Claude-skill-registry deepPlan
Atlas skill for Manus AI-inspired persistent planning. USE WHEN complex multi-step task, research project, extended implementation, OR need goal anchoring. Uses filesystem as memory with 3-file structure.
install
source · Clone the upstream repo
git clone https://github.com/majiayu000/claude-skill-registry
Claude Code · Install into ~/.claude/skills/
T=$(mktemp -d) && git clone --depth=1 https://github.com/majiayu000/claude-skill-registry "$T" && mkdir -p ~/.claude/skills && cp -r "$T/skills/data/deepplan" ~/.claude/skills/majiayu000-claude-skill-registry-deepplan && rm -rf "$T"
manifest:
skills/data/deepplan/SKILL.mdsource content
DeepPlan - Persistent File-Based Planning
Auto-routes for complex multi-step tasks requiring persistent memory.
Overview
DeepPlan implements the Manus AI "planning-with-files" pattern that uses filesystem as working memory. This combats:
- Context volatility - plans persist across session resets
- Goal drift - re-reading anchors objectives during long sessions
- Hidden failures - errors logged for pattern avoidance
- Memory fragmentation - findings stored externally, not in context
The 3-File System
| File | Purpose | Updates |
|---|---|---|
| Master tracking with phases and checkboxes | Before/after each phase |
| Research findings, errors, intermediate results | During exploration |
| Final output document | End of task |
Location: Created in current working directory.
Workflow Routing
| Workflow | Trigger | File |
|---|---|---|
| DiscoverPlans | On activation, check for existing plans | |
| InitializePlan | "start planning", "create plan" | |
| UpdateProgress | "update plan", "mark complete" | |
| ReviewGoals | Before major decisions | |
Core Behavior
On Activation
- Discover existing plans - Scan
for related project plans~/.claude/plans/ - Offer options - Resume, Reference, or Start Fresh
- Initialize files - Create task_plan.md, notes.md in current directory
- Sync with TodoWrite - Create in-session todos for visibility
During Work
- Store research findings in
(keeps context clean)notes.md - Update
checkboxes as phases completetask_plan.md - Keep TodoWrite synced for session visibility
- Log errors to
for pattern learningnotes.md
Before Major Decisions
- Re-read
- Refresh goals in attention windowtask_plan.md - Verify decision aligns with original objectives
- Check
for relevant findings or past errorsnotes.md
Integration with TodoWrite
DeepPlan expands TodoWrite, not replaces it:
| System | Purpose | Scope |
|---|---|---|
| TodoWrite | Session visibility | In-memory, current session |
| DeepPlan files | Persistent memory | On disk, survives restarts |
Sync behavior:
- Creating a plan phase → Creates matching todo
- Completing a phase → Marks todo complete
- Session restart → Reads files, rebuilds todos
File Formats
task_plan.md
--- project: project-name directory: /path/to/project created: 2026-01-07 status: in_progress --- # Task: [Task Description] ## Phases ### Phase 1: [Name] - [ ] Step 1 - [ ] Step 2 ### Phase 2: [Name] - [ ] Step 1 ## Status Updates - 2026-01-07 14:30: Started Phase 1
notes.md
# Research Notes ## Findings ### [Topic] - Finding 1 - Finding 2 ## Errors Encountered ### [Error Type] - What happened - Resolution
Examples
Example 1: Start a complex task
User: "/atlas:deep-plan Build authentication system with OAuth" → Checks ~/.claude/plans/ for existing auth plans → Creates task_plan.md with phases → Creates notes.md for research → Syncs phases to TodoWrite → Begins Phase 1
Example 2: Resume after session restart
User: "/atlas:deep-plan Continue the auth work" → Discovers task_plan.md in current directory → Reads current phase status → Rebuilds TodoWrite from checkboxes → Reads notes.md for context → Continues from last checkpoint
Example 3: Before major decision
[During implementation] → About to choose between JWT and session tokens → Re-reads task_plan.md to check original requirements → Reviews notes.md for relevant research → Makes decision aligned with goals