planning-workflow

Planning Workflow

A structured 7-step process for creating validated plans before implementation. Plans are critiqued adversarially by the plan-critic agent and enforced by the plan_gate hook.

When to Use

• Before any complex implementation (>3 files or >100 lines)
• When the approach is uncertain or multiple solutions exist
• When the feature touches multiple components
• When scope needs to be explicitly bounded

The 7-Step Workflow

| Step | Name | Description | Tools | |------|------|-------------|-------| | Step 1 | Problem Statement | Define WHY this change is needed and WHAT is in scope | Read, Grep | | Step 2 | Scope Check | Estimate files affected; halt if >50% drift from initial estimate | Glob, Grep | | Step 3 | Existing Solutions | Search codebase + web for prior art before building new | Grep, Glob, WebSearch | | Step 4 | Minimal Path | Design the smallest change that achieves the goal | Read | | Step 5 | Adversarial Critique | plan-critic agent reviews the plan (min 2 rounds) | plan-critic agent | | Step 6 | Issue Decomposition | Break into trackable issues if needed (via /create-issue) | Bash | | Step 7 | Plan Output | Write validated plan to .claude/plans/<slug>.md | Write |

Step 1: Problem Statement

Define:

• WHY this change is needed (not what, not how)
• SCOPE boundary -- what is IN and what is OUT
• Success criteria -- how do we know it's done?

Step 2: Scope Check

Estimate the number of files that will be created or modified. If during planning the actual file count exceeds the Step 1 estimate by >50%, halt the workflow and re-scope. This prevents scope creep before it starts.

Step 3: Existing Solutions

Before building anything new, search for existing solutions:

• Codebase search: Use Grep and Glob to find similar patterns in the project
• Web search: Use WebSearch to find libraries, patterns, or prior art
• Document findings: Even if nothing is found, document what was searched

This step prevents reinventing the wheel and catches cases where a library or existing code already solves the problem.

Step 4: Minimal Path

Design the smallest change that achieves the goal:

• What files need to change?
• What is the dependency order?
• What can be deferred to a follow-up?

Step 5: Adversarial Critique

The plan-critic agent reviews the plan across 5 axes:

• Assumption audit
• Scope creep detection
• Existing solution verification
• Minimalism pressure
• Uncertainty flagging

Minimum 2 critique rounds before PROCEED verdict.

Step 6: Issue Decomposition

If the plan involves multiple independent work items, decompose into GitHub issues using /create-issue. Each issue should be independently implementable.

Step 7: Plan Output

Write the final plan to .claude/plans/<slug>.md with all required sections.

Plan File Format

# Plan: <Feature Name>

## WHY + SCOPE
Why this change is needed and what is in/out of scope.

## Existing Solutions
What was searched, what was found, why existing solutions do/don't apply.

## Minimal Path
The smallest set of changes to achieve the goal.

## Files to Create/Modify
Ordered list of files with brief description of changes.

## Risks and Unknowns
What might go wrong and how to mitigate.

## Critique History
Summary of plan-critic feedback and resolutions.

Required Plan Sections

Every plan MUST contain these three sections (validated by plan_gate hook):

1. ## WHY + SCOPE -- Problem definition and boundary
2. ## Existing Solutions -- Prior art search results
3. ## Minimal Path -- Smallest viable approach

Verdicts

The plan-critic agent issues one of three verdicts:

| Verdict | Meaning | Action | |---------|---------|--------| | PROCEED | Plan is adequate | Continue to implementation | | REVISE | Plan has fixable issues | Address feedback and re-submit | | BLOCKED | Fundamental problems | Rethink the approach |

Escape Hatches

• SKIP_PLAN_CHECK=1 environment variable disables the plan_gate hook
• Simple edits (<100 lines) are never blocked by the hook
• Documentation files (.md, CHANGELOG, README) are always exempt