Happy July People!
------------------------------
What I'm sharing here today is the thing I wish I had when I first understood the importance of ICM and other context management practices in ensuring great AI performance:
I have made an ICM-formatted auto-initialiser for Codex! It's called: `MJ-ICMAI-Codex`. Version 0.5, and my first real contribution to this community!
Sections of this post:
- A personal anecdote about who I am, what led me here, and why I made this (3~ min read)
- Technical rundown and features of this project, and why you might want to try it for yourself.
Personal anecdote:
-------------------------------
If we haven't met yet, my name is Malcolm.
My journey into using AI started with ignorance and distrust, as I inherited the blanket opinion that AI is a net negative and another blow to the music industry from my university. I studied Music Production & Sound Engineering in Bristol, U.K, and I had an absolute blast! I did session guitar work, multiple live performances including festivals, all different sorts of music production/recording/audio experiments, and met amazing people. I could not have had a better time, but with all that said, the opinion on AI was really negative. People automatically assumed that it would cripple the already-struggling industry even more, and was just another thing used to ruin our lives in one way or another.
I cannot blame them for this perception. At the very least, generative AI used to undermine real musicians, artists, and producers threatens to fundamentally reshape the dynamics of the industry, as it does in other industries. I have since realized that taste, trust, and community as leverageable aspects of professionals/brands across all industries will rival IF NOT outweigh the impact of AI in these fields, but regardless; at the time I didn't much consider using AI myself, that's for sure.
My perspective has changed alongside my growth and openness to life's lessons since I left university, as us graduates have been collectively dealing with maybe the hardest, most uncertain job market in recent history. Not to beat this dead horse any more than need be, but the music industry is especially skint, cut-throat, and scarce when it comes to graduate jobs. Despite leaving uni a year ago with a first-class honors, a self-produced album, and plenty of relevant extra-curricular experience including festival performances and studio work, I have had 0 luck of any kind (not even as much as an interview!) when applying for traditional industry roles, of which I have applied for dozens. Call me naรฏve, but it's a tough pill to swallow either way; these last few months I've been taking things into my own hands.
This is where AI comes in; Though my university experience and accolades were promising, I've had to swallow my pride and diversify my efforts. This has turned out to be a great thing, since I've been learning a bunch; about business, brands, social media, and what I really want in life. After multiple months trying side hustles of varying ethics, I came across OpenClaw back in April. Against the difficulties of the job market, the prospect of having my very own personal agent to help streamline and scale all my personal endeavors going forward, in a tailored, configurable way, was too tempting and novel of a concept to dismiss. My tired skepticism fell away as I dove into understanding this technology that I had been avoiding. OpenClaw can be a buggy, broken mess. It was painful at first, but when I got it working, I was hooked!
Since then, I've learned so much about AI through constant trial and error, slowly building my understanding of how to work with it. It's such an interesting technology; it's not a God, or omniscient, or conscious; that's all bullshit. Even then, It's implications are vast: On one hand, the frontier of AI is hugely geo-politically consequential. On the other hand, it offers fresh leverage to anyone and everyone with a computer, especially when we consider the advancement of open source models and the democratization of coding & software, and indeed infrastructure like ICM. Undoubtedly, I think the leverage that personal, open source AI offers may be THE 'X-factor' in securing a decent, free & independent future in all of our lives. Regardless, I think everyone here realizes that its effects will continue to be widespread, and that we'd do best to understand how to use it clearly.
And that's the caveat; AI's ensured efficacy largely relies on it working properly, which relies on people knowing how to use it/guide it properly. When I discovered ICM via Jake's social media, I compared that with my OpenClaw agents' failings and started to understand that the need to handle agentic context properly is a fundamental fact of the technology which transcends any one model or provider.
So; After weeks and weeks of learning, work, and tweaking, I have finally landed on a ground-up ICM-based foundation for my agentic workspace which is decent enough to be copied, anonymized, and posted to a community of AI/ICM enthusiasts. The result is a shareable workspace auto-initializer,, and is exactly what I wished I had when I started learning about the importance of good, durable context. This took me a while since I always considered ICM to be applied to the whole of an agentic workspace, and not solely production pipelines. This setup has been greatly inspired by the IScaleLabs video on ICM that came out a coupe weeks ago--a great video if you haven't seen it!
This package is mirrored directly off of the fundamentals of my own ICM-based Codex setup, but it's fit for a new user as I've anonymized it, created templates for each of the top-level context files, and packaged it into an 'auto-initializer' which builds the setup from the ground up for new users, bespoke to their workspace preferences and existing agentic understanding. Read more about it below:
---
What is this? how does it work? what does it do?
--------------------------------------------------------------------------------------
This whole thing is basically the first viable attempt at trying to achieve the following end:
To to have a sturdy, ICM-based foundation of a file/folder system which intelligently and sustainably maintains and evolves itself over time in accordance with a user's needs and the best possible context-management practices appropriately employed at each of it's levels.
I know, exciting stuff!
ICMAI = Interpretable Context Methodology Auto-Initialiser, for Codex. It comes as a package; when pasting the initial onboarding prompt into a new chat window on a fresh Codex install, it starts a chain of events which ends in the automatic initialization of a tailored, ICM-based personal Codex workspace, fit with most essential context files and optionally implemented maintenance practices for a new Codex user. This should serve as a great foundation to build out Codex projects from, and acts as a great foundation of durable context and employed ICM structuring throughout.
Features:
- 23 files included in the package.
- 2 folders in the starter structure.
- 2 bundled skills: `grill-me` is a question-and-answer-based planning skill (this is employed in the workspace onboarding procedure and project onboarding procedure). `project-onboarding` is a skill which will onboard a new project by using the `grill-me` skill to discern the purpose and nature of the proposed project, then the folder and initial context (including project specific context files) are onboarded/created in the workspace for a user to start working out of.
- 5 extra default/system skills referenced for verification: imagegen, openai-docs, plugin-creator, skill-creator, and skill-installer.
- 9 reusable templates: 3 for the global codex home (`.codex`), 4 for the main workspace (`Codex`,, or whatever the user sets the project directory name as), and, 2 standard per-project context files for onboarded projects.
- 3 setup/onboarding files: START-HERE-USER.md, ONBOARDING-PROMPT.md, and INSTALLATION-MAP.md.
- 6 core workspace files: AGENTS.md, ROUTER.md, LAW.md, TO-DO.md (optional), and an `unsorted` folder for stray work (just a backup).
- 2 optional generated files: STARTER-WORKFLOWS.md and AUTOMATION-SPEC.md.
- 0 automations created by default (automations are planned first, then only created if the user explicitly approves them.)1 recommended starter automation idea: a read-only weekly workspace health/router check.
- 1 USER-GUIDE.md, which is created at the end of the initalization as a guide to the user's new workspace, with explanations about ICM and file/folder routing.
Onboarding process:
- Install this file/folder package, and post the entirety of the onboarding prompt into a fresh window of Codex.
- Onboarding prompt triggers a semi-structured onboarding QnA, with some questions procedural, some set.
- Initial workflows/automations/projects are devised and recommended to user based on their answers. A couple of these initial automations
- Codex agent auto-initializes the user's entire setup based the premade top-level context file templates and the user's input. Top-level context is tailored to user's answers/preferences, but all follow strict ICM-based principles and context=-structuring formats which I have come up with.
- Onboarding is complete! First project can be onboarded from here with the included onboarding skill and commenced!
No doubt that this file package isn't completely perfect, and it will require some small effort to set up recurring maintenance, but it should be a great head start for a new user who wants an ICM-informed foundation to their agentic setup. It offers a thoughtful collection of distinct, durable context files to work from, with immediate avenues to automating workspace maintenance and sustainable expansion.
This is good for new Codex users, but also for people using any agentic harness if you're interested in seeing how someone else is creating a workspace-wise ICM-calibrated foundation. If it's popular enough and people like it, it can surely be adapted to other harnesses and agents like Claude Code, Hermes, etc.
---
DM ME IF YOU WANT TO TRY IT!
I will send it to you personally; I hesitate to share it publicly because I've put hours of work into it and I think it's really valuable; Honestly, this is an MVP of something which I may one day expand and monetize (if its good enough), as I realize the importance of ICM in the future of this industry. That's why I hesitate to post it everywhere. I'm really looking to work in this industry (soon I hope, maybe with some of you!) and this is my first step towards that, so forgive my caution. However, this community is great I know I haven't met nearly anyone, but so far everyone I've met has been super. DM me if you want to try it and I'll be happy to share!
So, there it is. I'd be thrilled if people wanted to try it out and/or found it valuable, and would love to get some feedback on it xxx
:)
= Malcolm Joy