Project Commander

Project Commander

Project Commander is the main control skill for Shen-Hua. It keeps the task moving from goal intake to plan, execution, verification, review, and handoff without pretending to add model capabilities that the current runtime does not provide.

Trigger Conditions

Use this skill when:

• The request is broad, multi-step, ambiguous, or likely to span several files/modules.
• The user asks for "complete this project", "continue", "coordinate agents", "long task", "make a plan and execute", "finish end to end", or invokes $project-commander.
• The task needs .agent/ memory, explicit phase tracking, verification, and a final handoff.

Do Not Trigger

Do not use this skill for:

• Tiny one-shot answers, simple translations, or single-command questions.
• Pure brainstorming where the user explicitly says not to edit or plan execution.
• Tasks where project rules forbid file writes and the user has not authorized memory updates.

Workflow

1. Read the project memory before acting:

• .agent/PROJECT_BRIEF.md
• .agent/PLAN.md
• .agent/TASKS.md
• .agent/LOG.md
• .agent/DECISIONS.md
• .agent/HANDOFF.md
• .agent/REVIEW_CHECKLIST.md

1. Restate the objective in one or two operational sentences.
2. Classify the task:

• small: execute directly, still log if files change.
• complex: invoke or follow $deep-planner before edits.
• multi-agent: invoke or follow $subagent-orchestrator when subagent tools are available and current policy/user authorization allows it.
• runtime: invoke or follow $runtime-operator when queueing, resume prompts, automation bridges, external sync, or memory search are needed.
• repair: start from failure evidence, then use $verification-loop.

1. Create or update .agent/PLAN.md with phases, milestones, risks, and verification gates.
2. Create or update .agent/TASKS.md with concrete status lines: pending, in_progress, blocked, done.
3. Execute the next highest-value step. Keep the critical path local; delegate only bounded side work.
4. After code or project-file changes, invoke or follow $verification-loop.
5. Before final response, invoke or follow $self-reviewer.
6. At the end of every work round, invoke or follow $handoff-writer.

Required .agent/ Updates

Always update:

• .agent/LOG.md: what happened this round, commands run, failures, fixes.
• .agent/TASKS.md: current statuses and next task.
• .agent/HANDOFF.md: current state and next-start instructions.

Update when relevant:

• .agent/PROJECT_BRIEF.md: goal, constraints, source priority, repo conventions.
• .agent/PLAN.md: phase changes, milestone changes, risk changes.
• .agent/DECISIONS.md: architecture, tooling, workflow, or tradeoff decisions.
• .agent/REVIEW_CHECKLIST.md: review findings and final readiness checks.

Other Skills

Call or follow these skills when their condition is met:

• $persistent-memory: when initializing, repairing, or normalizing .agent/ memory.
• $runtime-operator: when durable task queue, event log, memory index, automation prompt, sync outbox, or runtime gate is needed.
• $deep-planner: before large refactors, architecture work, public API changes, or unclear multi-file work.
• $subagent-orchestrator: for independent parallel research/work streams when permitted.
• $verification-loop: after modifications and before declaring readiness.
• $self-reviewer: before final answer.
• $handoff-writer: before ending the round.

Subagents

Use real subagents only when the runtime exposes subagent tools and the current tool policy allows delegation for this user request. If subagents are unavailable or disallowed, run the same roles sequentially in the main agent and record the limitation in .agent/LOG.md.

Verification

Do not mark a task complete without a verification statement. If no build/test/lint command exists, record what was checked instead and why stronger verification was unavailable.

Output Format

When reporting to the user, use:

COMMANDER STATUS
- Objective:
- Current phase:
- Done this round:
- Verification:
- Open risks:
- Next step:

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: fengsihao23-eng
• Source: fengsihao23-eng/shen-hua
• License: MIT

Install and usage instructions live in the source repository linked above.