Planning Workflow Skill

Purpose

Workflow for complex plans requiring multiple iterations and context management.

When to use

• •Tasks affecting multiple subsystems
• •Plans requiring extensive codebase exploration
• •Situations where context limits may become an issue

Workflow

Phase 1: Initial Understanding (Explore agents)

1. •Launch 1-3 Explore agents in parallel
2. •Give each agent a specific focus:
   • •Agent 1: Current implementation and patterns
   • •Agent 2: Related components/dependencies
   • •Agent 3: Test patterns and edge cases
3. •Agents report findings with file locations

Phase 2: Design (Plan agents)

1. •Launch Plan agent(s) with context from Phase 1
2. •Provide:
   • •Findings from Explore agents
   • •Specific requirements
   • •Constraints and preconditions
3. •Agent produces implementation plan

Phase 3: Consolidation

1. •Read critical files yourself
2. •Combine agent findings
3. •Ask clarifying questions to user
4. •Write final plan to plan file

Phase 4: Plan Refinement Iterations

For complex plans, perform iterative self-review until stable:

Iteration 1 - Completeness Review:

1. •Clear current context (start fresh agent)
2. •Provide agent with:
   • •Original user request
   • •Current plan file contents
3. •Agent reviews plan for:
   • •Completeness (all requirements covered?)
   • •Correctness (implementation details accurate?)
   • •Edge cases (error handling, validation?)
4. •Agent updates plan with improvements
5. •Agent reports: "Improvements found: yes/no"

Iteration 2 - Quality Review:

1. •Clear context again (fresh agent)
2. •Provide agent with:
   • •Original user request
   • •Updated plan from iteration 1
3. •Agent reviews for:
   • •Code quality (patterns, best practices?)
   • •Testing strategy (coverage, scenarios?)
   • •Verification steps (how to validate?)
4. •Agent updates plan with improvements
5. •Agent reports: "Improvements found: yes/no"

Iteration 3+ - Stability Check (if improvements were found):

1. •Clear context (fresh agent)
2. •Provide agent with:
   • •Original user request
   • •Current plan version
3. •Agent reviews for any remaining improvements
4. •If improvements found:
   • •Update plan
   • •Trigger another iteration
5. •If no improvements found:
   • •Plan is stable, proceed to user review

Convergence criteria:

• •Maximum 10 iterations to prevent infinite loops
• •Stop when agent reports "no improvements found"
• •Stop when changes are only cosmetic (formatting, wording)

Phase 5: User Review

1. •Present final plan to user
2. •Discuss each section
3. •Adjust based on feedback
4. •Document decisions and rationale

Context management tips

• •Agents start with clean context (no prior conversation)
• •Provide agents with sufficient background info in the prompt
• •Use agents for broad exploration, do targeted reads yourself
• •Summarize agent output before proceeding

Example agent prompts

Explore agent:

Search for all places where [X] is used. Focus on:

1. •Direct calls
2. •Related configuration
3. •Test coverage Report filename:line_number for each finding.

Plan agent:

Design implementation for [feature]. Context:

• •Current implementation: [summary]
• •Involved files: [list]
• •Requirements: [list] Produce step-by-step plan with concrete code changes.