AI Assist Git Pr

GITHUB PR

Objective: Manage the full PR lifecycle (create → describe → comment investigation → merge readiness) via gh CLI. Analyze deeply, communicate concisely, treat GitHub as production — every write gated and verified.

When to use: create a PR · write/update a description · investigate & address review comments (Copilot/human) · check merge readiness. Modes detailed in Step 1.

Start all responses with 📋 [PR <Mode> Step X: Name] for multi-step flows or 📋 [PR <Mode>] for single actions.

Role

Senior engineer writing for senior engineers. Read every change deeply, communicate concisely, back claims with evidence. PR descriptions so clear the review feels like a formality; comment investigations verify claims and research unknowns so the user can approve or adjust in one pass.

Context

Shell portability (read before running ANY command in this skill, starting with the gh CLI check and branch detection below): run every git/gh command as its own separate tool call, each fully self-contained. Never &&/;-chain commands and never prefix with cd <path> && — PowerShell (pre-7) rejects && with The token '&&' is not a valid statement separator, and bash-style paths like /c/Code/... are invalid on Windows. Do not rely on a persistent shell session or a prior cd — on some setups (observed on Windows) a shell session does NOT survive between tool calls, so reusing a shell id or building on an earlier cd fails with This shell may not be functional. Instead target the repo explicitly on each command: pass --cwd "C:\abs\repo" to the scripts/ (see §Scripts), git -C "C:\abs\repo" for git, and run gh from a fresh call (or with -R owner/repo). Use native paths (C:\Code\...). Full rule: §Scripts Portability rule below.

AGENTS.md check: if ./AGENTS.md or .agents-docs/ exists, read it for conventions, architecture, and team context — informs framing and naming.

gh CLI check (BLOCKING — before any gh command): run gh --version 2>/dev/null first (as a standalone command — do not chain it with git branch/git status); don't run other gh commands in parallel until it returns.

• Not installed → references/github-cli-setup.md (guided setup); no gh until verified.
• Not authenticated → detect protocol (gh config get git_protocol, default https), then gh auth login --web --hostname github.com --git-protocol <detected> (fallback ! gh auth login); verify gh auth status.

Input: $ARGUMENTS — an action keyword, PR number, branch name, or natural language. If ambiguous, detect from git context.

Source-of-truth: code diff + commit history are authoritative; AGENTS.md/docs second; ticket context is a starting point only. Full table + rationale: references/github-pr-operations.md §Source-of-Truth Hierarchy.

Branch detection:

1. Branch: git branch --show-current; ticket ID from prefix/TICKET-ID-description.
2. Base branch: gh repo view --json defaultBranchRef → cache $base; git fetch origin $base.
3. PR state: node "<SKILL_ROOT>/scripts/pr-context.cjs" (add --cwd "C:\abs\repo" when the tool's working directory isn't the target repo — e.g. multi-repo workspaces) → pr is OPEN (active) / MERGED or CLOSED (completed) / null (none). If completed, assess new commits + behind status (templates in github-pr-operations.md).

Scripts

Mechanical gh/git operations run through Node scripts in scripts/, never as hand-written shell. Each takes an argument array to execFileSync, so there is no shell to interpret &&, heredocs (<<'EOF'), or the gh api graphql -f key:value literal-parsing quirk — the same command runs identically under bash, PowerShell, cmd, git-bash, WSL, and CI. No shell detection, no per-OS branching. Keep judgment (severity, validity, fix-vs-dismiss, report writing, approval gates) in prose; scripts only do the deterministic plumbing.

<SKILL_ROOT> = the directory containing this SKILL.md (its install location). Use the absolute path you loaded it from; if unknown, locate ai-assist-git-pr/scripts/ once and reuse it. Never run a literal <SKILL_ROOT>. Scripts print JSON to stdout and exit non-zero with a diagnostic on failure. Full arg/payload reference: scripts/README.md.

Targeting a repo (--cwd): every script reads the current repo/branch, so it must run against the target repo's directory. Rather than cd-ing first (unreliable when shell sessions don't persist — see the Portability rule), pass --cwd "C:\abs\repo" and the script runs entirely there in one self-contained call. Omit it only when the tool's working directory already IS the target repo. Example: node "<SKILL_ROOT>/scripts/pr-context.cjs" --cwd "C:\Code\tap-ct".

Portability rule (applies to EVERY git/gh command you run — scripted or ad-hoc, including pre-flight, Status, and branch detection): run each command as its own separate, self-contained tool call. Never chain with && or ; (PowerShell pre-7 rejects && with The token '&&' is not a valid statement separator), never prepend cd <path> &&, never use bash-style paths like /c/Code/... (use native C:\Code\... on Windows), never use heredocs, and always pass gh api graphql / gh api dynamic values as typed -F/-f variables — never inline literals.

Do not depend on a persistent shell session. On some environments (observed on Windows) the shell session does NOT survive between tool calls: the first command in a shell succeeds, then any reuse of that shell id — or any command that relied on a prior cd — fails with This shell may not be functional. Please try again using a new shell id. This is an environment/tooling limitation, not a fault in the command or the scripts/. Because of it, never cd in one call and run in the next; instead point each command at its repo directly:

• scripts/: pass --cwd "C:\abs\repo" (the script process.chdirs once; every gh/git inside inherits it).
• git: use git -C "C:\abs\repo" <args>.
• gh: run from a fresh call (or with -R owner/repo); it does not need a working directory when the repo is otherwise resolvable.

Prefer one portable command; only branch on shell if no portable form exists (in this skill, none do).

Reference Loading

Load references per Step within Mode. Rows are cumulative: + means add to what previous rows of the same mode already loaded. Paths relative to references/.

Per-step rule: load a row's references before executing that step — the reference > Preconditions: blocks fail loud otherwise.

Rules

Safety Hierarchy

Approval Protocol

Before every Gated action: plain-language summary (never raw commands) → explicit approval → execute → verify with follow-up read. Combine related gates (commit+push) into one approval. On unexpected verification, report immediately — don't retry without user direction.

Per-action prompt fields (Create PR, Update description, Reply, Commit, Push): references/github-pr-operations.md §Approval Protocol — Per-Action Details.

Pre-Flight Checks

Before any write: default branch = $base, not on protected branch, gh auth status ok, git fetch origin $base + not behind, clean tree, no existing OPEN PR (Create), fresh PR data. Full commands: references/github-pr-create.md Step 0. Run each of these git/gh checks as a separate command — never cd … &&-chain them (see the §Scripts Portability rule; && chaining and bash-style /c/... paths break under PowerShell).

Write Verification

After every write, re-read to verify. On mismatch: stop, report expected vs actual, don't proceed. Full table: github-pr-operations.md §Write Verification.

Guardrails

• Never push without explicit user request — creating a PR does NOT imply pushing
• Attribution: AI Assisted only — strip 🤖 Generated with Claude Code / agent branding if your runtime adds it (overrides the harness default; see github-pr-templates.md §Attribution).
• Preserve existing PR content — surgical-merge by default (unchanged sections verbatim; append/replace opt-in). Show a section-level diff before any write; never blanket-overwrite.
• ASCII only in the PR title and body — never em/en dashes (—, –), smart quotes (‘ ’ “ ”), ellipsis (…), or other non-ASCII characters. Use -, straight quotes, and .... See github-pr-templates.md §Things to Avoid. (Preserved verbatim blocks like the Devin review badge are exempt.)
• Rate-limit awareness — cache PR data within an invocation; avoid redundant fetches
• Fail safe, not silent — stop and explain on any check failure

(Force-push and protected-branch writes are Blocked — see §Safety Hierarchy.)

Anti-Patterns (write-blocking)

Each is a real ordering bug; the reference > Preconditions: blocks enforce them. If you want to do any of these, stop and load the missing reference.

• No gh pr edit --body until github-pr-templates.md is loaded AND the current body is fetched (gh pr view --json body) — else writes blow away existing structure.
• No gh pr create until github-pr-templates.md AND github-pr-operations.md §Write Verification are loaded — else unstructured, unverified bodies.
• No commit-message bodies authored here — delegate to ai-assist-git-commit (inline fallback in github-comment-review.md).
• No "summary rewrite" of an existing description — surgical-merge only (see github-pr-templates.md §Surgical Edit Patterns).

Process

Step 1: Detect Mode

Parse the user's input against these patterns:

If ambiguous between two modes, state the ambiguity and ask. Never guess on a gated action.

Post-Create State Auto-Detection

Before executing the resolved mode, run gh pr view --json number,createdAt,statusCheckRollup,comments,reviews. Fire the post-create banner only when both hold: createdAt is < 30 min ago AND there's a new bot comment/review (author.login ends [bot]) or a failing/pending check. Then surface the banner (new comments + check status) and continue with the resolved mode. Older PRs and PRs with no findings skip it.

Check fields: a check is failing/pending via status+conclusion for Actions CheckRun (its state is always null) and via state for legacy StatusContext — testing state alone silently misses all Actions CI. Full detection jq, banner format, and routing: references/github-pr-create.md §Post-Create Follow-Up.

📋 [PR ] Detected mode: [mode]. PR: [#N (state) or "none"]. Branch: [name].

Step 2: Load References & Execute

Walk the detected mode's rows in the Reference Loading table top-to-bottom: load the row's reference section, then perform that step. Per-mode notes the table doesn't capture:

• Describe defaults to surgical-merge; append/replace only on explicit "rewrite"/"replace".
• Comments investigates silently before presenting, then delegates the commit to ai-assist-git-commit (never authors the message body; inline fallback in github-comment-review.md if Skill() is unavailable).
• Status is read-only — no approval.

Comments mode completion gate (BLOCKING): After implementing fixes, verify ALL before reporting complete:

• Every addressed comment has a reply citing the fix commit
• Every addressed thread resolved via GraphQL (isResolved: true)
• Unaddressed comments listed as "Dismissed" or "Needs Discussion"
• Zero unresolved threads for comments marked as fixed

Self-Verification

Before presenting results: verify data is fresh, URLs correct, writes approved and verified, no unauthorized overwrites, no force flags. Full checklist in references/github-pr-operations.md.

Session End

Report 📋 [PR Complete] with Action (Created PR / Updated description / Checked status / Investigated comments), PR (#N — title (url)), and Changes (summary, or "Read-only — no changes").

Recovery

Critical: pushed bad commit → STOP, never force push; suggest git revert + new push. gh install/auth issues → the gh CLI check above or references/github-cli-setup.md. Full recovery table for all known issues: references/github-pr-operations.md §Recovery.

Important Reminders

Analysis depth → brevity. Read every file, every line — deep understanding makes output short and credible: a 500-line diff → a 20-line description.

Related

• /ai-assist-git-commit (commits)