AI Agents Lose Track—This Free Claude Skill Fixes It

What It Is

A developer has reverse-engineered the core planning mechanism from Manus—Meta’s $2 billion AI agent acquisition—and released it as an open-source Claude skill. The implementation addresses a fundamental problem in AI agent behavior: context drift during multi-step tasks.

The skill uses three markdown files as external memory structures. task_plan.md maintains a checkbox-based progress tracker, notes.md stores intermediate research and findings, and deliverable.md accumulates the final output. Before each tool call or decision point, the agent reads these files to reorient itself. This simple pattern prevents the common failure mode where agents lose sight of original objectives as conversation history grows and earlier instructions get buried under layers of new context.

Rather than relying on complex prompt engineering or fine-tuning, the approach treats planning documents as persistent state that exists outside the conversation window. The agent essentially consults its own todo list before taking action, similar to how developers might check project requirements before writing code.

Why It Matters

Multi-step agent tasks frequently fail not because of capability limitations but because of attention drift. An agent might start with clear instructions to research competitors, analyze pricing, and generate a report—but by the tenth tool call, it’s optimizing for the most recent subtask while forgetting the broader goal. This skill provides a lightweight solution that works within Claude’s existing architecture.

The timing is significant. Meta’s Manus acquisition validated that structured planning approaches have commercial value at scale. Having a free, MIT-licensed implementation means developers can experiment with similar patterns without enterprise budgets. Teams building agent workflows can now test whether external memory structures improve task completion rates before investing in proprietary solutions.

The approach also highlights an architectural principle: sometimes the best fix for AI limitations isn’t more sophisticated models but better scaffolding around them. Forcing agents to maintain and consult structured plans creates a form of procedural discipline that compensates for the lack of genuine goal persistence.

Getting Started

Installation requires Claude’s CLI and takes about ten seconds. Navigate to the skills directory and clone the repository:

cd ~/.claude/skills git clone https://github.com/OthmanAdi/planning-with-files.git

The skill becomes available immediately in Claude CLI sessions. When starting a multi-step task, the agent will automatically create the three markdown files in the working directory. Developers can inspect these files during execution to see how the agent is tracking progress and storing intermediate results.

For teams already using Claude for automation, the skill integrates without configuration changes. It works particularly well for tasks involving research synthesis, multi-file code generation, or any workflow where maintaining context across multiple tool invocations matters. The markdown format means humans can also read and edit the planning files if the agent goes off track.

Context

This pattern represents one approach among several for managing agent memory. Alternatives include vector databases for semantic search across past interactions, explicit state machines that define valid transitions between tasks, and prompt templates that repeatedly reinject goals into context windows.

The markdown file approach trades sophistication for simplicity. It won’t handle complex dependency graphs or parallel task execution as elegantly as dedicated agent frameworks like LangGraph or AutoGPT. But it also avoids their complexity overhead. For straightforward sequential workflows, three text files might be sufficient.

Limitations include the assumption that tasks can be decomposed into linear checklists. Creative or exploratory work that requires backtracking and revision doesn’t map cleanly to checkbox-based progress tracking. The skill also depends on Claude’s ability to consistently read and update the files correctly—file I/O errors or inconsistent formatting could break the planning loop.

The broader lesson is that effective agent design often involves constraining behavior rather than expanding capabilities. By forcing regular consultation with external planning documents, this skill creates a simple feedback mechanism that keeps agents aligned with their original instructions.