git-commit · thinkflowlab — Claude Code skill: what it does, how to install
git-commit is a Claude Code skill that creates Git commits or drafts commit messages exclusively from the staged index. Runs on requests to commit staged work or generate commit notes.
Apache-2.0 ★ 413 updated 2026-11-17
npx skills add ThinkFlowLab/OpenDirector --skill git-commit -a claude-codeWhat git-commit does
The skill runs git diff --staged before executing any Git action. If the staging area is empty, it halts immediately without adding files or making empty commits. Working tree files and general conversation history are ignored when evaluating what was changed.
Commit titles follow the [type] scope: imperative description pattern using defined types like [feat] or [bugfix], while the commit body starts with an exact ## summary header. The message is written to a temporary file outside the working tree, then committed via git commit --file <message-file>.
Commands such as git add, git stage, and git commit -a are prohibited during the commit sequence. If a push is explicitly requested, it rebases onto the upstream branch first and stops if conflicts occur instead of using force push.
What it helps with
- commit staged index changes without touching working tree files
- format commit messages with standard type tags and summary blocks
- prevent unreviewed changes from being staged automatically
- handle pre-push rebase conflict resolution safely
Example prompts for the agent
- Draft a commit message for the current staged changes
- Commit the staged changes
- Create a chore commit from the current index
How to install git-commit
1Skills CLI (Vercel)
npx skills add ThinkFlowLab/OpenDirector --skill git-commit -a claude-code2Manually: put the files in place
~/.claude/skills/git-commit/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/git-commit/
1Skills CLI (Vercel)
npx skills add ThinkFlowLab/OpenDirector --skill git-commit -a codex2Manually: put the files in place
.agents/skills/git-commit/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add ThinkFlowLab/OpenDirector --skill git-commit -a cursor2Manually: put the files in place
.cursor/skills/git-commit/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install ThinkFlowLab/OpenDirector .agents/skills/git-commit2Manually: put the files in place
.github/skills/git-commit/SKILL.md1Skills CLI (Vercel)
npx skills add ThinkFlowLab/OpenDirector --skill git-commit -a gemini-cli2Platform CLI
gemini skills install https://github.com/ThinkFlowLab/OpenDirector --path .agents/skills/git-commitManually: put the files in place
.devin/skills/git-commit/SKILL.mdThe legacy .windsurf/skills/ path still works
Claude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about git-commit
How to install git-commit in Claude Code?
Run npx skills add ThinkFlowLab/OpenDirector --skill git-commit -a claude-code.
What does git-commit do?
It reads git diff --staged, structures a title with a change type, adds a ## summary body, and commits via a temporary file.
What is required for git-commit to work?
An active Git repository with staged changes. If nothing is staged, the skill halts without modifying files.
Source text
SKILL.md · Apache-2.0
--- name: git-commit description: Create a OpenDirector Git commit, or draft its message, from the current staging area only. Inspect git diff --staged first; never stage files or include unstaged work. --- # Git Commit Commit only changes the user has already staged. Creating or editing this skill does not authorize a commit. A message-only request authorizes only drafting the message; execute a commit when the user requests it. ## Workflow 1. Run `git diff --staged` before any other Git inspection command. If command output is truncated, read the remaining staged diff before describing it. For binary files, describe only what can be established from staged metadata or staged content. 2. If the diff is empty, stop and report that nothing is staged. Do not stage files, create an empty commit, or infer a commit from prior work. 3. Base the message exclusively on this staged diff. Additional inspection may use `git diff --staged --stat`, `git diff --staged --name-status`, or `git show :path` to read the index version. Working-tree files, previous conversation, and the product roadmap are not evidence of staged work. 4. Choose a type prefix and, when useful, one scope from the guidance below. Write a title and a body whose first line is exactly `## summary`. 5. Before committing, run `git diff --staged` again. If its content changed during preparation, reread the new diff and revise the message. If it became empty, stop. Do not include newly staged material without review. 6. Save the exact message to a temporary file outside the working tree and run `git commit --file <message-file>` without a pathspec or any option that changes the staging area. Respect normal sandbox permissions. 7. After success, report the commit hash and title. Inspect the resulting commit with `git log -1 --format='%h %s'` and `git show --stat --oneline HEAD` as needed. If hooks or validation fail, report the failure and stop; leave the staging area as returned, without automatic fixes or retries. ## Staging and commit boundaries - While preparing or creating a commit, never run `git add`, `git stage`, `git commit -a` / `--all`, or any equivalent staging operation, including `git update-index` or `git apply --cached` / `--index`. - Never use `git commit --include`, `--only`, or path arguments. They can commit working-tree content or bypass part of the reviewed index. - Do not unstage, reset, stash, restore, edit, or regenerate project files as part of committing. Partially staged files must retain that boundary. - Do not split staged work, amend commits, bypass hooks, or push unless separately requested. Mixed staged changes may use an umbrella title and summary bullets; do not silently select a subset. - Before any separately requested `git push`, you must first run `git rebase` onto the current branch's configured upstream. If no upstream is configured, do not push; report the blocker instead. If the rebase stops with conflicts, follow the workflow below and do not push until it completes successfully. - If a hook changes the index or files, report the observed effect. Do not undo its changes or claim that the index necessarily remained unchanged. ## Rebase conflict handling When the required pre-push rebase stops with conflicts: 1. Run `git status` and `git diff --name-only --diff-filter=U` to identify every conflicted path. Inspect both sides of each conflict before editing. 2. Resolve each conflict deliberately, preserving intended local and upstream changes. Never discard user work or select an entire side merely to make the conflict disappear. 3. Stage only resolved conflict paths with explicit `git add -- <path>...` commands. Do not stage unrelated working-tree changes; this is the only exception to the no-staging rule above. 4. Confirm `git diff --name-only --diff-filter=U` is empty and run `git diff --cached --check`. Run relevant focused validation for the resolved files when available. 5. Run `git rebase --continue`, preserving the existing commit message. Repeat these steps if a later commit conflicts. 6. After successful completion, run `git status` and ensure no rebase is in progress before pushing. If a conflict cannot be resolved confidently, run `git rebase --abort`, report the affected paths and blocker, and wait for instructions. Never use a force push to bypass conflicts. ## Title Use `[type] scope: imperative description`, omitting `scope:` when a change spans the repository or a scope would add little information. Keep the full title about 70 characters or less. Use English by default for titles and bodies; follow an explicit user language preference. Choose the type by the staged change's purpose, not its file extension: | Prefix | Use | |---|---| | `[feat]` | New user-visible behavior or a newly usable capability. | | `[bugfix]` | Correctness fixes, regressions, stale-result protection, or recovery fixes. | | `[refactor]` | Code or directory restructuring that preserves behavior. | | `[perf]` | A performance improvement supported by the staged implementation; quote measurements only when verified. | | `[docs]` | Documentation, tutorials, comments, examples, and skill instructions. | | `[test]` | Tests, fixtures, or test infrastructure. | | `[build]` | Dependencies, lockfiles, packaging, and build tooling. | | `[ci]` | Automated integration, validation, or delivery pipelines. | | `[chore]` | Repository or operational maintenance not covered above. | UI additions and visual fixes use `[feat]` or `[bugfix]`; CSS files alone do not imply a formatting-only change. When a commit includes supporting tests and docs, choose the type of its main behavior change. Scopes describe the affected area, independently of the type: | Scope | OpenDirector area | |---|---| | `web` | Browser workbench, layouts, components, and local interactions. | | `api` | HTTP / SSE entrypoints, routes, and request handling. | | `agent` | Server-side creative decisions, tools, and evaluation. | | `workflow` | Durable orchestration, dispatch, waiting, and recovery. | | `media` | Probing, proxies, audio, rendering, and export processing. | | `providers` | Language-model and image, video, or speech adapters. | | `storage` | Database, object storage, repositories, and migrations. | | `contracts` | API, event, and shared timeline contracts. | | `repo` | Repository layout, shared tooling, and repository skills. | Use a narrower product scope such as `projects`, `assistant`, `timeline`, or `exports` when clearer. The browser assistant UI is distinct from the server-side `agent`. Future architecture is context for naming, not evidence that a capability already exists. Avoid prefixes such as `[agent]` or `[media]` that mix the type of change with its subsystem. ## Body The required body starts with `## summary`, followed by a short paragraph or Markdown bullets. For multiple distinct points, use bullets starting with `- `. Explain the concrete change and its purpose; state an important preserved invariant when it helps reviewers assess a refactor or fix. Wrap prose and bullet continuation lines manually near 70 characters. Keep commands, identifiers, and URLs intact when wrapping would break them. Use a structured tool argument or a temporary message file with actual newlines; do not interpolate message text into shell command strings. Add the following only when relevant and supported: - `## validation`: commands or checks actually performed against the staged content. A passing working-tree build does not verify a partially staged version. If that distinction cannot be resolved, omit the section or explicitly describe the verification limit. Never invent passing tests. - `## breaking changes`: incompatible API, event, timeline, storage, or command changes, their impact, and a migration action when known. - `Refs: #123` or `Fixes: #123`: only for a real issue supplied by the user or verified from repository context. Use `Fixes` only when addressed. Do not require empty optional sections, repeat the title as the entire summary, include secrets, or claim simulated generation is a real model integration. Describe only the staged result, not the work chronology. Example format (not a claim about the current staging area): ```text [refactor] web: organize the workbench into feature modules ## summary - Move the frontend into apps/web and update resource imports. - Separate layout controls, project operations, and assistant types while preserving existing browser storage keys and schemas. - Update installation, development, and build instructions. ```
© ThinkFlowLab · Apache-2.0 · ThinkFlowLab/OpenDirector
Similar skills
All skills for Git and GitHub →git-commit by other authors comes first.