Chapter 05 | Physical Hack: planning-with-files

4 MIN READ | UPDATED: 2026-05-15

If the previous chapters were about upgrading the Agent's "persona" and "discipline," this chapter focuses on a core technology that fundamentally shatters the theoretical ceiling of all current Large Language Models (LLMs): Memory Capacity and Context Collapse.

When our T-Block calendar management system grows to 50 files and hundreds of API communications, even the most powerful model will inevitably forget a critical constraint you mentioned in round 1 by the time you reach round 50.

This pattern is born from top-tier products like Manus: the faintest ink is better than the best memory. Replace token-based conversational history with persistent text files.

1. Principle: Using the Filesystem as the LLM's Hippocampus

Context Window (Chat Context) = Computer RAM — Cleared constantly, extremely limited capacity.
Filesystem (Project Files) = Computer Disk — Persistent, searchable by name at any time.

We must instill a mandatory, black-and-white logical concept into the Agent: "Every important decision, discovery, and plan MUST be written to the hard disk."

2. Establishing the Planning Law

Create a skill specifically designed to manage this process. The planning-with-files skill teaches the Agent to maintain a set of "Iron Triangle" files in the docs/ or root directory:

  1. task_plan.md: What are we going to do? (Broken down into checkboxes)
  2. findings.md: What did we discover while researching the current codebase or APIs?
  3. progress.md: What has been completed, what is pending, and what errors were encountered?

By forcing the Agent to read these files at the start of a session and write to them at the end, you grant it the power of continuous, infinite memory across days and sessions.