Git flow

Branch naming rules, protected branches, whether to merge directly or through a PR / MR, whether merging triggers CI, and how to choose versions / who creates tags are project-specific rules governed by this repository's AGENTS.md / CLAUDE.md. This skill provides only general discipline.

Version / tag closeout (only for projects that manage versions with tags + CHANGELOG; a tag is a version bookmark—whether it also means release / deployment is defined by the project guide document)

• Choose the version before creating the branch: first confirm the intended tag with the owner—state the previous tag, what remains in the backlog, and whether this work is a feature or a fix, then recommend a version. Wait for the owner's confirmation before creating the branch; do not finish the work only to discover there is no agreed release target.
• Follow SemVer: major = breaking change, minor = backward-compatible feature, patch = backward-compatible fix. For 0.x, use the looser convention (minor = capability, patch = fix), and when uncertain prefer the larger bump. The project guide document defines what counts as a minor change.
• Freeze once decided: after the target version is confirmed, freeze its scope. New requirements arising midway go to the backlog / next version by default. Including one in the current version requires explicit owner approval and a clear statement of what it displaces or defers.
• Use a component, not a version number, as the commit scope (feat(auth):, not feat(v1.2):). Branches are deleted; commit history is not. The history remains accurate if the version changes.
• Version closeout sequence: complete the scope → merge the branch into the baseline → get CI green → move the CHANGELOG entry from pending into the new version → create the tag (who creates it and how are project-specific).

Branches and commits

• Create a new branch for each development effort. Never edit directly on a protected main branch or someone else's branch. If the user or client has explicitly provided a working branch, keep using it instead of creating a pointless one. Keep branches short-lived: integrate them quickly rather than letting them drift.
• Use a type + short description in branch names (for example, feat/search-pagination or fix/login-timeout); follow the project's naming convention when it has one.
• Use conventional commits: type(scope): description (feat/fix/refactor/docs/chore…). Commit in small steps, each independently reversible.

Parallel worktrees (when running several workstreams at once or isolating work from the current workspace)

• First check whether you are already in an isolated workspace. If so, do not nest another one.
• Prefer built-in worktree management in clients such as Claude Code / Codex. Do not manually run git worktree add and create a workspace the client cannot see; fall back to Git only when the client lacks this capability.
• Before creating a project-local worktree under .worktrees/, confirm that the directory is gitignored; otherwise its contents may be committed accidentally.
• After creating it, install dependencies and run the baseline tests. Report existing failures before starting work.

Wrap-up

• Verify that all tests pass before wrapping up (see verify-before-done). A failing branch must not proceed to merge / PR.
• Offer explicit choices instead of asking an open-ended “What next?”: ① merge into the baseline branch ② push and open a PR/MR ③ keep as-is ④ discard.
• Always in this order: merge and verify success → remove the worktree → delete the branch. A branch still referenced by a worktree cannot be deleted; before running git worktree remove, cd back to the main repository.
• Clean up only worktrees you created; do not touch client-created ones. Discarding requires typed confirmation. Do not force-push unless explicitly asked.