AgentSkillsCN

arch-start

从第一步开始，正式执行架构规划。

SKILL.md
--- frontmatter
name: arch-start
description: Start executing the architecture plan from step 1
disable-model-invocation: true

/arch-start Skill

Status: Active

Begin executing the architecture plan by starting from step 1. This skill transitions from planning to execution and implements the first step.

Usage

code
/arch-start

Prerequisites

Run /arch [request] first to generate the architecture artifacts.

Step 1: Verify Artifacts Exist

Confirm these files exist:

  • •.claude/arch/TASK_GRAPH.json
  • •.claude/arch/state.json

If either is missing, stop and output:

code
Error: Architecture artifacts not found.
Run /arch [request] first to generate the plan.

Step 2: Start the Session

Run this command:

bash
python scripts/arch_session.py start

This will:

  • •Set phase to "executing"
  • •Set current_step to the first step ID
  • •Print the current step details

Step 3: Read Current Step Details

Read .claude/arch/TASK_GRAPH.json and find the current step. Display:

code
Current Step: [id] - [title]
Goal: [goal]
Allowed Paths: [allowed_paths]
Verification: [verification commands]
Success Criteria: [success_criteria]

Step 4: Implement the Step

Implement the goal described in the current step.

CRITICAL CONSTRAINTS:

  • •ONLY modify files that match allowed_paths for this step
  • •Do NOT touch any files outside the allowed paths
  • •Do NOT skip ahead to other steps
  • •Keep changes minimal and focused on the step's goal

Step 5: Run Verification

After implementing, run each verification command listed for the step:

bash
# Run each command from the step's "verification" array
[verification command 1]
[verification command 2]
...

Step 6: Report and Stop

After verification, output the results and stop:

If verification passes:

code
Step [id] implementation complete.
Verification passed.

Run /arch-next to advance to the next step.

If verification fails:

code
Step [id] implementation complete.
Verification FAILED:
[error output]

Fix the issues and re-run verification, then run /arch-next.

Constraints

  • •Write boundaries enforced: Only modify files matching allowed_paths for the current step
  • •One step at a time: Do not implement multiple steps in one invocation
  • •Verification required: Always run verification commands before stopping
  • •User controls progression: Wait for user to run /arch-next to advance