AI Guides › Workbench
By Nigel Guy · 8 min read
Most people start every AI chat from zero. They retype who they are, paste in last month's brief, and then spend three rounds of edits removing phrases they would never write. It feels like normal chatting while quietly costing you the same ten minutes each time.
The rule: write the standing context once, as plain files, and load it every time, so each chat starts from "who you are, what you sell and what is live" rather than from nothing.
They are ordinary text or Markdown documents, each doing one job. The assistant does not "install" them. You attach them to a Project (or paste them in), and the assistant reads them as context. That is the whole mechanism, and it is why the quality of the files matters more than the filenames.
The ten, grouped as the reader's three questions: who you are, what you are doing, and how it all gets loaded.
| # | File | Job | Update how often |
|---|---|---|---|
| 1 | Identity.md | Who you are and how you work | Twice a year |
| 2 | BrandDNA.md | Voice, with real samples | When your voice shifts |
| 3 | AntiAI.md | Banned phrases and habits | Whenever you spot a new tic |
| 4 | Audience.md | Who you write or sell to | Quarterly |
| 5 | Business.md | What you sell, prices, boundaries | When offers change |
| 6 | Projects.md | What is live this month | Weekly or fortnightly |
| 7 | SOPs.md | Your repeatable procedures | When a process changes |
| 8 | Decisions.md | Choices already made, and why | After each decision |
| 9 | Examples.md | Finished work you are proud of | When you ship something good |
| 10 | Index.md | The loader: what to read first | When you add or drop a file |
Files 4, 8, 9 and 10 are our additions to the usual six; use them only if they earn their keep (see "How to choose").
| Place | What it does | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|
| Claude Project | Files go in project knowledge; instructions apply to every chat in that project | Free accounts can create up to five projects per Anthropic's help centre; Pro was listed at $20 a month or $17 on annual billing (about £15 at time of writing; check the £ price at checkout) | Most people | Context is not shared between chats unless it is in project knowledge |
| ChatGPT Project | Upload files or paste text, plus project instructions | OpenAI says projects are available on free and paid plans; file count depends on plan | If you already live in ChatGPT | Check your plan's file limit |
| CLAUDE.md in a folder (Claude Code) | Loaded at the start of every session | Needs a paid Claude plan or API usage; check the current page | Developers and anyone working in folders of files | Treated as context, not enforced rules |
1. Identity.md. Your role, what you are good at, what you are bad at, how you like to receive answers (short first, reasoning after), and your working hours or constraints. Keep it to one page. This is the file that stops the assistant explaining your own job to you.
2. BrandDNA.md. Describe your voice in observable terms: sentence length, formality, what you say and never say, spelling variant. Then paste three to five real samples you wrote. Samples beat adjectives: "warm but professional" fits ten thousand writers, whereas your actual paragraphs fit one.
3. AntiAI.md. A list of the phrases, structures and habits you refuse: stock openers, "not X but Y" constructions, inflated adjectives, forced three-item lists, emoji, closing summaries that repeat the body. Write it from your own edits: every time you delete the same phrase, add it.
To draft the first version of these three, paste this into a chat with your own old writing attached.
You are an editor who builds a style profile from evidence.
I will attach [NUMBER] pieces of my own writing: [LIST_OF_PIECES]. My role is [YOUR_ROLE] and my readers are [READERS].
Do the following, in order:
1. Read the samples and note concrete patterns: typical sentence length, how I open and close, words I lean on, words I never use, spelling variant, level of formality.
2. Write Identity.md: a one-page profile of how I work and how I like answers delivered. Mark anything you inferred rather than saw as "(inferred)".
3. Write BrandDNA.md: voice rules as observable behaviours, each backed by a short quote from the samples.
4. Write AntiAI.md: a list of phrases and structures that do not appear in my samples but which AI writing commonly produces, so they can be banned.
Rules: quote only from the samples; do not invent facts about me; if the samples are too few to support a rule, say so and ask me for more instead of guessing.
Before you answer, check that every rule in BrandDNA.md is traceable to a quote, and remove any that is not.
Fill in: the number and names of your samples, your role, your readers.
4. Audience.md. Who you serve, what they already know, what they worry about, the words they use. Two or three real segments, not a persona with a stock-photo name.
5. Business.md. What you sell, prices, what you will not do, your terms, claims you may and may not make. Put dates on anything volatile and a "last checked" line at the top. An assistant will repeat an out-of-date price with total confidence.
6. Projects.md. One short block per live project: goal, deadline, status, next action, who is waiting. This is the file that goes stale fastest, so give it a weekly slot in your diary or accept that it will mislead you.
7. SOPs.md. Your step-by-step procedures: how you write a proposal, how you onboard a client, how you review a draft. Write the steps you actually follow, including the check you do at the end.
8. Decisions.md. A dated list: "Chose X over Y because Z." It stops the assistant reopening settled arguments.
9. Examples.md. Three to five finished pieces you would be happy to see again, each with a line on why it worked.
A prompt to turn a messy process into an SOP entry:
You are an operations writer. Turn my description of a process into one entry for SOPs.md.
Process name: [PROCESS_NAME]
How I do it today (rough notes, in any order): [ROUGH_NOTES]
What a good result looks like: [DEFINITION_OF_DONE]
Tools involved: [TOOLS]
Output exactly this format:
- Name and purpose (one line)
- Trigger: when this process starts
- Numbered steps, each beginning with a verb
- Check before finishing: three yes/no questions
- Never do: up to three items
Rules: use only what I have told you; where a step is unclear, list it under "Questions for you" rather than filling the gap; keep the whole entry under 250 words.
Before answering, check that the steps are in the order my notes imply and that nothing in the entry came from outside my notes.
10. Index.md. A short file that says what each other file is for and the order of precedence if they clash (for example, Business.md beats Examples.md on prices). Add a standing instruction: "Before answering, tell me which files you used; if you need something that is not in them, ask me."
Then load it:
Anthropic also describes account-wide "Instructions for Claude" for things that should apply everywhere. Put only your Identity basics and AntiAI list there, and leave project detail inside projects.
Hypothetical: Priya runs a three-person bookkeeping firm and writes a monthly client email. Her Business.md lists three fixed-price packages with a "checked 1 October" line. Projects.md notes the VAT-deadline reminder is due on the 20th. Her AntiAI.md bans "I hope this finds you well" and "navigating the complexities". She asks, "Draft the October email." The assistant names the three files it used, writes in her voice, and asks which package to feature because Projects.md does not say. That question is the system working.
Start with files 1, 2, 3, 5 and 6. That is a working system. Add Audience.md once you write for more than one type of reader, SOPs.md once you have repeated work, and Decisions.md and Examples.md when you notice the assistant reopening old arguments or missing your standard. Add Index.md once you pass six files.