The files that run an AI practice
Most people think working well with AI is about clever prompts. For anyone doing it seriously, it isn't — it's about a handful of plain text files that tell the AI how to work, what to remember and which jobs it can do. Here's what they are, and why getting them right is the difference between an assistant that drifts and one that behaves like a colleague.
Separate the concerns, in plain text.
A capable AI needs several different kinds of guidance: how to work on your material, what to remember between sessions, how things should look, and which repeatable jobs it can carry out. Put all of that into one long prompt and it loses the thread.
The fix is the one good organisations have always used: separate the concerns, and write each down — here, in plain, version-controlled files. Each has one job and one reader. Together they are the standing brief the AI reads before the conversation even starts.
Prompts are the conversation. These files are the standing orders the AI reads before it begins.
Six files, six jobs, six readers.
Each answers one question. Knowing which question is which is most of the skill.
The standing orders
How the AI should work here — the rules, context and guardrails for your project or firm.
Read by — The AI assistant, every session
The long memory
Durable facts and decisions it carries from one session to the next, so nothing is re-learned from scratch.
Read by — The AI, at the start of each session
The playbooks
A job you repeat — a kind of draft, a review against a checklist — written down once and triggered by name.
Read by — The AI, on demand
The brand bible
How things should look and feel — colours, type and the rules behind them — so anything it builds stays on-brand.
Read by — A design-focused AI
The build manual
A tool-agnostic brief on how to build the project, readable by any AI coding tool, not just one.
Read by — Coding assistants, across tools
The front door
What the project is and how to start. The one file in the set written for people, not the AI.
Read by — Humans
Standing layers load every time; playbooks load on demand.
Two of these load automatically at the start of every session — the standing orders (CLAUDE.md) and the long memory (MEMORY.md). They're the context the AI always has. A playbook (SKILL.md) is different: it's pulled in only when a task calls for it, so the AI isn't carrying every procedure at once.
That split has a practical consequence: keep the standing files lean. Every line in them is paid for on every run, so detail belongs down in the on-demand playbooks — not in the brief the AI re-reads each time.
Where does this belong?
The skill is direction, not prompting.
In Kramer Consulting's AI-fluency framework, the competency at work here is Direction — communicating a task clearly enough that the tool can act on it. These files are how you give that direction once instead of every time: you encode how you work, what matters and what good looks like into the setup itself, and move from micro-managing each prompt to macro-managing the work.
It pairs with Delegation — deciding what to hand the AI in the first place — and it never removes the human: you write these files (often with the AI's help), then read, tune and own them. Getting an AI to know you and your standards is the same instinct, one level up: see getting AI to know you. The look-and-feel file in this set, DESIGN.md, gets its own deep dive in a companion piece.
Set your team's AI up to work like a colleague.
Kramer Consulting runs hands-on sessions that put this into practice — on your own tools, your real work, with your people owning the result.
Related guides
Getting AI to know you
Most people think the skill is prompting. The real unlock is the settings, projects and agents you’ve never opened — so ChatGPT or Copilot knows your role and acts on it without being told twice.
Read the guide The shiftFrom chatbots to agents
AI is moving from answering questions to doing tasks. What’s changing, why it matters, and the skills that stay valuable.
Read the guide