agentic-engineering-framework: Command for Claude Code

/plan - Lightweight Implementation Plan

Create a concise implementation plan for the current task. Use after exploration is done and direction is chosen, before starting to build.

Prerequisites

1. Verify an active task exists (cat .context/working/focus.yaml)
2. Task should have a clear direction — if not, run /explore first

Workflow

Step 1: Assess Scope

Read the active task file. Count deliverables. If more than 3 independent deliverables, suggest decomposition before planning.

Step 2: Write the Plan

Create docs/plans/{date}-{task-slug}.md with this structure:

# Plan: {Task Name}

**Task:** T-XXX
**Goal:** One sentence.
**Approach:** One paragraph explaining the chosen approach and why.

## Tasks

1. **{What to build}** — {what to verify} — {files touched}
2. **{What to build}** — {what to verify} — {files touched}
...

## Technical Notes

- {Key decision or constraint, one line each}

## Open Questions

- {Anything unresolved, if any}

Plan rules:

• Target: under 100 lines. If longer, the plan is too detailed.
• One line per task. No step-by-step instructions — the agent doing the work decides HOW.
• No complete code in plans. Code belongs in implementation, not planning.
• No TDD ceremony per task. Testing is part of each task's verification, not a separate step.
• No embedded skill directives ("REQUIRED SUB-SKILL", etc.) — the plan is for humans and agents alike.

Step 3: Present for Approval

Print a summary (task list + key decisions) in conversation. Ask: "Plan written to {path}. Ready to start, or changes needed?"

Wait for user approval before any implementation begins.

Rules

• This skill does ONE thing: plan. It does not start building.
• No skill chaining — do not invoke /explore or any other skill from within this skill.
• Plans are living documents — update them as work progresses, don't treat them as contracts.
• If the task is small enough to not need a plan (single file, obvious change), say so instead of writing one.