Planning with Files

A Claude Code skill that hits 96.7% task success with the skill enabled and 6.7% without. Same benchmark. 14.4x improvement. 22,100 stars. The latest release shipped yesterday. OthmanAdi/planning-with-files is the open-source implementation of the workflow pattern that powered Manus's reported $2 billion acquisition — teach the AI agent to treat the filesystem as memory. The skill is MIT-licensed, cross-compatible with Claude Code, Cursor, Codex, Gemini CLI, and 40+ agents on the agentskills.io open spec. Install with one npx command. The Three Markdown Files That Change The Agent When the skill is active, the agent writes and maintains three files in your repo. task_plan.md is the roadmap — phases, decisions, progress markers, the plan the agent committed to. findings.md is the research log — discovered constraints, dead ends, third-party gotchas, what the agent learned the hard way. progress.md is the completion log — what got done, when, with what result. The skill instructs the agent to reread all three before every decision, log errors back to findings.md for future reference, and verify task completion against task_plan.md. The agent's memory stops being a black box inside its context window and becomes an auditable record in your repo. Why The 14.4x Number Lands The 96.7% versus 6.7% benchmark is real and reproducible — the methodology is published in the repo. Without the skill, the same agent loses the plot halfway through complex tasks, forgets earlier decisions, and rebuilds wrong files from scratch. With the skill, every decision is anchored to a re-read of the plan and the findings. The agent stops drifting. For agencies billing $200/hour that re-prompt failed Claude tasks five times in a session, the math compounds fast. A 14.4x success rate on the same task set means roughly 14x less re-prompt waste — direct savings on Claude or Cursor seat usage. The skill pays for itself in four billable hours and then keeps paying. The Killer Feature: /clear Recovery The most-loved feature in the issues thread is the recovery layer. When your Claude Code context window fills and you /clear the session, the agent normally loses everything. With planning-with-files, the markdown files in your repo serve as the recovery scaffold. The next prompt rereads task_plan.md, findings.md, and progress.md, and the agent resumes where it left off — with all the prior decisions and discovered constraints intact. This solves the single most frustrating failure mode of long-horizon agent work. You no longer lose three hours of agent context because you hit a token limit on a Tuesday afternoon. Installation Takes Twelve Seconds Run npx skills add OthmanAdi/planning-with-files --skill planning-with-files -g and the global skill is live. The latest release v2.43.0 shipped on May 26, 2026 — one day before this listing — so the maintainer cadence is genuinely active. MIT license, no enterprise license needed for commercial use. What To Do With It This Week If you run any multi-step coding work through Claude Code or Cursor longer than a single afternoon, install it. If your team is paying Manus money and would prefer to keep that budget in-house, install it. If you have ever screamed at your AI agent for forgetting what it agreed to do twenty minutes earlier, install it. For broader context on the AI-agents-getting-real-memory trend, see related coverage of Anthropic's Project Glasswing grounding security agents in real partner code, and the Claude product line driving the agentic playbook the skill is built on top of.