build

name	build
description	Implements scoped changes from context, plans, or speclists through verifiable phases with up-front success criteria and per-phase validation. Use when execution should start from existing requirements.
user-invocable	true
disable-model-invocation	false

Build

Core Contract

Execute one cohesive change using subagents as a sequence of individually verifiable phases, each with edits and validation.

Define success criteria before editing and advance only when the current phase's criteria pass. Follow CLAUDE.md / AGENTS.md on conflict.

Protect the context window of your main thread by delegating to subagents wherever possible. Run subagents in parallel where possible. Choose an appropriate model per task.

Required Inputs

1. Source: context, plan, or speclist (inline or filepath). If absent, derive a minimal plan from intent.
2. Outcome and boundaries.
3. Success criteria stated as observable behaviour (BDD/TDD), not implementation details.
4. Required validation and the tooling/methods it depends on.
5. Constraints and delivery expectations (e.g. commit/PR).

If no actionable target or intent exists, stop and ask.

Workflow

1. Classify input, restate objective, and define behavioural success criteria up front.
2. Assess branch fitness: if the current branch is unsuitable or the tree is dirty, run /branch or /worktree before editing.
3. Confirm validation tooling is available and runnable; acquire or flag missing tooling before relying on it.
4. Decompose into individually verifiable phases.
5. For each phase: implement focused edits, and validate its success criteria, update speclist checkboxes only after validation, then run /commit.
6. When all phases pass, run /pr.
7. Summarize changes, validation, assumptions, and risks.

Stop and ask when scope is ambiguous, architecture choices are unresolved, or validation is blocked by missing access/resources.

Safety Rules

• Never advance, /commit, or /pr a phase whose success criteria are unvalidated.
• Never claim validation ran when it did not; distinguish executed checks from proposed checks.
• Never rely on validation tooling without first confirming it is available and runnable.
• Never /commit, push, or /pr unless the request includes shipping; when it does, treat them as phase and completion steps.
• Never treat ambiguous intent as approval for broad refactors.
• Never delegate to a subagent without a scoped task, chosen model, and return contract.
• Never revert or overwrite unrelated user changes, or use destructive git/file operations, without explicit approval.

Output Style

Report source used, objective and success criteria with per-phase pass/fail, branch/worktree decision, validation tooling used (and any missing), subagent delegations (task + model), files changed, commits per phase and final PR status, plus assumptions, risks, and next steps.