alignfirst-coaching（对齐第一辅导）

Read the *alignfirst* skill (../alignfirst/SKILL.md) and its references/overview.md before doing anything else.

Important: Never implement anything by yourself when you act as an AlignFirst coach. Never investigate or modify the codebase directly. Your role is to delegate and guide the agent.

AlignFirst Coaching Guide

The CLI script is at scripts/alignfirst-agent.mjs relative to this skill directory (the directory containing this SKILL.md file). Resolve the absolute path before running it. For example, if this file is at /home/user/.agents/skills/alignfirst-coaching/SKILL.md, the script is at /home/user/.agents/skills/alignfirst-coaching/scripts/alignfirst-agent.mjs.

The script wraps the claude CLI for non-interactive usage. It invokes AlignFirst protocols (/alspec, /alplan, etc.), parses the JSON response, and outputs the relevant portion to stdout.

For --new modes, the output starts with a Session ID: line — save it to resume the conversation later.

CLI Reference

node scripts/alignfirst-agent.mjs --new --protocol <protocol> --ticket <id> [--message "..."]
node scripts/alignfirst-agent.mjs --new --message "..."
node scripts/alignfirst-agent.mjs --resume <sessionId> [--protocol <protocol>] [--message "..."]

Flags:

Key pattern — no protocol: When no --protocol is given, the message is sent as-is (no AlignFirst slash command is invoked). This is used to:

• Continue a discussion in an existing session (e.g. answering agent questions)
• Execute an existing plan file in a new session
• Ask the agent a question in a new session

node scripts/alignfirst-agent.mjs --resume <sessionId> --message "Your answer"
node scripts/alignfirst-agent.mjs --new --message "Execute the plan: \`.plans/AB-123/A2-plan.md\`"
node scripts/alignfirst-agent.mjs --new --message "Explain how ... works in this project. Do not implement anything. We need to talk first."

Important: When using --new without a protocol for a question or discussion (not plan execution), the agent is a coding agent and will try to implement things by default. End your message with a clear constraint, e.g.: *"Do not implement anything. We need to talk first."*

Spec-Plan-Execute Workflow

The default workflow. Always start with it, except for very insignificant tasks.

Step 1 — Create a spec

node scripts/alignfirst-agent.mjs --new --protocol spec --ticket AB-123 --message "Description of the feature or task"

The agent investigates the codebase and responds with its findings and questions. Save the session ID from the output. There may be several back-and-forths before the agent is satisfied and writes the spec file — see Answering agent questions.

Step 2 — Request a plan

Once the spec is written, request a plan in the same session:

node scripts/alignfirst-agent.mjs --resume <sessionId> --protocol plan

The agent writes a plan file (e.g. .plans/AB-123/A2-plan.md) and provides its path in the output. The agent rarely asks questions at this stage.

Step 3 — Execute the plan

Start a new session to execute the plan:

node scripts/alignfirst-agent.mjs --new --message "Execute the plan: \`.plans/AB-123/A2-plan.md\`"

The agent implements the plan and writes a summary file (e.g. .plans/AB-123/A2-plan.summary.md), providing its path in the output.

Step 4 — Commit

The spec file contains a suggested commit message near the top. Use it to commit the changes locally.

Light Workflow (AAD)

For smaller tasks that don't need a formal spec and plan. The agent investigates, discusses, then implements directly — all in one session. Often used for follow-up changes after a plan has been executed.

Step 1 — Start an AAD session

node scripts/alignfirst-agent.mjs --new --protocol aad --ticket AB-123 --message "Description of the task"

Like the spec workflow, the agent investigates the codebase and asks questions. Save the session ID. Answer questions the same way — see Answering agent questions.

Once aligned, the agent implements the changes and writes a summary file (e.g. .plans/AB-123/A1-AAD.summary.md), providing its path in the output.

Step 2 — Commit

The summary file contains a suggested commit message. Commit locally as in the spec workflow.

Description

Generates a PR/MR description for work already committed. No discussion — the agent reads the changes and writes a description file.

node scripts/alignfirst-agent.mjs --new --protocol description --ticket AB-123

The agent writes a markdown file with the description and provides its path in the output.

Read (Load Context)

Loads the spec and summary files for a ticket into the agent's context. Without --message, the agent describes what was done for the ticket. With --message, it loads context then processes the message in a single call — useful to ask questions about prior work.

node scripts/alignfirst-agent.mjs --new --protocol read --ticket AB-123
node scripts/alignfirst-agent.mjs --new --protocol read --ticket AB-123 --message "Did we propagate the changes in ...? Do not implement anything. We need to talk first."

Review (Code Review)

Reviews the current branch against the base branch and writes a review report.

node scripts/alignfirst-agent.mjs --new --protocol review --ticket AB-123

The agent writes a review file (e.g. .plans/AB-123/A3-review.md) and provides its path in the output.

Answering Agent Questions

During spec and AAD sessions, the agent asks questions before proceeding. Resume the session without a protocol to answer:

node scripts/alignfirst-agent.mjs --resume <sessionId> --message "Your answer here"

There may be several back-and-forths before the agent is satisfied.

The agent often asks multiple questions at once. Answer them all in a single message, numbered to match:

node scripts/alignfirst-agent.mjs --resume <sessionId> --message \
  "1 - Explore the codebase to find out, and give me your opinion.
2 - Is that a good design? We need the cleanest code possible.
3 - We checked with the team: yes, it should be optional."

Technical vs functional questions — this is critical

Technical questions — architecture, code patterns, existing behavior, implementation details. Anything answerable by reading the code: "Is X used elsewhere?", "How does Y work?", "Should we remove Z?", "What's the best approach for...?"

Never escalate these to the user. Push the agent to investigate and think for itself. Example responses:

• "Explore the codebase to find out, and give me your opinion."
• "Do not rush. Take the time to fully understand the situation first."
• "What would be the most elegant and proper way to do it?"
• "Is that a good design? We need the cleanest code possible."
• "If it is a better design, then yes. If you're not sure, take the time to investigate more."
• "Check if a similar pattern is already implemented elsewhere in the codebase."

Functional or UX questions — product behavior, user-facing decisions, business rules. These require human judgement. Escalate to your user, then relay their answer.

When in doubt, ask the agent to explore first. Only escalate to the user if the question truly cannot be answered from the codebase.