linear:
   • Attach PR URL to parent issue via update_issue (links array)
   • Or post a create_comment on the parent issue listing all chunk PRs once execution finishes

Between chunks, return to the base branch before creating the next chunk's branch. Never branch off a previous chunk unless the plan explicitly says so ("depends on chunk N").

7. Report

Chunked {source} into {count} PRs:

  1. [#{N}]({url}) — {title}  ({lines} lines, depends on: none)
  2. [#{N}]({url}) — {title}  ({lines} lines, depends on: #{N-1})
  ...

  Parent branch:   {branch} (untouched)
  Parent PR:       {original, if any}
  Linear:          {issue-id} — {count} PR(s) linked
  Original HEAD:   {sha}  ← rollback anchor

Recommended merge order: top to bottom. Rebase each after the previous lands.

8. Learn

Save preferences silently when the user:

• Overrides --max-lines or --strategy twice in a row → update defaults
• Repeatedly merges or splits chunks the same way → nudge the classification weighting
• Picks a branch-naming pattern that differs from the saved one → update preference
• Always chooses "plan only" → switch default action to dry-run
• Consistently chooses "ready" over "draft" → update default PR type

Mention: "Noted: you prefer {X}. Saved for next time."

Principles

1. Suggest, never decide. The user approves the plan before any branch is created. They approve again before any push. No exceptions.
2. Preserve the original. Never modify or delete the parent branch. Always cherry-pick into new branches. Record the original HEAD so rollback is trivial.
3. Dependency-aware ordering. Schema before API before UI. Refactor before the feature that depends on it. Config before the code that reads it. Call out dependencies explicitly in the plan.
4. Each chunk stands alone. Every chunk must be reviewable on its own and mergeable in order — no chunk depends on an un-landed sibling except through the declared Depends on.
5. Use git, don't regenerate. Cherry-pick commits or git checkout {ref} -- {path}. Don't reconstruct diffs by hand — it loses authorship, dates, and commit messages.
6. Never bypass global rules. No pnpm build/pnpm dev. No git add -A. No generic commit messages. No regex for IDs. If a test or typecheck fails, fix it — don't skip.