AI Guides › Playbooks
By Nigel Guy · 6 min read
Most people automate by picking the cleverest-sounding task: an agent that "runs your marketing", a pipeline with eleven connected apps. It demos well, breaks in week two, and nobody notices because nobody was checking. The better default is duller. Find the work you did last week, and the week before, in the same shape, and hand over only that.
The rule: automate the part that repeats, keep the part that decides, and write down which is which before you build anything.
A ledger is a table with one row per repeating job. Five columns, filled in before you touch a tool:
| Column | What you write |
|---|---|
| Job | The chore in plain words ("Monday competitor check") |
| Trigger | What starts it: a day, an email arriving, you pasting something in |
| Hand-off | The part the AI does, with the exact output shape |
| Your check | The 2-minute human step before anything leaves your hands |
| Stop rule | When you pull the job back (two bad outputs in a row, say) |
If you cannot fill in "Your check", the job is not ready to hand off. Four jobs are a sensible first page, because they are boring, frequent and low-stakes.
Trigger: once a week, same day. Hand-off: read what you paste in (articles, notes, a competitor's pricing page) and return a fixed-shape summary. Your check: open the sources for any claim you will repeat.
The trap is asking for "research" with no inputs, so the model fills gaps from memory. Give it the material, and tell it to say when the material does not answer the question.
You are a research assistant for [YOUR_ROLE] at [YOUR_BUSINESS], which [ONE_LINE_DESCRIPTION].
Goal: turn the material below into a one-page weekly digest I can read in five minutes and act on.
Material (pasted below, each item labelled with its source name and date):
[PASTED_SOURCES]
What I need to know this week: [QUESTION_OR_TOPIC]
Steps:
1. Pull out only points that bear on my question.
2. Group them under no more than four headings.
3. Under each, give the point, the source label it came from, and one line on why it matters to me.
4. End with "Not covered": anything my question needs that the material does not answer.
Rules: use only the pasted material. If a figure or date is not in it, do not supply one. If an input is missing or unreadable, ask me before writing.
Before you answer, check that every point has a source label and that nothing came from outside the material.
Fill in: your role, business, the question, and the pasted sources.
Trigger: you have a topic or a rough note. Hand-off: a first draft in your voice from your own notes. Your check: read it aloud once; cut anything you would not say.
The trap is a blank-page prompt that returns generic copy. The fix is a voice sample and a banned list. If you use Claude, a Project can hold your voice notes and examples as project knowledge, with instructions applied to every chat inside it. Projects are available on the free plan (up to 5) and unlimited on paid plans, per Anthropic's help centre.
You are an editor-writer who drafts in my voice for [AUDIENCE].
Goal: a first draft of [FORMAT, e.g. 400-word blog post / LinkedIn post] about [TOPIC] that I can finish in ten minutes.
My raw notes: [ROUGH_NOTES]
Two samples of my writing: [SAMPLE_1] / [SAMPLE_2]
Phrases and habits I avoid: [BANNED_LIST]
Steps: (1) describe my voice in three bullets from the samples; (2) propose a one-line angle and wait only if the notes are too thin to support one; (3) write the draft using only facts from my notes; (4) list every claim you could not support from my notes under "Needs checking".
Rules: British English. No invented statistics, quotes or examples. Where I have not given you a fact, leave a [GAP] marker.
Before answering, check the draft against the banned list and confirm each [GAP] is marked.
Trigger: the day you send it. Hand-off: assemble, not invent. Feed it the week's drafts, links and notes; it returns subject-line options, an intro and a tidy structure. Your check: click every link and read the subject line as a stranger would.
You are a newsletter editor for [NEWSLETTER_NAME], read by [AUDIENCE].
Goal: assemble this week's issue from the items below, in the fixed structure: subject line options (3), two-sentence opener, [NUMBER] items each with a one-sentence summary and its link, one closing line.
Items and links: [WEEKS_ITEMS]
Tone: [TONE]. Length cap: [WORD_LIMIT] words.
Rules: use only the items supplied; never add a link I did not give you; if an item is unclear, ask me. Do not write urgency language ("last chance") unless I have given a real deadline.
Before answering, check that every link in your output appears in my list.
Trigger: an email arrives that you have answered before. Hand-off: a draft reply from a saved set of your own past answers. Your check: you read it and send it yourself.
You are my correspondence assistant at [BUSINESS].
Goal: draft a reply to the message below using only the approved answers I supply.
Approved answers (policies, prices, opening hours, turnaround): [APPROVED_ANSWERS]
Message received: [PASTED_EMAIL]
Steps: (1) say which approved answer applies, or "none applies"; (2) if none applies, stop and tell me what you need; (3) otherwise draft a short, polite reply in my usual tone: [TONE_NOTE].
Rules: never promise refunds, dates or prices that are not in the approved answers. Flag anything involving a complaint, a legal threat or a personal-data request instead of answering it.
Before answering, check that each factual statement in the draft appears in the approved answers.
Do it by hand first, three weeks running, pasting the prompt yourself. Only then schedule it. ChatGPT has scheduled tasks; at time of writing OpenAI's help page lists active-task limits of 3 on Free and Go and 5 on Plus, with Free limited to one-time tasks or recurring tasks no more than once a day. Claude's pricing page lists task scheduling among Pro features. Claude Pro is $20 a month, or $17 a month billed annually (about £15 and £13 at time of writing; check the £ price at checkout). Scheduling details change often, so read the current help page before you rely on a limit.
Priya runs a three-person bookkeeping firm. Her ledger has the four jobs above. The research digest covers HMRC and competitor pages she pastes in on Monday. Content drafts come from her client-question notes. The newsletter goes out monthly. Repeat replies cover "what do you charge" and "do you do payroll". Her stop rule: if a draft contains one wrong fact twice in a row, that job returns to manual for a month. After three weeks she keeps jobs 1 and 3 on a schedule and leaves job 4 as paste-and-draft, because email carries more risk than the other three.