Master the planning-with-files AI Skill: Persistent Thinking Like Manus

What Is planning-with-files?

planning-with-files is a newly popular Claude Code skill whose core idea comes from the AI‑agent company Manus. It externalizes the AI’s thought process, research notes, and task progress into local Markdown files, allowing the AI to recover its "soul" simply by reading those files.

Why You Need It

Standard AI chats forget details as the conversation grows, leading to hallucinations, omissions, or repeated mistakes. By persisting information in files, the AI can maintain context across long, complex tasks such as building a full app or conducting deep research.

Core Architecture: The Three‑File Pattern

task_plan.md – the brain/registry that records the ultimate goal, checklist, current progress, errors, and next steps.

notes.md – a temporary memory that stores all search results, code snippets, and long‑text references.

[deliverable].md – the final output file (e.g., report.md or main.py).

Installation Guide (for Claude Code users)

Prepare the directory : locate Claude’s skill folder.

Mac/Linux: ~/.claude/skills/ Windows: %USERPROFILE%\.claude\skills\ Deploy the code : copy the planning-with-files folder into that directory.

# Clone and move (example)
git clone https://github.com/OthmanAdi/planning-with-files.git
mv planning-with-files/planning-with-files ~/.claude/skills/

Activate : restart Claude or invoke the skill directly in Cursor/Antigravity.

How to Force Trigger in Cursor or Antigravity

Method A – Keyword Prompt

Start the conversation with an explicit command such as “Please start planning-with-files mode to research …”. Including trigger words like plan, research, or multi-step also works.

Method B – Configuration File Lock (Recommended)

Add the following rule to a .cursorrules file:

"For any complex or multi‑step tasks, follow the planning-with-files pattern: strictly maintain task_plan.md , notes.md , and a deliverable file. Update the plan after each phase."

When to Use It (Typical Scenarios)

Building a project from scratch that requires architectural decomposition and multi‑module collaboration.

Conducting deep research that involves many web sources and synthesis of core conclusions.

Major refactoring where you must record change steps and avoid forgetting original logic.

Typical Case: Researching 2025 Front‑End Frameworks

User request: “Use planning-with-files to research the three hottest front‑end frameworks for 2025 and write a comparative report in markdown.”

AI first creates task_plan.md with phases, then populates notes.md with sourced data, and finally generates a deliverable report. The generated files (shown below) contain concrete data such as adoption rates, performance numbers, and feature summaries.

# Task Plan: 2025 Front‑End Framework Comparison Report
## Goal
Research the three most popular front‑end frameworks for 2025 and produce a markdown comparison.
## Phases
- [x] Phase 1: Create plan and select frameworks
- [x] Phase 2: Research latest versions, features, performance, ecosystem
- [x] Phase 3: Consolidate results into notes.md
- [x] Phase 4: Generate comparison report
- [x] Phase 5: Review and polish report
## Key Questions
1. Which three frameworks are hottest in 2025?
2. Latest versions and core features?
3. Performance, ecosystem, learning‑curve comparison?
4. Suitable scenarios and pros/cons?
5. Market trends and future direction?
## Decisions Made
- Selected React, Vue.js, and Svelte for comparison.
## Errors Encountered
- (to be recorded)

# Notes: 2025 Front‑End Framework Research
## Sources
### Source 1: Market Survey
- React, Vue.js, Svelte are the top three.
- Latest versions: React 19, Vue 4.0, Svelte 5.0.
- Enterprise adoption: React 78%, Vue 65%, Svelte 35%.
- Svelte growth: 42% of new projects (2024‑2025).
### Source 2: Performance Tests (10 k rows list)
- React: first‑paint 3200 ms, 18 FPS, 480 MB.
- Vue.js: 2800 ms, 22 FPS, 420 MB.
- Svelte: 950 ms, 58 FPS, 210 MB.
### Source 3: Architecture Comparison
- React: virtual DOM, JSX, Hooks, Server Components.
- Vue.js: virtual DOM, template syntax, Composition API.
- Svelte: no virtual DOM, compile‑time optimization, Runes.
### Source 4: Ecosystem
- React: richest ecosystem, highest enterprise adoption.
- Vue.js: strong community, popular in Asia.
- Svelte: smaller but rapidly growing ecosystem.
## Synthesized Findings
- Svelte leads performance, especially for large‑data rendering.
- Vue.js balances performance and developer experience.
- React lags slightly but improves via Server Components.
- Learning curve: Vue.js easiest, Svelte moderate, React steepest.
- Adoption: React 78%, Vue 65%, Svelte 35%.
- Latest features: React 19 (concurrency, Server Components), Vue 4.0 (enhanced Composition API), Svelte 5.0 (Runes, partial hydration).

Advanced Tips: Why It’s a “Life‑Saving” Skill

1. Quick Recovery

If the AI crashes or you close the window, simply ask it to “continue with task_plan.md ” and it instantly knows where to resume.

2. Transparent Multi‑Person Collaboration

Colleagues can open task_plan.md to see completed items, blockers, and next phases, eliminating the need for verbal hand‑offs.

3. Fast Adjustment to Requirement Changes

Add a new phase to task_plan.md and update notes.md; the AI will integrate the new feature without losing prior context.

4. Avoid Context Pollution

All final conclusions are written to files, while discarded drafts stay in the chat history, preventing outdated information from contaminating new answers.

5. Versioning and Auditing

Since task_plan.md is a plain text file, you can track its history with Git to pinpoint when a bug was introduced.

6. Let the AI Self‑Correct

Mark failed phases directly in task_plan.md (e.g., “Phase 2 failed due to API rate‑limit, switch to alternative”) so the AI skips repeated dead‑ends.

Conclusion

Core Insight

planning-with-files is less about a specific technology and more about enforcing a disciplined thinking model that turns the AI from a “one‑off writer” into a “methodical project manager.” Files act as shared memory, making the workflow traceable, interruptible, and recoverable.

Who Benefits

Solo developers handling complex projects who need externalized to‑do lists and knowledge bases.

Content creators conducting deep research who must synthesize large amounts of information.

Limitations

The pattern is overkill for trivial tasks like renaming a variable; it shines on multi‑step, multi‑file, iterative projects. It also relies on the AI’s discipline to update the files—manual nudges may be required.

Future Outlook

As AI improves, the external‑memory model could become smarter: automatic deduplication and categorization in notes.md, template generation for task_plan.md, and richer inter‑file relationships.

Ultimately, the value lies in externalizing thought, persisting work state, and making collaboration transparent—principles distilled from high‑end AI‑agent companies like Manus and applicable to any developer.