返回 Skill 列表
extension
分类: 内容与媒体无需 API Key

deepPlan

Atlas技能用于Manus AI启发的持续规划。在复杂的多步骤任务、研究项目、长期实施或需要目标锚定时使用。使用文件系统作为内存,并采用三文件结构。

person作者: jakexiaohubgithub

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 | |------|---------|---------| | task_plan.md | Master tracking with phases and checkboxes | Before/after each phase | | notes.md | Research findings, errors, intermediate results | During exploration | | [deliverable].md | Final output document | End of task |

Location: Created in current working directory.

Workflow Routing

| Workflow | Trigger | File | |----------|---------|------| | DiscoverPlans | On activation, check for existing plans | Workflows/DiscoverPlans.md | | InitializePlan | "start planning", "create plan" | Workflows/InitializePlan.md | | UpdateProgress | "update plan", "mark complete" | Workflows/UpdateProgress.md | | ReviewGoals | Before major decisions | Workflows/ReviewGoals.md |

Core Behavior

On Activation

  1. Discover existing plans - Scan ~/.claude/plans/ for related project plans
  2. Offer options - Resume, Reference, or Start Fresh
  3. Initialize files - Create task_plan.md, notes.md in current directory
  4. Sync with TodoWrite - Create in-session todos for visibility

During Work

  1. Store research findings in notes.md (keeps context clean)
  2. Update task_plan.md checkboxes as phases complete
  3. Keep TodoWrite synced for session visibility
  4. Log errors to notes.md for pattern learning

Before Major Decisions

  1. Re-read task_plan.md - Refresh goals in attention window
  2. Verify decision aligns with original objectives
  3. Check notes.md for relevant findings or past errors

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