dev-create-pr

Examine all uncommitted changes, identify what belongs to the current task, prompt the user about anything ambiguous, create a branch (if on main), commit with a descriptive message, and open a pull request.

When to use this skill:

• New or updated lessons under lessons/
• Library code, skills, CI, config, scripts, tests
• Mixed infrastructure work
• Bug fixes, refactors, or tooling changes

Arguments

• --all (optional): Include all uncommitted changes without prompting. Skips the relevance check in Phase 2. Use when you know every change in the tree belongs to the same task.
• branch-name (optional): Explicit branch name to use. If omitted, the skill infers one from the changes (see Phase 3).

Workflow

Work through each phase in order.

Phase 1 — Survey the working tree

Gather a complete picture of what has changed.

Run these commands in parallel:

git status --short
git diff --stat
git diff --cached --stat
git branch --show-current
git log --oneline -5

If there are no changes (status and both diffs are empty):

• Report "Nothing to commit — working tree is clean." and stop.

Collect:

• The current branch name
• List of modified tracked files (staged and unstaged)
• List of untracked files
• List of deleted files
• Recent commit history (for context)

Phase 2 — Classify changes by relevance

Goal: Every file in the commit must belong to the same logical unit of work. Unrelated changes must not silently slip in, and must not be silently excluded.

If --all was passed, skip this phase — stage everything and go to Phase 3.

2a. Identify the task

Determine what the current work is about by examining:

1. The current branch name — e.g. lesson-30-new-topic or fix-checkbox-click tells you the scope.
2. Already-staged files (git diff --cached) — these are the most explicit signal of intent.
3. Recent commits on this branch (if not on main) — shows what has been committed so far for this task.
4. File clustering — files in the same directory or module likely belong together.

Formulate a one-line description of the task, e.g.: "Add UI Lesson 11 — Widget ID System" or "Fix checkbox click detection".

2b. Classify each changed file

For every modified, added, deleted, or untracked file, assign one of:

Heuristics for classification:

• Files in the same directory as staged changes → likely Relevant
• Files in a completely different module → likely Unrelated
• Build artifacts, compiled shaders, .o files → Unrelated (and likely gitignored already)
• Root-level project files (CMakeLists.txt, README.md, PLAN.md) → check if the changes reference the current task; if so Relevant, otherwise Unclear
• .claude/skills/ files → Relevant if the skill matches the task topic
• Test files for the module being changed → Relevant

2c. Handle each classification

Relevant files: Stage them. No prompt needed.

Unclear files: Present each to the user with context and ask:

This file changed but may not belong to the current task:

  M  common/ui/forge_ui.h  (+12 -3)

  Diff summary: Added forge_ui_widget_id() declaration

Include in this PR?

Use AskUserQuestion with options:

• "Include" — stage the file
• "Exclude" — leave it unstaged (it stays in the working tree for a future commit)
• "View diff" — show the full diff for this file, then re-ask

Unrelated files: Present them as a group and confirm exclusion:

These files appear unrelated to the current task and will NOT be included:

  ?? build/CMakeCache.txt          (build artifact)
  M  lessons/gpu/05-mipmaps/main.c (different lesson)
  ?? .DS_Store                      (OS file)

Proceed without these files?

Use AskUserQuestion with options:

• "Proceed" — leave them out
• "Let me pick" — switch to per-file prompting for this group

CRITICAL — never silently exclude. If a file has real code changes (not build artifacts or OS junk), the user must be told it is being left out. The purpose of this skill is to prevent both accidental inclusions AND accidental omissions.

Phase 3 — Create a branch (if needed)

If already on a feature branch (not main), skip to Phase 4.

If on main:

1. Ensure main is up to date:

   bashCopy

   git pull origin main
   

2. Derive a branch name from the changes. Use the task description from Phase 2 to generate a kebab-case name:

   Task description	Branch name
   Add UI Lesson 11 — Widget ID System	ui-lesson-11-widget-id
   Fix checkbox click detection	fix-checkbox-click
   Add dev-create-pr skill	add-dev-create-pr-skill
   Update math library with cross product	math-cross-product

   If the user provided a branch-name argument, use that instead.

3. Create and switch to the branch:

   bashCopy

   git checkout -b <branch-name>
   

Phase 4 — Stage, review, and local review

MANDATORY: Run /dev-local-review before committing. No exceptions. Do NOT proceed to Phase 5 until the local review passes with zero findings.

1. Stage all relevant files (those classified as Relevant in Phase 2, plus any Unclear files the user chose to include):

   bashCopy

   git add <file1> <file2> ...
   

   Stage files by explicit name — never use git add -A or git add ..

2. Run quality checks on staged content:

   • If any .md files are staged, lint only those files:

     bashCopy

     git diff --cached --name-only -- '*.md' \
       | xargs -r npx markdownlint-cli2
     

     If errors are found in staged files, fix them, re-stage, and re-check. Never bypass lint rules.

   • If any scripts/*.py or pipeline/*.py files are staged, lint only those files:

     bashCopy

     STAGED_PY=$(git diff --cached --name-only -- 'scripts/*.py' 'pipeline/*.py' 'tests/pipeline/*.py')
     [ -n "$STAGED_PY" ] && ruff check $STAGED_PY && ruff format --check $STAGED_PY
     

     Fix issues if found.

   • If any Python files are staged (scripts/, pipeline/, or tests/pipeline/), run Pyright type checking:

     bashCopy

     STAGED_PY=$(git diff --cached --name-only -- '*.py')
     [ -n "$STAGED_PY" ] && pyright
     

     Fix issues if found.

3. Show the final staging summary:

   bashCopy

   git diff --cached --stat
   

   Report the file count and total lines changed.

4. Run /dev-local-review.

   This is a mandatory gate. Run the local review skill and fix any findings it reports. Do NOT proceed to Phase 5 until it completes with zero findings. This catches problems before they reach the PR and burn a GitHub review round.

Phase 5 — Compose the commit message

Analyze the staged diff to write a descriptive commit message.

git diff --cached

Commit message format:

<type>: <short summary under 70 chars>

<body — what changed and why, wrapped at 72 chars>

Files changed:
- path/to/file.c — brief description of change
- path/to/other.h — brief description of change

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Type prefixes:

Rules for the message:

• The summary line describes the what concisely
• The body explains the why — what motivated the change, what problem it solves, what the user will see differently
• List every changed file with a brief note (helps reviewers)
• Do not pad the message with filler — be direct
• If the changes span multiple concerns, note that honestly rather than forcing a single narrative

Present the message to the user before committing. Allow them to edit it.

Phase 6 — Commit and push

0. Block secrets before commit:

   Scan staged file names for .env files or credential patterns. If any are found, refuse to proceed:

   bashCopy

   # Block .env files
   git diff --cached --name-only \
     | grep -E '(^|/)\.env(\.|$)' && {
       echo "Refusing to commit .env files. Unstage them first."
       exit 1
     }
   
   # Block credential and key files
   if git diff --cached --name-only \
     | grep -Eq '(^|/)(credentials\.json|[^/]+\.pem|[^/]+\.key)$'; then
     echo "Refusing to commit credential/key files. Unstage them first."
     exit 1
   fi
   

1. Commit:

   bashCopy

   git commit -m "$(cat <<'EOF'
   <commit message from Phase 5>
   EOF
   )"
   

2. Push with upstream tracking:

   bashCopy

   git push -u origin <branch-name>
   

3. Verify the push succeeded. If it fails (e.g. network error, permission denied), report the error and stop.

Phase 7 — Create the pull request

1. Compose the PR title and body.

   • Title: Same as the commit summary line (under 70 chars).
   • Body: Structured description derived from the commit message.

2. Create the PR:

   bashCopy

   gh pr create --title "<title>" --body "$(cat <<'EOF'
   ## Summary
   <2-4 bullet points describing what changed and why>
   
   ## Changes
   <bulleted list of files with descriptions>
   
   ## Test plan
   - [ ] <relevant verification steps>
   
   🤖 Generated with [Claude Code](https://claude.com/claude-code)
   EOF
   )"
   

3. Report the PR URL to the user.

Error handling

• No changes to commit: Report and stop — do not create an empty commit.
• Branch already exists: If the inferred branch name already exists, append a numeric suffix (-2, -3) or ask the user for a name.
• Push rejected: If the remote rejects the push (e.g. branch exists remotely with different history), report the error. Do not force-push.
• gh CLI not authenticated: Provide setup instructions (gh auth login) and stop.
• Markdown lint fails: Fix the issues — never bypass. See Phase 4.
• Merge conflicts: Should not happen on a fresh branch from main, but if they do, report and stop.

Safety guarantees

• Never uses git add -A or git add . — always stages files by name.
• Never silently ignores changed files — every non-trivial change is either staged or explicitly acknowledged by the user as excluded.
• Never silently includes unrelated changes — unclear files are always prompted.
• Never force-pushes.
• Never commits secrets or .env files — if detected in the staged set, block the commit and exit so the user can unstage them.
• Never bypasses quality checks — lint must pass before commit.
• Always shows the commit message for user approval before committing.