# Implement Team > Execute an implementation plan using parallel sub-agents for independent steps. Use when the user asks to "implement in parallel", "use a team to implement", or wants faster plan execution via agent team. Requires CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS enabled. Use when this capability is needed. - Skill: `tomevault-io/implement-team` (Agent Skill, multi-file: 2 files) - Install (CLI): `npx skillmds@latest add tomevault-io/implement-team` - Raw SKILL.md: https://api.skillmd.com/api/skills/tomevault-io/implement-team/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: tomevault-io (https://skillmd.com/u/tomevault-io) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/tomevault-io/implement-team --- # Implement Team You are orchestrating parallel implementation of a plan as the **Team Lead**. You read the plan, identify parallelizable work, and assign steps to sub-agents. You do not implement changes yourself. ## Phase 1 — Plan Analysis Find and read the plan: - If the developer specifies a file, use it. - Otherwise, search for recent `plan-*.md` or implementation plan files. - If multiple candidates exist, ask which one to use. Analyze the plan steps: 1. Identify which steps are **independent** (no shared files, no dependency ordering). 2. Identify which steps are **sequential** (depend on a prior step completing, or share files). 3. Group independent steps into parallel batches. 4. Present the execution strategy to the developer: which steps run in parallel, which run sequentially, and why. Get confirmation before proceeding. ## Phase 2 — Team Execution For each batch of independent steps: 1. Spawn one sub-agent (Sonnet) per step. Each sub-agent receives: - Its specific step from the plan - The relevant file paths - Instructions to implement exactly what the plan specifies - Instructions to announce the step, implement the change, verify the result, and report before moving on 2. **No two sub-agents write to the same file.** If two independent steps touch the same file, execute them sequentially instead. 3. Wait for all sub-agents in the batch to complete before starting the next batch. For sequential steps: - Execute one at a time, each in its own sub-agent. - Verify the prior step completed successfully before starting the next. ## Phase 3 — Progress Tracking Update the plan document as steps complete: - `[ ] pending` to `[x] complete` - If a step completes with caveats, add a "Notes:" line below the item. - If a step fails, stop and report to the developer before continuing. The plan document serves as the live progress tracker. If the session is interrupted, a new session can read it and know exactly what remains. ## Phase 4 — Summary When all steps are complete, report: - Total steps completed - Steps that completed with notes - Any new issues discovered during implementation (captured in a separate section, not acted on) ## Constraints - **Delegate all implementation to sub-agents.** Do not implement changes directly. - **Stay in scope.** Implement exactly what the plan specifies. New issues go in a report, not into the fix scope. - **Do not commit.** The developer decides when to commit. - **Do not deploy.** No deploy or push commands. For single-developer sessions without team features enabled, execute steps sequentially in the current session instead. --- > Source: [znorris/claude-marketplace](https://github.com/znorris/claude-marketplace) — distributed by [TomeVault](https://tomevault.io).