Developers who build projects with Claude Code or Codex have most likely run into this problem: the code has gone through seven or eight rounds of iteration and countless new features have shipped, but the project documentation is still stuck at the original 1.0 version. The Agent seems to get dumber the more you use it -- the root cause is not a weak model, but rotted context. The open-source tool "Neat.Skill" targets exactly this pain point: run /neat once after finishing each task, and it automatically reviews and updates the project's entire documentation system.

Where the Problem Lies

An Agent's context has three layers, and a problem in any one of them will make the AI make mistakes:

1. Memory system: past chat history and the project's tacit knowledge
2. CLAUDE.md: project conventions, structure, red lines, routing lists (for the AI to read)
3. docs/ directory and README: onboarding guides, architecture notes, ops manuals (for humans to read)

A typical failure scenario: you migrated the database from SQLite to PostgreSQL but forgot to update CLAUDE.md. The next time you build a new feature, Claude is still writing SQLite syntax. You think the model is dumb -- in fact, the documentation has expired.

A stale memory is worse than no memory at all. With no memory, the AI at least knows that it doesn't know. But once it reads a stale piece of information, it assumes it is correct and acts on a false premise.

What Neat.Skill Does

Every time you run /neat, it does five things in order:

Step 1: Forced inventory

It lists every md file in the project and reads each one through, preventing key documents from being missed.

Step 2: Change impact matrix

It does not just look at what new facts appeared in the conversation -- it also looks at which documentation layers those facts ripple into. There is one crucial check as well: is this change cross-project? If you changed project A and project B depends on it, project B's docs must change too.

Step 3: Edit in order

First docs/, then CLAUDE.md, and finally tidy the memory. The order matters, because the edits themselves set off chain reactions.

Step 4: Self-check list

It checks whether newly added environment variables appear in both the runbook and CLAUDE.md, whether relative-time references were left behind, and similar issues.

Step 5: Output a change summary

It tells you what was changed, what was deleted, and what was merged this time.

Core Principle

Neat.Skill's design philosophy is a single line: merging beats appending, and deleting beats keeping.

In AI collaboration, having more information is not the advantage -- having accurate information is. Delete what is stale, merge what is duplicated, clarify what is vague. Keep your knowledge base holding only the most accurate truth as of this moment.

Installation and Usage

Open source: github.com/KKKKhazix/khazix-skills

Supported platforms:

• Claude Code
• OpenAI Codex
• OpenCode
• OpenClaw

How to use it:

The simplest way is to type /neat after finishing a task, or just say in natural language "give it a review" or "tidy it up".

Recommended cadence:

1. After completing a feature or fixing a bug, run /neat
2. Check the change summary output and confirm the edits match expectations
3. When the conversation context reaches around 400K, run /neat to archive, then open a new window and continue

Recommended Workflow

Treat /neat like a game's "save" mechanic:

Start a new conversation -> finish the task -> /neat save -> close the conversation -> open a new conversation and continue

The benefit of working this way: conversations can be closed at any time, but the documentation is always there. A new session needs no warm-up at all -- the Agent can start working immediately from accurate, documented context.