How to Audit Your Claude and Codex Setup and Clear Out AI Clutter

Key takeaways

• Clutter builds up when you hop between AI tools and install every trending skill or repo
• Too many instruction files, memory files and skills can fill the context window before your first prompt
• A short, clear audit request works; you do not need a giant prompt
• Ask for a script that scans your conversation logs, not sub-agents reading every chat
• Approve low-risk fixes first and review anything that deletes or rewrites important files
• One shared AGENTS.md plus folders per business or client keeps context where it belongs
• Before adding anything new, check need, scope and maintenance

If you have used AI tools seriously for a few months, your setup has probably grown in every direction. A skill from one video, a rule from another, a memory file you meant to tidy up, a plugin you tried once. Each one made sense at the time. Together, they can make your AI slower, more confused and more expensive.

This guide shows you how to audit your Claude and Codex setup, clear out what is not helping and keep it tidy. It is based on a recent video by Mark Kashef, explained in our own words with practical steps you can follow today, whether you use Claude Code, Codex or both.

Why does an AI setup turn into a mess?

Mostly because things are easy to add and nobody ever removes them. In the video, Mark describes a pattern he sees again and again on calls with founders:

• Tool hopping. People move from one AI coding tool to another and back again, and each move leaves its own instruction files behind
• Trend chasing. Every week brings a new skill, plugin or GitHub repo that promises to fix everything, so it gets installed and then forgotten
• Duplicate memory. Files named “memory”, “memory 2”, “final” and “real memory” all compete to tell the AI what is true
• Half-built systems. A content system, a launch system and a growth system, each started and none finished
• Stale notes. A notes vault from months ago, full of files that are now out of date, still connected to the AI

None of this is a personal failing. It is what happens when tools change this quickly. But it does need a regular clear-out.

What does the clutter actually cost you?

Every instruction file, rule and always-on skill can be loaded into the AI’s working memory, known as the context window. Load too much and you have used a chunk of it before you have typed a single prompt.

How do you audit your Claude and Codex setup?

You might expect a huge, complicated audit prompt. The video makes the opposite point: current models are capable enough that you only need to explain what to audit and what to judge it against. Here is a request in the same spirit, written in our own words:

I use both Claude and Codex and switch between them often.
Review my whole setup in both tools and tell me:
1. What is working and should stay
2. What is dead weight (rarely or never used, outdated, duplicated)
3. What is missing
Include skills, rules, instruction files (CLAUDE.md, AGENTS.md and similar),
memory files, settings, hooks, plugins, connected tools, agents, scheduled
tasks and every project I run as its own system.
Point out rules that conflict, and anything set up in one tool but not the other.
Do not read every past conversation in full. Instead, write a small script
that scans my conversation logs for tools I call (or never call) and for
moments where I pushed back or got frustrated, and use that as evidence.
Present the findings as a simple web page, scored by area, with the
top 10 changes first. Do not change or delete anything yet.

Two parts of this matter most:

• Use a script, not a reading marathon. Asking sub-agents to read hundreds or thousands of old chats end to end burns a lot of usage. A script that searches the log files for specific signals is far cheaper and still gives the AI real evidence about how you work
• Ask for a visual report. A web page is much easier to check than a long Markdown file, and you need to check it before anything is removed

What should the audit report show?

In the video, the report took around 30 minutes to produce and listed the top 10 changes, a score for each area and a long tail of smaller findings. Some were surprising, including eight old sound hooks left over from an early setup that made the computer ping when a task finished.

A good report should cover:

One useful recommendation from the video was to put all your writing rules, such as the phrases you never want the AI to use, into one shared file both tools load. That stops Claude and Codex drifting apart in tone.

How do you clean up safely?

Treat the report as a list of suggestions, not orders. Work through it in this order:

• Back up first. Copy your instruction files, skills and settings folders somewhere safe before any changes
• Approve the low-risk fixes. In the video, the Claude desktop app offered to run the low-risk configuration cleanups straight after the review. You can approve all of them, or say “everything except these”
• Review anything bigger yourself. Merging instruction files, rewriting rules or deleting projects deserves a proper look
• Test one normal task. Run something you do every week and check the result is as good as before, or better

He also points out that scanning both tools this way used only about a third of the context window, because the script did the heavy lifting. That is a nice example of working smarter, not harder.

How do you stop the clutter coming back?

A one-off clean-up helps for a while. To keep it tidy, schedule the audit. You can use a scheduled task in Codex or the schedule command in Claude, and have it re-run the report on a regular basis.

You can let the scheduled run fix the safe items without asking, for example removing temporary files you have not touched in months, and leave anything important for you to approve. The video calls the result a self-cleaning system.

The other half is your habits. If you install every new skill that a video or post promises will change everything, the pile will grow back. Our guide Should You Install a New AI Skill? A Simple 3-Question Test gives you a quick check to run before you add anything.

Can one AGENTS.md work across Claude, Codex and Gemini?

Increasingly, yes. AGENTS.md is a plain instruction file that Codex reads at the start of a session. According to the video, Claude can now also look for an AGENTS.md instead of only a CLAUDE.md, and Gemini uses a version of the same idea. That means you can aim for one set of shared instructions rather than three copies that slowly drift apart.

Think of it as overlapping circles. Most of your rules, such as your writing style and how to handle files, belong in the shared middle. Only keep something tool-specific if that feature genuinely exists in only one tool.

What does a tidy setup look like in practice?

The video walks through two founders, both given made-up names. Their problems and fixes are common, so they are worth borrowing.

Several businesses in one workspace

The first founder runs several businesses, each with its own voice and projects. Her setup had separate instruction files for three different AI tools competing at the start of every session, 15 skills of which only 3 really drove the business, a notes vault of about 1,400 notes with only around 10% still relevant, and a global instruction file that leaked into every project.

The fix was one unified AGENTS.md at the top acting as a pointer, plus one folder per business, each with its own nested AGENTS.md that only loads when you work in that folder.

Many clients and lots of know-how

The second founder runs launches and campaigns for clients: funnels, cold-traffic webinars and paid ads. She had a folder per client but no consistency, two “second brain” systems she did not use, and a real problem with one client’s launch material seeping into another’s work.

The fix was a unified AGENTS.md, a clients folder with a nested instruction file for each client, and a new playbooks folder. She recorded about an hour of dictation explaining exactly how she plans a launch, builds a funnel and runs a webinar, which became simple files such as launch.md, webinar.md and emails.md. The AI reads them when a task needs them, instead of every playbook becoming another skill.

Our guide How to Organise Your AI Workspace for Multiple Clients or Businesses turns this into a step-by-step folder plan you can copy.

What are the three rules for a useful AI setup?

Before anything new goes into your setup, the video suggests three questions. In our words:

1. Need: will this help on every task, or most of them? Would your work be worse without it? If not, could you get the same benefit by asking the AI to read the repo or skill once, pick out the one useful instruction and add just that?
2. Scope: if it is only useful sometimes, load it only when that job comes up, inside one project folder, not globally
3. Maintenance: keep running the scheduled audit, because what helped last month may now be out of date

There is a bigger point here too. As models improve, they tend to need fewer rules and skills, not more. Old instructions written for weaker models can quietly hold a better model back.

Frequently asked questions

The bottom line

A messy AI setup is normal, and it is fixable. Audit your Claude and Codex setup with a clear request and real evidence, approve the safe fixes, schedule a regular check, and be choosy about what you add next. Fewer, clearer instructions in the right places will do more for you than any new skill.

Related posts

• How to Export a Claude Code Session and Reuse It in Another AI Agent
• Claude Code Mods: How to Build Your Own Automation Shortcuts
• How to Create a Brand Context File You Can Reuse in Any AI Tool

Garry Knight

I'm Garry Knight, the person behind Prodify Digital. I write about email list building, email marketing, SEO, AI search and the tools that connect them. My aim is to make online marketing easier to understand, so creators and small business owners can make informed decisions about building an audience and keeping people engaged. Here you'll find straightforward guides and product reviews that explain what something does, where it fits and which limitations matter. The focus is on clear explanations and useful next steps—not hype, shortcuts or promises of easy earnings.

See Full Bio