AI Guides › Playbooks
By Nigel Guy · 7 min read
Most people who use AI to produce content repeat the same three mistakes. They start every session from a blank chat and re-explain themselves badly. They ask the model for facts it has no way of knowing. And they publish the first draft because it reads smoothly. Each one feels fine at the time, because the output is fluent, arrives fast and looks finished. The cost turns up later as drift, invented details and a slow loss of trust in your own work.
This isn't a study. Each mistake has a mirror-image rule that cancels it. Written down, the three rules make one card.
The rule: for every mistake you keep making, write down the rule that makes it impossible, and put that rule where the work happens, not in your head.
No, and that matters. None of the three comes from bad writing or weak prompting. They are defaults: they are what happens when nothing in your setup stops them. A skilled writer working from a blank chat with no sources and no review step will make all three. A beginner with the card in place will make far fewer. So the fix is structural. You change the setup and stop relying on trying harder.
| # | The mistake (default) | What it looks like | The mirror rule | Where the rule lives |
|---|---|---|---|---|
| 1 | Blank-page restart | Every chat opens with "write a post about…" and a slightly different description of your audience and tone | No piece starts without the written brief loaded | Project instructions and project knowledge |
| 2 | Unsourced draft | You ask for "five stats about X" or "what's new with Y" and use whatever comes back | No factual claim survives without a source you supplied | The drafting prompt, plus the review step |
| 3 | One-pass publish | The first draft goes out because it reads well | Draft, review and fix are three separate steps | A saved review prompt you run every time |
What it looks like: you describe your audience, voice and banned phrases from memory at the top of each chat, a bit differently every time. Output drifts because the input drifts.
The mirror rule: write the brief once and make it load automatically. In Claude this is what Projects are for. Each project has its own instructions and its own knowledge files, and every chat inside the project starts with them. At time of writing, Anthropic's help centre says Projects are available on free accounts (up to five projects) and on paid plans. The paid plans can also switch to retrieval mode when project knowledge gets large.
How to set it up:
What it looks like: asking the model to supply facts, figures, quotes or "latest news" from its own memory, then publishing them. The output is confident whether or not it's right.
The mirror rule: you supply the facts and the model supplies the drafting. Anthropic's own guidance on reducing hallucinations sets out the mechanics. Explicitly allow the model to say it doesn't know. Restrict it to the documents you provide. Ask it to back each claim with a direct quote, and remove any claim it can't support. The same page says these techniques reduce hallucinations but don't eliminate them, so you still check anything that matters.
What it looks like: you read the draft once, it sounds right, and it ships. Fluency gets mistaken for accuracy.
The mirror rule: separate the steps. Anthropic's prompting guidance names "self-correction" as the most common chaining pattern: generate a draft, review it against criteria, then refine based on the review. The criteria are the point. A review with nothing written down is just a reread.
The drafting prompt enforces rule 2. Fill in the bracketed parts and paste your source material between the tags.
You are drafting content for [PUBLICATION OR BRAND NAME], written for [AUDIENCE IN ONE LINE].
Goal: a [FORMAT, e.g. 600-word newsletter section] on [TOPIC] that a reader can act on.
Use only the material inside <sources>. Do not add facts, figures, dates, names or quotes from your general knowledge.
<sources>
[PASTE YOUR NOTES, LINKS' TEXT, DATA OR TRANSCRIPTS]
</sources>
Steps:
1. List the factual claims the sources actually support, each with the exact supporting quote.
2. Draft the piece using only those claims, following the voice rules in the project instructions.
3. Where the piece needs a fact the sources don't contain, write [NEEDS SOURCE: what's missing] instead of filling the gap.
Output: the claim list first, then the draft.
If the topic, audience or format is missing or unclear, ask me before drafting. Do not guess.
Before you answer, check: does every figure, date and name in the draft appear in the claim list? If not, replace it with a [NEEDS SOURCE] marker.
Fill in: your publication, audience, format, topic and the source text.
The review prompt enforces rule 3. Run it as a separate message, or better, in a fresh chat in the same project, so the review doesn't just agree with the draft it wrote.
You are a sceptical editor reviewing a draft for [PUBLICATION OR BRAND NAME]. Your job is to find problems, not to praise.
Review the draft inside <draft> against these criteria:
1. Every factual claim is supported by the material inside <sources>. Quote the support or flag the claim.
2. It follows these voice rules: [PASTE 3-6 OF YOUR VOICE RULES].
3. It does what the brief asked: [ONE-LINE GOAL].
4. Nothing is padding: flag any paragraph that could be cut without losing meaning.
<sources>
[SAME SOURCES AS THE DRAFT]
</sources>
<draft>
[PASTE THE DRAFT]
</draft>
Output a table with columns: Location | Problem | Criterion broken | Suggested fix. Then a one-line verdict: "ready", "fix then ship" or "rewrite".
If the sources or the draft are missing, say so and stop. Do not review from memory.
Before answering, confirm you checked every number and proper noun in the draft against the sources.
Fill in: your publication, voice rules, the one-line goal, the sources and the draft. Then make the fixes yourself, or ask for a revised draft that addresses only the flagged rows.
Imagine a small bakery that writes a monthly email to customers. Before the card, the owner opens a new chat each month, types "write a friendly newsletter about our autumn menu", and gets copy that invents a "family recipe since 1952" the bakery doesn't have. They publish it because it sounds lovely.
With the card in place:
The trap is review theatre: running the review prompt, skimming the table and shipping anyway. The card only works if a "fix then ship" verdict actually stops you.
What it won't do: