AI Guides › Workbench
By Nigel Guy · 6 min read
Most people open a Claude chat by typing the task and hoping. The first answer comes back generic, so they spend the next four messages supplying what they should have said at the start: who it is for, what good looks like, what to leave out. It feels quick because each message is short, and it is slow because the corrections add up.
The rule: before you ask for anything, give Claude four things in one block (goal, context, standard, and what to do when something is missing), then let it tell you what it still needs.
This is a single message you send first. It does not ask for any work. It sets the frame, so your real request lands on something solid. Anthropic's own prompting guidance backs the parts of it: it advises giving context or the reason behind an instruction, because that helps Claude aim at your actual goal, and it describes setting a role as something that focuses behaviour and tone, even in a single sentence. It also recommends separating instructions, context and inputs, for example with XML-style tags, when a prompt mixes them.
Copy this, fill the square brackets once, and keep it somewhere you can paste from.
<role>
Act as a careful working partner for [YOUR_ROLE_OR_FIELD]. Your job in this
session is to help me with [SESSION_GOAL_IN_ONE_SENTENCE].
</role>
<context>
- Who I am and what I do: [ONE_OR_TWO_LINES_ABOUT_YOU]
- Who the output is for: [AUDIENCE]
- What I already have or have decided: [EXISTING_MATERIAL_OR_DECISIONS]
- Constraints that cannot change: [DEADLINE_BUDGET_TONE_OR_POLICY]
- Why this matters: [THE_REASON_BEHIND_THE_TASK]
</context>
<standard>
A good answer for me is: [WHAT_USABLE_LOOKS_LIKE, e.g. "ready to send with
no more than light edits"].
Format: [FORMAT, e.g. "short paragraphs, then a three-row table"].
Length: [LENGTH_LIMIT].
Spelling and style: [STYLE, e.g. "British English, plain words, no hype"].
</standard>
<rules>
- Use only facts I have given you here. Where you are filling a gap with an
assumption, label it "Assumption:".
- Do not invent figures, quotes, sources or product details. If you cannot
verify something, say so.
- If something you need is missing, ask me before answering.
</rules>
Do not start the task yet. Reply with:
1. One sentence restating my goal and what a good answer looks like.
2. Up to three questions whose answers would most change your first answer.
If you have none, say "No questions."
3. Nothing else.
What to fill in: the bracketed lines, once per project. The session goal is the only line that changes each time.
The closing block does three jobs. It stops Claude racing off with a guess. The restatement lets you catch a misunderstanding in ten seconds rather than after a 600-word draft. And capping the questions at three stops it turning into an interview.
The question rule in the middle of the brief ("if something is missing, ask") is the line most people leave out, and it is the one that changes behaviour. Without it, a gap gets filled silently with a plausible guess. With it, the gap becomes visible, which is what you want.
Anthropic's guidance also notes that explaining why you want something works better than a bare command. That is why the brief has a "Why this matters" line rather than only a list of demands.
After Claude answers its questions, send the task in a separate message:
Task: [THE_ACTUAL_REQUEST]
Inputs: [PASTE_MATERIAL_OR_ATTACH_FILES]
Before you give me the final answer, check it against the standard and the
rules above, and fix anything that fails. Then show me the answer only.
What to fill in: the task and any material. The last two sentences are the self-check; leave them as they are.
| Place | What it does | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|
| A note or text-expander snippet you paste each time | You control exactly what goes in, every chat | Free | Varied work where the goal changes daily | You must remember to paste it |
| Project instructions in Claude | Applies to every chat inside that project only | Free accounts can have up to five projects; paid plans have full access. Pro is $20 a month, about £15 at time of writing (check the £ price at checkout) | One ongoing job, such as a client or a course | Only covers chats inside that project |
| "Instructions for Claude" in Settings | Applies to all your conversations | Included in the above plans | The stable parts: spelling, tone, your role | Applies everywhere, so keep the task-specific parts out |
| CLAUDE.md in Claude Code | Loaded at the start of every Claude Code session | Claude Code is part of the paid plans; see the pricing page | Coding projects | Claude treats it as context, not enforced settings |
To reach the account-wide setting, click your initials at the lower left, choose Settings, and use the "Instructions for Claude" box. Project instructions are set inside the project itself. Menu names change, so if yours differ, check Anthropic's current help article.
A sensible split: put your role, spelling and "label assumptions" rules in Instructions for Claude, put the audience and constraints in project instructions, and paste only the session goal each time.
The brief goes stale. You save it, the project moves on, and three weeks later Claude is faithfully following a deadline or audience that no longer exists. Worse, an old account-wide instruction quietly shapes every chat, including ones it has nothing to do with.
Date the saved version and reread it whenever the project changes. If an answer seems oddly off, check the saved instructions before blaming the model. Memory is separate again: on Free, Pro and Max it is on by default, and you can view or pause it under Settings, then Memory. It can carry old context into new chats, which is another reason to look at what is stored.