SkillCommunity★ 25k

Planning With Files

Updated Jul 5, 2026

A skill that encodes the Manus AI context-engineering pattern: write task_plan.md, findings.md, and progress.md to disk so the agent survives context resets and crashes. Hooks re-inject the plan before each tool call. A completion gate prevents the agent from stopping until all phases are checked off. Supports 18+ platforms via the Agent Skills standard.

View on GitHub

What it does

npx skills add OthmanAdi/planning-with-files --skill planning-with-files -g
Install the skill globally via npx — works across Claude Code, Cursor, Codex, Gemini CLI, and 18+ other platforms.
/planning-with-files:plan
Start a planning session: creates task_plan.md, findings.md, and progress.md in the project directory.
/planning-with-files:status
Show current planning progress — phases complete, phases remaining, and active plan file.
/plan-attest
Lock the active task_plan.md with a SHA-256 attestation; hooks reject injection if the file is tampered.
PreToolUse hook (plan injection)
Re-injects the active plan before each tool call so the agent never loses its goal context mid-task.
Stop hook (completion gate)
Blocks the agent from stopping until all phases in task_plan.md are checked off.

Files (1)

OVERVIEW.mdprimary · markdown · 2.0 KB

## What it does

Planning With Files encodes the context-engineering pattern that made Manus AI worth $2B: treat the filesystem as working memory rather than stuffing everything into the context window.

For every complex task the agent creates three files:

- **task_plan.md** — Phases with checkboxes, dependencies, and acceptance criteria.
- **findings.md** — Research, discoveries, and reusable knowledge accumulated during the task.
- **progress.md** — Session log, test results, and error history so the same mistake is never repeated.

Lifecycle hooks re-inject the active plan at the start of each turn (UserPromptSubmit), remind the agent to update progress after file writes (PostToolUse), and verify completion before the agent stops (Stop hook). If the session is interrupted or context is cleared, the agent reads the plan files from disk and recovers its state automatically.

**Benchmark (v2.21.0, claude-sonnet-4-6):** 96.7% pass rate on 30 file-pattern fidelity assertions vs 6.7% without the skill. Blind A/B wins: 3/3. Competitive benchmark: 2.7x faster recovery than 6 other planning methods.

## Install

**One-line install (works on Claude Code, Cursor, Codex, Gemini CLI, and 18+ platforms):**

```bash
npx skills add OthmanAdi/planning-with-files --skill planning-with-files -g
```

**Claude Code plugin (adds /plan and /plan:status slash commands):**

```
/plugin marketplace add OthmanAdi/planning-with-files
/plugin install planning-with-files@planning-with-files
```

## Key slash commands (after plugin install)

| Command | Description |
|---------|-------------|
| /planning-with-files:plan | Start a planning session — creates task_plan.md, findings.md, and progress.md |
| /planning-with-files:status | Show current planning progress at a glance |
| /plan-attest | Lock task_plan.md with a SHA-256; hooks block injection on tampering |

## When to use

Multi-step tasks (3+ steps), research tasks, building/creating projects, or any task spanning many tool calls. Skip for simple questions, single-file edits, and quick lookups.