AI Guides › Playbooks
By Nigel Guy · 6 min read
You asked for a client email, got something polite and generic, and spent twenty minutes rewording the request. Add "act as an expert", swap "write" for "draft", try again. It feels like progress because the wording is the only part you can see. Usually the problem is upstream: the model never knew who the client is, what went wrong last month, or what you want them to do next.
The rule: before you rewrite a prompt, check what the model has been given to work with. If a smart stranger could not do the job from this page alone, better wording will not save it.
Anthropic's prompting guide describes Claude as "a brilliant but new employee who lacks context on your norms and workflows", and its test is blunt: show your prompt to a colleague with minimal context on the task, and if they would be confused, the model will be too. A new colleague who gets an oddly phrased brief but has the facts will do fine. One who gets elegant phrasing and no facts will guess.
Anthropic's engineering team now frames the wider discipline as context engineering: curating "the optimal set of tokens" available to the model at each step, not only the instruction text. That covers everything the model can see: your message, earlier chat, uploaded files, saved project instructions and tool results.
Context is everything the model can read when it answers. For you, in a chat tool, that is five things:
| Layer | Example | Who controls it |
|---|---|---|
| Your message | The request itself | You, each time |
| Conversation so far | Earlier turns, corrections | You, but it grows on its own |
| Files and pasted text | A brief, a past report | You |
| Standing instructions | Project instructions, saved preferences | You, once |
| Anything it looks up | Search or connected tools | The tool, within your settings |
Claude Projects, for example, let you upload documents to a project's knowledge base and add custom instructions, so you stop re-explaining the same background. Free accounts are limited to five projects at time of writing; check the current plan page for anything newer.
The Context Brief is a short block you put at the top of any task that matters. It has five parts, in this order.
Placement matters too. For long documents, Anthropic's guidance is to put them near the top of the prompt, above your instructions and question, and to ask the model to quote the relevant passages first before it answers.
Paste this when you are staring at a task and cannot tell what is missing. Fill in the square brackets.
You are a careful briefing editor. I want to hand a task to an AI assistant and I need to know whether I have given it enough to do the job well.
My task, in my own words: [TASK]
What the result is for and who will read or use it: [PURPOSE_AND_AUDIENCE]
What I have already written or gathered (pasted below or described): [MATERIALS]
Constraints I know about (length, tone, deadline, things to avoid): [CONSTRAINTS]
Do this in order:
1. Restate the task and the finish line in two sentences, and tell me if they are not clear enough to judge a result against.
2. List the facts, files or decisions a capable stranger would still need that I have not supplied. Rank them from most to least important.
3. Ask me up to five questions to fill those gaps. Do not answer them yourself and do not guess.
4. Once I reply, write the final Context Brief with these labelled sections: Finish line, Facts, Example, Constraints, Output format.
5. Flag anything in my materials that looks irrelevant, contradictory or out of date, so I can cut it.
Before you answer, check that every question you ask is something only I can answer. If my inputs are too thin to even ask good questions, say so instead of inventing detail.
Priya runs a small bookkeeping firm. She asks an assistant: "Write a firm but friendly message to a client who is late with receipts." She gets a generic nudge and rewrites the request four times.
She runs the brief builder instead. It asks: how late, what is at stake, what have you already said, what is the client like? Her answers: three weeks late, her filing deadline is the real pressure, she has chased twice, the client responds well to short messages with one clear task. The finished brief names the finish line (client sends receipts by Friday), includes the facts, and attaches one earlier message that worked. The first draft is usable. Nothing in the wording changed; the inputs did.
Once you accept that context matters, the instinct is to paste the whole archive. Anthropic's engineering post describes context rot: as the amount of text in the window grows, the model's ability to recall specific information from it decreases. More is not safer. Long chats degrade for the same reason, with early instructions buried under later detail.
Signs you have over-fed it: the model quotes irrelevant parts of your files, repeats an instruction you retired ten turns ago, or contradicts something in the middle of the thread. Cut, summarise, and restart with the carry-forward note.