AI Guides › Workbench
By Nigel Guy · 7 min read
You save a list of punchy two-word prompts, type "zoom out" under a draft, and get something that looks thoughtful. The next time, the same two words give you something completely different, because Claude had to guess what you meant both times. The guessing is the problem, and it shows up exactly when you rely on it.
The rule: two words make a good handle, but they're a poor instruction. Write down once what each handle should make Claude do, then load that written version every time you use the two words.
A good two-word prompt changes the approach, not the wording: "Argue against" asks for an opponent instead of an editor, and "Cut half" sets a hard limit instead of a vague wish for brevity.
What the two words can't do is carry context. Anthropic's own prompting guidance makes the test simple: if a colleague with no background on the task would be confused by your instruction, Claude will be too. The fix is the expansion card: a two-word handle, plus two or three sentences saying exactly what it triggers and what the output looks like.
Six families of three. The expansion is what Claude actually follows.
| Family | Handle | Expansion: what Claude does | Catch |
|---|---|---|---|
| Diagnose | Find assumptions | Lists unstated assumptions, marks each safe, shaky or unknown, and names the riskiest | Only sees what's on the page |
| Diagnose | Spot gaps | Names missing steps, numbers, owners or dates a careful reader would expect | Can flag trivia. Ask for a ranking |
| Diagnose | Name trade-offs | For each choice, states what I give up and who pays | Sounds balanced even when one side clearly wins |
| Challenge | Argue against | Makes the strongest unsoftened case against my position | You have to actually read it |
| Challenge | Steelman opposite | States the opposing view as its best advocate would, then says where mine still wins, if anywhere | Can over-credit weak positions |
| Challenge | Pre-mortem this | Assumes the plan failed six months on; gives three likely reasons, each with a warning sign | Gives likely stories, not probabilities |
| Shrink | Cut half | Rewrites at half the length, keeps every fact and decision, lists what it dropped | Drops caveats. Check the list |
| Shrink | One sentence | States the point in 25 words or fewer, then the one detail that leaves out | Loses nuance by design |
| Shrink | Simplify further | Shorter words, one idea per sentence, no unglossed jargon | Can become patronising. Name your reader |
| Reframe | Zoom out | Names the larger problem and whether I'm solving the right one | Can wander. Cap it at a paragraph |
| Reframe | Invert it | Lists what would guarantee failure, then turns each into something to avoid | Works best on goals, not facts |
| Reframe | First principles | Sorts the problem into fact, convention and preference, then rebuilds from fact only | "Definitely true" still needs your check |
| Audience | Beginner version | Re-explains for a newcomer with one concrete example | Accuracy can slip in the simplifying |
| Audience | Sceptic version | Leads with evidence and names limits up front | Can't supply evidence you didn't give it |
| Audience | Busy-reader version | The ask in line one, then three supporting points | Hides context the reader may need |
| Decide | Rank options | Scores options against my criteria (asks for them if missing) and shows the scoring | Garbage criteria give a garbage ranking |
| Decide | Pick one | Commits to one recommendation, its main reason, and what would change its mind | Sounds sure. It isn't a guarantee |
| Decide | Next step | Names the smallest action I can take today and what it will tell me | Ignores bigger moves on purpose |
Every card costs £0 and works in any chat assistant.
Paste this at the top of a chat, or put it in a Claude Project's instructions so it applies to every chat in that Project. Projects are available on the free plan (up to five projects) and on paid plans.
You're working with me as a thinking partner. I'll sometimes type a two-word handle instead of a full request. Treat each handle as shorthand for the expansion below and apply it to [WHAT_TO_APPLY_IT_TO: "your last answer" or "the text I paste"].
[PASTE THE HANDLE AND EXPANSION ROWS YOU WANT TO USE, ONE PER LINE]
Rules:
1. If I use a handle that isn't on the list, don't guess. Ask me what I want it to do.
2. If the handle needs something you don't have, such as criteria for "Rank options" or a reader for "Simplify further", ask for it in one question before you start.
3. Start each reply with the handle name in bold, so I can see which expansion you applied.
4. Don't invent facts, figures or sources to fill a gap. Mark anything uncertain as uncertain.
Before replying, check that your output matches the expansion's stated format (count, length or structure) and fix it if it doesn't.
Fill in: what the handles act on, and the six or so rows you'll actually use.
"Pre-mortem this: a paid webinar for 40 people" beats the bare handle. The handle picks the move; the line supplies the real constraints.
"Argue against" then "Pick one" is a sound pair. Three or more blur, and you can't tell which card shaped the answer.
Claude Code is included in paid Claude plans at time of writing. A weekly card can become a real slash command: create ~/.claude/skills/pre-mortem/SKILL.md for a personal command available in all your projects:
---
name: pre-mortem
description: Run a pre-mortem on a plan the user names
disable-model-invocation: true
---
Treat this plan as having failed six months from now: $ARGUMENTS
Write the three most likely reasons it failed. For each one, give an early warning sign I could spot within the first two weeks, and one cheap way to reduce the risk. If the plan description is too thin to judge, ask me for the missing detail instead of guessing. Keep the whole answer under 250 words.
Fill in: nothing. Type /pre-mortem [YOUR PLAN IN ONE LINE]; that text replaces $ARGUMENTS, and disable-model-invocation: true stops Claude running it unprompted.
Use a keep-or-cut column. Add one column to your copy of the table and mark it after each real use:
| Mark | Meaning | What to do |
|---|---|---|
| K | Changed my decision or saved a rewrite | Keep. After three Ks, consider making it a command |
| E | Useful, but the output drifted | Edit the expansion, not the handle |
| C | Didn't change anything | Cut it after two Cs |
Hypothetical example: "Spot gaps" scores C, C, E on three proposals because it flags formatting. You narrow the expansion to "missing decisions, owners and dates only" and the next two runs score K. Same handle, better instruction.
To build a card of your own:
I want a reusable two-word prompt for [THE THINKING STEP I KEEP SKIPPING], used on [TYPE OF WORK]. Propose three two-word handles. For the best one, write an expansion of no more than three sentences that says what you'll do, the output format, and one thing you won't do. Then show it working on this sample: [SHORT SAMPLE]. If the step I've described is too vague to turn into a card, ask me one clarifying question instead.
Fill in: the step you skip, the kind of work, and a short real sample.