/cp - Continuous Planner - 3-File Pattern (v9)

Create a living plan that survives session breaks and enables multi-model coordination with explicit skill sequences.

Core Philosophy

Context Window = RAM (volatile, limited)
Filesystem = Disk (persistent, unlimited)
→ Anything important gets written to disk.

The 3-file pattern keeps plans alive across sessions, /clear operations, and even model switches.

When To Use

User says:

• /cp [project] - Quick shortcut for continuous planner
• "continuous plan", "living plan", "3-file plan", "plan with files"
• "create a continuous plan for [project]"
• "I want a plan that survives /clear"
• "multi-model coordination for [task]"
• "working with subagents on [project]"

The 3 Files

File Locations

Project-based (preferred):

$PROJECT_DIR/.claude/continuous/
├── task_plan.md
├── findings.md
└── progress.md

Fallback for no project:

~/.claude/plans/<name>-continuous/
├── task_plan.md
├── findings.md
└── progress.md

Workflow

For Main Agent (You)

READ before decisions:

1. Read task_plan.md to understand current state
2. Read findings.md for existing research
3. Read progress.md for last checkpoint

UPDATE after actions:

1. Update progress.md with what you did
2. Update findings.md with new discoveries
3. Update task_plan.md only when requirements change

For Subagents

Always inject this context when spawning:

prompt = f"""
You are working on a continuous planning project.

CRITICAL: You MUST follow this workflow:

1. BEFORE ANY DECISION: Read task_plan.md to understand current state
2. BEFORE INVESTIGATION: Read findings.md for existing research
3. AFTER ANY ACTION: Update progress.md with what you did
4. ON NEW DISCOVERY: Update findings.md with what you learned
5. ON REQUIREMENT CHANGE: Update task_plan.md decisions section

Plan Files Location: {plan_files}

Your Task: {task_description}

Remember: The 3 files are your single source of truth.
"""

Creating a New Continuous Plan

Step 1: Gather Requirements

Ask these questions:

1. What are you building? (One sentence)
2. What problem does this solve?
3. Who is the user?
4. What does "done" look like?
5. Any technical constraints?

Step 2: Create Directory Structure

# Determine location
if [ -n "$PROJECT_DIR" ]; then
    PLAN_DIR="$PROJECT_DIR/.claude/continuous"
else
    PLAN_DIR="$HOME/.claude/plans/$(date +%Y%m%d)-${slugified_name}-continuous"
fi

mkdir -p "$PLAN_DIR"

Step 2.5: Discover Required Skills (v9)

Use skill discovery to find which skills to use:

# Run skill discovery for the goal
python3 ~/.claude/skills/skill_discovery.py "$GOAL"

This will:

1. Match local skills to the goal
2. Identify skill gaps
3. Suggest SkillsMP searches if needed

Add skill sequence to task_plan.md:

## Skill Sequence (The Playbook)

1. **front-door** → requirements
   - Interview and triage
   - Status: ⏸️ Pending

2. **[discovered-skill]** → [output]
   - Description of what it does
   - Status: ⏸️ Pending

Step 3: Initialize Files from Templates

Copy templates and fill in:

• task_plan.md: Fill summary, problem, goals, decisions, phases
• findings.md: Initialize with empty sections
• progress.md: Add first session entry

Step 4: Sync with Beads

# Create epic
bd create "Continuous: [project-name]" -t epic --json > /tmp/epic.json
EPIC_ID=$(jq -r '.id' /tmp/epic.json)

# Create phases as subtasks
for phase in "Phase 1" "Phase 2" "Phase 3"; do
    bd create "$phase" --deps parent=$EPIC_ID --json
done

Multi-Model Coordination

The 3 files are model-agnostic. Any AI can read/write them:

┌─────────────────────────────────────────────────────────┐
│                   Shared Files                          │
│  .claude/continuous/{task_plan,findings,progress}.md   │
└─────────────────────────────────────────────────────────┘
          ▲           ▲           ▲           ▲           ▲
          │           │           │           │           │
     GLM 4.7      Opus 4.5     Codex      Gemini     MCP Tools
   (execution)  (planning)   (code)     (research)   (ZAI)

Coordination flow:

1. Opus 4.5 (planning): Reads task_plan.md, updates phases/decisions
2. GLM 4.7 (execution): Reads task_plan.md, updates progress.md after actions
3. Codex (code): Reads findings.md for context, writes code
4. Gemini (research): Updates findings.md with research
5. ZAI Vision MCP: Analyze screenshots/diagrams, diagnose errors from images
6. ZAI Web Search MCP: Real-time research for findings.md
7. ZAI Zread MCP: Explore GitHub repos for dependencies and code patterns

MCP Integration for Enhanced Planning

The ZAI MCP servers add powerful capabilities to continuous planning:

Vision MCP (Image & Video Analysis)

Usage in findings.md:

## Visual Analysis (via Vision MCP)
### [Date] - Screenshot/Diagram Analysis
- **Image**: filename.png
- **Tool Used**: diagnose_error_screenshot
- **Findings**: [analysis results]

Web Search MCP (Real-time Research)

Usage in findings.md:

### MCP-Enabled Research
#### Web Search (via webSearchPrime)
- **Query**: "[search query]"
- **Results**: [key findings with URLs]
- **Relevance**: [how this applies]

Zread MCP (GitHub Repository Intelligence)

Usage in findings.md:

#### Zread Repository Analysis
- **Repo**: owner/repo
- **Documentation**: [key docs found]
- **Code Structure**: [relevant files]

MCP Quota Awareness

Check quota before heavy MCP usage.

Session Recovery

After /clear or session break:

1. Read the 3 files:

   # Main context
   cat .claude/continuous/task_plan.md
   cat .claude/continuous/findings.md
   cat .claude/continuous/progress.md
   

2. Locate last checkpoint:

   ## Checkpoint
   **Context Level**: X%
   **Beads Sync**: Complete
   **Files Committed**: Yes
   

3. Continue from next step: Read "Next Steps" section from progress.md

Beads Integration

Plan → Beads (on creation)

# Create epic
bd create "Continuous: [project]" -t epic

# Create phases
bd create "Phase 1: [name]" --deps parent=<epic-id>
bd create "Phase 2: [name]" --deps parent=<epic-id>

# Create steps
bd create "[step 1.1]" --deps parent=<phase1-id>

Beads → Plan (on completion)

# Update progress.md with checkpoint
echo "## Checkpoint" >> progress.md
echo "**Beads Completed**: $(bd list -s completed | wc -l)" >> progress.md

Templates

See templates/ directory for:

• task_plan.md - Plan structure with phases, decisions, dependencies
• findings.md - Research log with MCP sections
• progress.md - Session log with checkpoints

Best Practices

1. ALWAYS read before writing: Don't duplicate work
2. Update progress.md frequently: After every action
3. Keep findings.md organized: Group by date/topic
4. Only update task_plan.md for decisions: Not for every small change
5. Use MCP tools: Prefer Web Search MCP over built-in WebSearch
6. Sync with beads: Keep plan and beads in sync
7. Create checkpoints: Before risky operations

Keywords

continuous plan, living plan, 3-file plan, plan with files, continuous planning, multi-model coordination, persistent plan, session recovery