AI Guides › Workbench
By Nigel Guy · 8 min read
When one person's AI output looks like a small team's, people assume a better tool or a secret prompt. Usually it's neither: it's a few habits that stop the model guessing, plus not retyping the same setup every morning. The default is the opposite: a fresh chat, a bare question, a generic answer, and a verdict that the tool is average.
The rule: three habits change the answer, two habits buy back the time. Learn them in that order, because saving a bad setup only repeats it faster.
| Habit | What it does | Cost at time of writing | Best for | The catch |
|---|---|---|---|---|
| 1. Context Card | Tells the model who you are, who it's for and what you've already tried | Free in any chat tool | Any question where "it depends" is the honest answer | Stale cards give confidently wrong advice |
| 2. Decision Page | Turns a long document into one page aimed at a single decision | Free; long files work better on paid plans | Contracts, proposals, reports, supplier quotes | It summarises what's there, not what's true |
| 3. Objection Pass | Makes the model argue against its own recommendation | Free | Anything you're about to sign, send or spend on | Can manufacture doubt if you ask it to "find problems" |
| 4. Standing Project | Stores your Context Card and files so every chat starts briefed | Claude: free accounts get up to five projects. ChatGPT: Projects on all plans, 5 files per project on Free | Ongoing clients, products or roles | One project per topic, or the context muddles |
| 5. Saved Skill | Packages a task you repeat into a reusable instruction set | Claude Skills listed for Free, Pro, Max, Team and Enterprise | Weekly reports, standard replies, recurring checks | Needs code execution switched on; you maintain it |
Nothing here needs a paid plan to start. I couldn't confirm current UK pound prices for Claude Pro or ChatGPT Plus on 4 October 2026, so check the pricing pages below before you pay.
Most weak answers are sensible replies to a question the model had to fill in itself. Write this once, keep it in a note, and paste it at the top of any conversation that matters.
You are advising me on my work. Before I ask anything, read this background and use it for every answer in this conversation.
About me: [YOUR_ROLE], [YOUR_BUSINESS_OR_TEAM], doing this for [HOW_LONG].
Current focus: [WHAT_YOU_ARE_WORKING_ON_THIS_MONTH].
The people this is for: [AUDIENCE_OR_CUSTOMER].
What matters most right now, in priority order: [E.G. TIME, THEN MONEY, THEN QUALITY].
Already tried, so please don't suggest it again: [WHAT_YOU_TRIED_AND_WHAT_HAPPENED].
How to reply: [E.G. BLUNT AND SHORT / DETAILED WITH REASONING].
Rules for this conversation:
- If an answer depends on something missing from this background, ask me one short question rather than assuming.
- If my priorities pull against each other on a question, say which one you favoured and why.
- Before each answer, check it doesn't repeat anything on my "already tried" list.
Reply only with "Understood" and any gap in this background you'd want filled first.
Fill in every bracket; a blank one gets invented. Update "current focus" when your month changes.
A summary written for nobody is reading theatre: shorter, but no closer to a choice. This ties it to one decision.
You are a careful analyst preparing a one-page brief for a busy decision-maker.
The material: [PASTE_TEXT_OR_ATTACH_FILE]
The decision I have to make: [THE_DECISION, E.G. "SIGN THIS SUPPLIER CONTRACT OR NOT"]
My deadline and constraints: [DEADLINE / BUDGET / NON-NEGOTIABLES]
Steps:
1. Read it all, then write exactly these four headed sections, one page maximum:
- The gist: five lines, plain English, no jargon.
- What matters for this decision: three points, each with the section or page it came from.
- What's missing: questions the document leaves open that would change my decision.
- Recommendation: one line, followed by the strongest case against it.
2. Mark anything inferred rather than stated with [INFERRED].
Rules:
- Quote figures and dates exactly as written; do not round or convert.
- If the document doesn't cover my decision at all, say so and stop.
- If the decision or deadline is missing above, ask before starting.
Before replying, check that every point under "What matters" has a location reference.
Fill in the material, decision and constraints. The location references let you check the three claims that matter without rereading forty pages.
Models tend to agree with how a question is framed. This turns that round once before you act, straight after Habit 2 or on any plan of your own.
Act as a sceptical but fair reviewer whose job is to stop me making an avoidable mistake.
What I'm about to do: [YOUR_PLAN_OR_RECOMMENDATION]
Why I think it's right: [YOUR_REASONS]
What it costs if I'm wrong: [MONEY / TIME / REPUTATION AT STAKE]
1. Give the single strongest argument against this plan, in under 100 words.
2. Name the one fact that, if it turned out differently, would most likely flip the decision, and how I could check it this week.
3. Say plainly whether the objection is serious or minor. If you can't find a serious one, say that rather than inventing one.
Do not list every possible risk.
Fill in the plan, reasons and downside. Ask for "all the risks" instead and you get a padded list that makes every plan look shaky.
Once the Context Card works, stop pasting it: store it, with your files, in a project.
In Claude: open Projects, create a project, add your Context Card to the project instructions and upload the reference files you keep reusing. Anthropic's help centre says projects are available to all users and free accounts can create up to five; its personalisation page also describes project instructions as a paid-plan feature, so if the instructions box is missing on a free account, that's why. Account-wide instructions live under your initials (lower left) > Settings > Instructions for Claude.
In ChatGPT: create a project from the sidebar, then open the ••• menu and choose Project settings to add instructions. File limits per project are 5 on Free, 25 on Go and Plus, and 40 on Pro, Business, Enterprise and Edu. Projects can use default memory or project-only memory; pick project-only for client work so details don't leak between projects. Account-wide custom instructions are under Settings > Personalization (1,500 characters on Free and Go, 5,000 on paid tiers).
Use one project per client, product or role; a single "Work" project is just a slower blank chat.
For a task you run weekly with the same steps, write the steps once as a Skill; Claude loads it when a task matches its description.
name (64 characters maximum) and a description (200 characters maximum) saying when Claude should use it.To draft the SKILL.md body, run this in a normal chat:
You are helping me turn a task I repeat into written instructions another assistant can follow without me.
The task: [WHAT_YOU_DO_EACH_TIME]
How often: [FREQUENCY]
Inputs I start with: [FILES_OR_INFORMATION]
What a finished result looks like: [PASTE_A_GOOD_PAST_EXAMPLE]
Mistakes I've made or seen before: [KNOWN_PITFALLS]
Write numbered steps in the order I'd do them, a description of the output format, and a final checklist to run before handing over. Ask me about any step you'd otherwise have to guess. Keep it under 500 words.
ChatGPT users without Skills can paste the same steps into a dedicated project's instructions; it does the same job.
Start with Habit 1 today; it improves every other habit. Add Habits 2 and 3 at your next real decision with a document attached. Move to Habit 4 once your card has survived a couple of weeks unchanged, and build a Skill only for a task you've done the same way at least three times. Saving something unsettled just means editing it in two places.