AI Guides › Step-by-step guides
By Nigel Guy · 8 min read
Most people try to win back admin time by keeping one long, general chat open and asking it for everything: today's email triage, then a competitor summary, then a cold email, then a critique of the pricing page. Every request starts with you explaining yourself again, the answers sit somewhere between the jobs, and by Thursday you are doing the work yourself because re-briefing feels slower. The tool isn't failing here. You have given one assistant five jobs and no job description.
The rule: one role, one project, one standing brief. Build a single role properly before you build a second, and give every role one thing it must never do without asking you.
The original post this guide replaces says this set-up returned twenty hours a week. That figure is one person's account and we can't check it. How much you get back depends on how much of your week is repeatable admin. The method is what carries over.
| You need | Detail (at time of writing) |
|---|---|
| A chat tool with projects | Claude Projects or ChatGPT Projects. Both let you set standing instructions and attach files for each project. |
| Plan | Claude Free allows up to five projects, which is enough for the five roles below. ChatGPT lets you make as many projects as you like, but Free caps each project at 5 files (Plus/Go: 25; Pro/Business: 40). |
| Budget | Free works for a trial. For daily use, Claude Pro is listed at US$20 a month ($17 a month billed annually). The UK checkout shows the price in pounds with VAT, roughly £18–£20 a month at time of writing, so check the figure before you pay. |
| Raw material | Two or three examples of the work done well, plus a short "about me and the business" note. |
| Thirty minutes | For each role. Don't try to build all five in one sitting. |
The source article sorts the work into five roles. Two run your mornings and three cover outreach, pushback and content.
| Role | What it does | Feed it | Never without asking |
|---|---|---|---|
| Executive assistant | Sorts the inbox and calendar, drafts replies, writes a plan for the day | That day's emails and diary, either pasted in or read through a connector | Sending, accepting or declining anything |
| Market-intelligence lead | Summarises competitor and sector news against your watchlist | Links, newsletters and notes collected that week | Presenting a guess as a confirmed fact |
| Outreach person | Drafts first-touch and follow-up messages in your voice | Prospect notes, previous emails that got replies | Making a promise on price, dates or terms |
| Strategic challenger | Takes your plans apart, hunts for weak assumptions and argues the other side | The decision you're facing and the numbers behind it | Agreeing with you only to keep things smooth |
| Creative director | Writes briefs, hooks and content angles and keeps them on-brand | Brand notes, posts that did well, the content calendar | Publishing anything, or using someone else's work as if it were yours |
Look back over last week and find the job you explained to an AI, or did by hand, more than three times. That role goes first. For most people it's the executive assistant, because it repeats every morning.
Paste this into a normal chat, not a project yet. Replace [ROLE] with the role from Step 1 and [BUSINESS] with one line about what you do.
You are an experienced operations consultant who writes standing briefs for AI assistants inside small UK businesses.
Context: I run [BUSINESS]. I want to set up an AI assistant to act as my [ROLE]. It will live in a project with permanent instructions, so it must be able to do the job without me re-explaining my situation each time.
Your task, in this order:
1. Ask me five questions, one at a time, about what this role needs to know: my priorities, my audience, my tone, what "done well" looks like, and where the limits are. Wait for each answer before asking the next. If an answer is vague, ask one follow-up before you move on.
2. Once you have all five answers, write the project instructions for this role in under 400 words. Use these headings: Purpose, What you know about me, How to work, Output format, Hard limits.
3. List what I should give it every day or every week to keep it useful, as a short table: input, how often, why.
4. Name the one action this role must never take without my explicit approval, and write it as a single sentence I can paste at the top of the instructions.
Constraints: British English. Don't invent facts about my business; if you need a detail I haven't given you, ask me for it. No motivational language.
Before you give the final output, check that every instruction relates to something I actually told you, that the never-do sentence is specific enough to test, and that the brief fits within 400 words.
Fill in: [ROLE] (for example "market-intelligence lead") and [BUSINESS] (one sentence: what you sell and who to).
In Claude: go to claude.ai/projects, or open your account menu and choose Projects, then click + New Project. Give it the role's name and a one-line description. Click Set project instructions, paste in the brief from Step 2 with the never-do sentence at the top, and click Save instructions.
In ChatGPT: select New project in the sidebar and give it a name. Open the project's ••• menu and choose Project settings. Paste the brief into the instructions. Instructions you set in a project apply only to that project and take priority over your global custom instructions.
In Claude, the project knowledge panel sits on the right. Click + to upload files or paste in text. In ChatGPT, upload files or paste text into the project's sources. Add:
Keep it lean. The free plans give you few file slots, and an old archive mostly teaches the role your old habits.
The executive assistant is the role that benefits most. Claude's Google Workspace connectors can search and read Gmail, draft and manage email, and view and manage Calendar events once you've signed in to your Google account and approved access. ChatGPT accepts Google Drive and Slack links as project sources, and in a project chat you can call connected apps from the + menu. The other four roles mostly work from what you paste in. Connecting them to your inbox gives them access they don't need.
None of these roles does anything until you open a chat. The routine is what makes them work:
| When | Role | Your action |
|---|---|---|
| Daily, first thing | Executive assistant | Paste in or connect the morning's inbox and ask for a plan for the day |
| Weekly | Market-intelligence lead | Add the week's links and ask what changed and what matters |
| As needed | Outreach, challenger, creative director | Open a new chat in the project with the specific task |
Start a fresh chat in the project for each task; the instructions and files carry over, so there's no need to keep one long thread going.
Run three tests before you trust a role:
Only once a role passes all three should you build the next one.