AI Guides › Playbooks
By Nigel Guy · 7 min read
Most people's AI habit is a blank chat box. You open it, re-explain who you are and what you want, get something decent, close the tab, and do the same thing tomorrow. Each answer is fine, but nothing accumulates: context, corrections and working steps vanish when the tab closes. That is chat theatre, and it looks like progress right up until you notice you've explained your business to a machine forty times this month.
The rule: a system is anything you set up once that makes the next run better or cheaper without you starting from scratch. Build the four rungs in order, and don't climb to the next one until the one below is earning its keep.
A chat is a single exchange. A system has three properties a chat doesn't:
| Property | Chat | System |
|---|---|---|
| Memory | Starts blank each time | Holds your context, files and rules |
| Repeatability | Depends on how you phrase it today | Same inputs, same shape of output |
| Improvement | Corrections are lost | Corrections are written back in |
If you can't point to where the context lives, what the fixed steps are, and where you record fixes, you have a habit, not a system.
| Rung | System | What it holds | Where it lives today | What it costs at time of writing |
|---|---|---|---|---|
| 1 | Second brain | Your reference material | A Claude Project or ChatGPT Project with files | Free tiers exist; Claude Free caps you at five projects |
| 2 | Teammate | A role, a voice and standing rules | Project instructions on the same project | Same as above |
| 3 | Workflow | A fixed sequence: input, steps, output | A saved prompt you run inside the project | Same as above |
| 4 | Automation | The workflow on a timer, without you opening it | Claude scheduled tasks or ChatGPT scheduled tasks | Claude scheduled tasks need a paid plan |
Claude Pro is listed at $20 a month on Anthropic's pricing page at time of writing; the UK price including VAT is shown at checkout, so check the £ figure there before you subscribe.
This is the knowledge layer. In Claude, a Project is a workspace with its own chat history and knowledge base: you upload files once and every chat inside the project can draw on them. Anthropic's help centre says free accounts can create up to five projects, and paid plans switch to retrieval mode to stretch knowledge capacity when you get near the limit. ChatGPT Projects do the same basic job.
How to build it:
The second brain knows things; the teammate knows how to behave. This is the project instructions field: the standing brief every chat in the project inherits. Write it like a job description, not a wish list.
You are the [ROLE, e.g. operations assistant] for [BUSINESS NAME], a [ONE-LINE DESCRIPTION] serving [AUDIENCE].
Your job in this project: [CORE RESPONSIBILITY].
Always:
- Use British English and [TONE, e.g. plain and direct].
- Treat the uploaded files as the source of truth; say which file you used.
- Flag anything time-sensitive (prices, laws, product features) as "needs checking".
Never:
- Invent figures, quotes, client names or results.
- [ANY HOUSE RULE, e.g. promise turnaround times].
When a request is missing something you need, ask one short question before starting.
Before you reply, check: does this follow the rules above, and is every factual claim traceable to a file or marked as needing a check?
Fill in the role, business, audience, core job and any house rules. Keep it under a page; long instructions get followed patchily.
A workflow is a fixed chain where each step's output feeds the next. The test is whether someone else could run it and get the same shape of result. A good default shape for content work is: idea in, research against current reality, a first draft in the right voice, then edit notes specific enough to act on. Here it is as a reusable prompt to run inside your teammate project.
Role: You are a researcher and scriptwriter working inside this project, using its files and instructions.
Goal: Turn one idea into a researched first draft plus edit notes I can act on immediately.
Input:
- Idea: [IDEA IN ONE OR TWO SENTENCES]
- Format and length: [e.g. 60-second video script / 600-word article]
- Audience: [WHO IT IS FOR]
- Voice: [VOICE NOTES, or "use the project style guide"]
Steps, in order:
1. Research. List the claims the piece will rely on. For each, say whether it is confirmed by a current source (give the source), confirmed by a project file (name it), or unconfirmed. Drop or flag anything unconfirmed; do not fill gaps with assumptions.
2. Draft. Write the first draft using only the confirmed claims, in the voice given.
3. Edit notes. Give 5 to 8 numbered notes. Each must name the exact line or section, the problem, and the specific change. No general advice such as "make it punchier".
Output format: three headed sections: "Claims check" (a table), "Draft", "Edit notes".
If the idea, format or audience is missing, ask for it before starting.
Self-check before answering: is every factual line in the draft backed by the claims table, and could each edit note be carried out without asking me what it means?
Fill in the idea, format, audience and voice. If you're on a tool without web access, step 1 will only check against your files; say so in the input so it doesn't pretend otherwise.
The final rung puts a proven workflow on a schedule. Both main assistants now offer this:
/schedule in a Cowork task. Manual set-up asks for a name, the prompt, an approval mode, a frequency (hourly, daily, weekly, weekdays or manual) and optionally a model and folder. Tasks that only use connectors can run with your computer off; tasks that need local files run on your machine and are picked up late if it was asleep.The sensible route, which Anthropic's own course also suggests: run the task by hand first, confirm the output is right, then schedule that exact process.
Priya runs a two-person bookkeeping practice in Leeds. Rung 1: a project called "Client comms" holding her engagement letter, fee schedule and five past newsletters. Rung 2: instructions making it a client-comms assistant that never quotes a tax deadline without marking it "check against GOV.UK". Rung 3: the research-draft-edit workflow, run each month on one topic for her newsletter. Only after three months of drafts she barely changed does she reach rung 4: a scheduled task on the first weekday of each month that produces a draft and edit notes for her to review. It drafts; it never sends.
Building rung 4 first. An automated version of a workflow you haven't tested just produces wrong output on time. Stacking half-set-up tools is the same trap: it's another job, not a system.