Planning with Files
A skill that keeps the task plan in files on disk and re-injects it every turn so work survives context clears, compaction and crashes
Medium risk
We rate an entry medium when the tool runs code, makes network calls or reads project files. Check what exactly it does before installing.
Why this level
- Registers lifecycle hooks that run scripts every turn
- Reads project files and writes plan files to disk
Install
Manual install
/plugin marketplace add OthmanAdi/planning-with-files
/plugin install planning-with-files@planning-with-filesThe plugin ships the skill, hooks and slash commands.
This is third-party code. Review the repository files before installing.
What it does
The skill makes the agent keep its plan not in the context window but in three markdown files in the project: task_plan.md with phases and checkboxes, findings.md with notes and decisions, and progress.md with a session log. Lifecycle hooks at session start, on prompt submit and before compaction re-inject the current phase into context, so work is not lost after /clear, compaction or a crash. Recovery reads only the project's plan files; a separate catchup mode, run explicitly, can read the same project's local session records for aggregate counts or bounded excerpts. There is a gated mode: the Stop hook holds the stop while an unfinished phase remains, with a cap and stall protection. Plan integrity is checked with a SHA-256 hash: a plan body changed past the approved hash is flagged as tampering. Parallel tasks get separate plan directories. It installs in Claude Code as a plugin and in dozens of other agents through the Agent Skills standard.
Who it is for. For developers and anyone running long, multi-step agent tasks that must not reset when the context clears.
Good fit when
- The task is multi-step and spans several sessions or more than five tool calls
- You need the agent to resume after /clear, compaction or a crash instead of starting over
- You need to keep the agent from declaring the work done too early
Not a fit when
- The task is short and needs no plan
- Your agent supports neither hooks nor the Agent Skills standard
Example request
Plan this task in files and resume from the current phase even after the context is clearedLimitations
The full effect comes from the lifecycle hooks, which are registered by the plugin or npx skills install but not by the npm package alone. The completion gate and continuation mode work only on hosts that support them. Plan files are gitignored by default and live locally; hosts do not sync them with each other. The recovery-turn figures in the README are described by the author as historical and from their own runs rather than a fresh measurement of current behavior.
How to disable. Remove the planning-with-files plugin via /plugin in Claude Code, or delete the skill folder from the agent's skills directory. The task_plan.md, findings.md and progress.md files can be deleted from the project by hand.
Security check
- Registers lifecycle hooks that run scripts every turn
- Reads project files and writes plan files to disk
README in short
The README presents a persistent planning skill for coding agents: the task plan lives in three markdown files on disk, and lifecycle hooks re-inject the current phase into context every turn so work survives /clear, compaction and crashes. It describes optional modes: a completion gate that holds the stop until the plan is ready, a SHA-256 hash that catches plan tampering, and an autonomous mode that saves tokens. The approach cites Manus's principle of the filesystem as working memory. It installs in Claude Code via a plugin and in other agents via npx skills and the Agent Skills standard, with native plugins for Codex, Pi, Hermes, OpenCode and DeepSeek Harness. The author's own benchmarks are included. MIT licensed.
SKILL.md
--- name: planning-with-files description: "Persistent file-based planning for multi-step AI-agent work. Keeps task_plan.md, findings.md, and progress.md on disk; lifecycle hooks inject selected project planning context. Automatic recovery reads project planning files only. Explicit session-catchup.py --metadata reads same-project local agent session records and emits aggregate counts only; --replay may emit bounded nonce-framed excerpts. Optional gated mode can request continuation only when the host supports it and never runs commands declared in Markdown. The skill has no network upload path. Use for research or work needing 5+ tool calls." user-invocable: true allowed-tools: "Read Write Edit Bash Glob Grep" metadata: version: "3.20.1" --- # Planning with Files Work like Manus: Use persistent markdown files as your "working memory on disk." ## FIRST: Restore Project State Before continuing, resolve the plan this task owns: use the installed resolve-plan-dir script with the task's PLAN_ID and PWF_PLAN_ROOT, and read task_plan.md, progress.md and findings.md from that one selected directory. Run git diff --stat to see code changes not yet recorded in the planning files.
FAQ
What exactly survives /clear?
The plan lives in task_plan.md on disk. The prompt-submit hook re-injects the current phase into context, so after a clear the agent resumes from the right place instead of re-asking the task.
Does the skill send my data anywhere?
No. Per its description the skill has no network upload path; the plan files stay local to the project.
Related
Open-source personal AI assistant on your own machine: answers in Telegram, Slack, Discord and WhatsApp, extended with skills and plugins
A self-improving agent from Nous Research with a TUI, messaging gateway, cron jobs and skills it writes itself
An open source coding agent for the terminal and desktop with build and plan modes
Open prompt library with a Claude Code plugin, MCP server and CLI: search, fetch and improve prompts and skills from an agent