AI Guides › Workbench
By Nigel Guy · 8 min read
You ask a one-line question and get a warm-up sentence, three paragraphs of context, a bulleted recap of the context, and an offer to help further. Most people respond by adding "be concise" to the odd message, which works for that reply and is forgotten by the next. That is brevity theatre: you feel you have fixed it, and the padding comes back the moment you stop asking.
The rule: put the length rule where the tool reads it on every message, write it as what to do rather than what to avoid, and give the model a clear exit for when it genuinely needs more room.
The standing instruction in this kit borrows from two style books rather than from prompt folklore.
Anthropic's own prompting guidance adds a third ingredient. It advises telling Claude what to do instead of what not to do, and matching the style of your prompt to the output you want. A rambling, hedge-filled instruction tends to get you rambling, hedge-filled answers.
| Item | What it does | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|
| Answer-First Card (standing instruction) | Sets length and structure for every chat | Free; the instruction fields exist on free plans | Anyone using ChatGPT, Claude or Gemini daily | It is an instruction, not a hard limit |
| ChatGPT "Efficient" personality | Built-in concise tone setting | Free on all plans | ChatGPT users who want a one-click start | Does not govern drafts, code or emails you ask for |
| Claude Code "Concise" output style | Leads with the result, drops narration and recaps | Needs Claude Code, which needs a paid Claude plan or API billing (check current £ pricing) | Developers in the terminal | Requires Claude Code v2.1.237 or later |
| Trim prompt | Cuts one long reply down after the fact | Free | Rescuing a single answer | Treats the symptom, not the cause |
This is the standing instruction. Paste it once into the account-wide field:
Fill in the two bracketed lines, then paste:
Treat this as a standing rule for how you write replies to me.
Who I am: [YOUR_ROLE_AND_TYPICAL_WORK, e.g. "I run a small UK online shop and mostly ask about tools, copy and spreadsheets"].
Words I already know and don't need explained: [TERMS_YOU_KNOW].
How to write:
1. Open with the answer itself. If the question is yes/no, the first word is yes or no.
2. Keep each sentence to a single point, with a real subject and an active verb.
3. Default to short sentences and short replies. Run longer only when the problem has real steps or real trade-offs.
4. Say each thing once. Skip any summary of what you just said.
5. Begin with content, not with praise for the question, thanks or an offer to help.
6. Use plain technical English: the precise term when it matters, everyday words otherwise.
When to break the rule:
- If you need a fact from me to answer properly, ask one short question instead of guessing.
- Keep warnings, error details, costs and anything I need to act safely in full, even if it makes the reply longer.
- When I write "more", "explain" or "full version", drop the length limit for that reply.
Before sending, check: Is the answer in the first sentence? Can any sentence be deleted without losing information? Did I cut a warning I should have kept?
Fill in: your role and the terms you already know, so the model stops defining them.
On the web, select your profile icon, then Personalization, then use the dropdown next to Base style and tone and pick Efficient. OpenAI describes it as "concise and plain", giving a direct answer first and then brief reasoning or steps. On iOS and Android it sits under Settings, Personalization, Base style and tone.
Pair it with the Answer-First Card. OpenAI's help page is explicit that the personality does not control content you ask it to produce: an email draft, code or a CV will follow your request and context, not the Efficient tone. Your custom instructions and saved memories can also override or soften the personality.
If you work in Claude Code, run:
/output-style concise
Or run /config and choose Output style. To make it your default across projects, set "outputStyle": "Concise" in ~/.claude/settings.json. The name is case-sensitive in the settings file, and a mismatch such as concise silently gives you the Default style; the slash command ignores case.
Anthropic's documentation says the first sentence states the result, simple questions get one to three sentences, and the engineering work stays as thorough as Default. It also keeps error reports, failing test output, security warnings and confirmations for destructive actions at full length, which is the same escape hatch the Answer-First Card builds in by hand.
If you want your own version, a custom style is a Markdown file in ~/.claude/output-styles/. Add keep-coding-instructions: true in the frontmatter if you still want Claude coding normally, or it drops Claude Code's software engineering instructions.
For a single bloated reply you already have:
You are an editor who follows Zinsser and Williams: cut clutter, keep real subjects and active verbs, put the point first.
Rewrite your previous reply to me so that:
1. The first sentence is the answer or the decision.
2. Anything repeated, restated or announced is removed.
3. Every number, step, warning and caveat that affects what I do next is kept.
4. The result is no longer than [TARGET_LENGTH, e.g. "120 words" or "5 bullet points"].
Output only the rewritten reply. If cutting to that length would drop something I need to act safely, go over the limit and tell me in one line what you kept and why.
Fill in: the length you want.
Start with Item 1: it works across ChatGPT, Claude and Gemini. Add Item 2 on ChatGPT, Item 3 only if you already use Claude Code, and keep Item 4 for the reply that slips through.