ISSUE 23 · ME+MACHINE
For a long time, my entire AI setup ran on one simple strategy: I kept highly detailed instructions in Markdown files, then ran prompts like: "Follow the instructions here [file path] to do [task]."
I could describe the differences between a skill, a prompt, an agent, and a file. But as far as I was concerned, at the end of the day, they were all just Markdown files.
My methods worked great for a while. But for the last few weeks I've been bouncing back and forth between Fable and Opus 5, getting more and more frustrated as my outputs got worse and worse.
With earlier models, my default response was to write an even more detailed process. But no matter what I tried, it just wasn’t working.
So I finally read the model guides:
Opus 5 wants the entire job handed over at once, and then to be left alone to run it. The guide says it can widen a task, add steps nobody asked for, or apply its own judgment about what the task should be.
Fable 5 has the opposite problem. Its guide says skills written for older models are often too prescriptive and can make the output worse. It recommends removing your old instructions to see whether the default is better without them.
Bouncing between the two meant every adjustment for one model made the other harder to use, and my processes and files that used to produce great work were becoming increasingly degraded.
Issue 18 explained how your AI considers instructions instead of obeying every line equally. This was different. The same process was giving me opposite failures depending on which model I picked, and I never had that problem with Opus 4.7 or 4.8.
My old setup worked partly because the models went along with the rule I gave them: read this file and follow it. Putting instructions in a Markdown file stored them, but it never gave them any real authority to enforce what the AI did next.
The newer models didn't make Markdown instructions invalid. "Everything is files" is true when you're looking at what they're made of. But once the AI starts working, they do different jobs:
The same words can move between those jobs. A process document sitting in a folder is a file. Paste its instructions into today's request and they become part of the prompt. Put them in a skill, and the AI can load them as a reusable method when the right task comes up. None of those files is the agent doing the work.
This is also where knowledge and instructions separate:
Knowledge helps the AI understand the work: facts, background, examples, source material, and decisions you've made.
Instructions tell it what to do: today's assignment, the method, the format, the boundaries, and when to stop or ask you.
Say the job is analyzing customer interviews. The transcripts and client brief are files holding knowledge. "Find the three objections costing us sales" is the prompt. The method you reuse every time is the skill. The agent is the active AI opening the files, using the method, and producing the work.
And none of it is a guarantee. A skill can hold prompts and ordinary files and still not make the model obey every line inside it.
So I stopped writing longer processes and started sorting by what each piece is for..
Facts, examples, source material and past decisions stay in reference files. That is knowledge, and it was never the problem.
Methods I reuse go into skills, and I re-read those whenever I switch models.
The prompt carries what I want today: the outcome, the limits, and what finished looks like.
And I decide separately what the AI (agents) is allowed to do on its own.
For any piece of text you've written for your AI, ask which of the four it is: background, a method you'll reuse, today's request, or a limit on what it may do.
My instinct with a new model is to carry every old correction forward. Both Anthropic and OpenAI now tell you to do the opposite. Run the model on its defaults first, then add back only the instructions that fix a problem you can see.
If your setup has grown into a folder full of overlapping prompts and process files, paste the material into your AI with this:
Help me clean up the way I instruct my AI. Do not rewrite anything yet.
First, sort each passage into:
- background information the AI needs
- a method I will reuse
- directions for the current job, including when to stop
Flag repeated or conflicting instructions. Flag older instructions that may be worth testing again, but do not guess why a rule exists. Ask me if it is unclear.
Then tell me separately what the AI can access, change, or send, and which actions need my approval.
For every recommendation, tell me what should stay, what should move, and why. If a passage serves two jobs, say so.THE 30-SECOND VERSION
Opus 5 and Fable 5 broke my setup, and OpenAI's guides point the same way: fewer old instructions, clearer outcomes.
Older models needed more hand-holding. Newer ones run much further on what you asked for, so a detailed route reads as optional.
Knowledge helps your AI understand the work. Instructions tell it what to do with that information.
Files hold material, prompts give today's request, skills save reusable methods, and agents carry out the work.
Test each new model on its defaults first, then add back only the instructions that fix a problem you can see.
Now that my files are a mess, I'm reworking all of it. The material is still text, but it's going into skills, agents, and automations now, instead of just folders I point at. The mess is also why I don't mind experimenting with new tools like Grok Bot.
me+machine.