Prompting that actually works

Imagine handing this task to a smart freelancer who started this morning. Brilliant, fast, zero context. Whatever they'd need to nail it first time is exactly what belongs in your prompt.

That one test dissolves most prompting advice into common sense. Would the freelancer need to know who it's for? Say so. Whether there's a house style? Attach it. What "done" looks like? Describe it. Would they benefit from an example of previous work? Paste one.

And the reverse holds: anything they wouldn't need, cut. Long wind-ups, apologies, "act as a world-class expert" incantations. Claude doesn't need warming up; it needs ingredients.

Five ingredients. Not all are needed every time, but when an answer misses, one of these is missing.

Before and after

Nothing clever happened there. The second prompt just answers the questions the freelancer would have asked.

Adjectives are lossy. "Professional but warm" means fifty different things; an example of it means exactly one. The single highest-value prompting technique is pasting an example and saying "like this".

• Voice: two or three things you've written and like, then "match how I write in these".
• Structure: a past report you're proud of, then "same structure and rigour, new topic".
• Judgement: one example you love and one you hate, and say which is which. Contrast teaches faster than praise.
• Format: a filled-in sample row, so it extends a pattern instead of inventing one.

Here are two emails I've written that sound like me,
and one draft that doesn't. Rewrite the draft so it
sounds like the first two: same warmth, same sentence
length, same level of directness. Don't smooth out
my quirks.

Unspecified format is how you get eight paragraphs when you needed a table. Decide the shape of "done" before you send, and say it.

• Structures: "a table with these exact columns", "numbered steps", "a one-line answer, then your reasoning".
• Lengths: word counts work. So do sharper units: "three bullet points", "one paragraph I can paste into Slack", "a single sentence for the subject line".
• Multiples: "give me three versions: safe, bold, weird" is the cheapest way to explore. You'll usually take the middle one and steal a phrase from the third.
• Templates: paste a skeleton with headings and let it fill it in. Total control, zero surprise.

Claude's most expensive default: faced with ambiguity, it guesses helpfully rather than interrogating you. One line flips that.

Before you write anything: ask me the 5 questions whose
answers would most improve the result. Then wait for
my answers.

Use it for anything with stakes or fuzz: proposals, plans, difficult messages, briefs you haven't fully thought through. The questions routinely surface things you hadn't decided, which is the point: the interview improves you, not just the prompt.

It scales up beautifully: "interview me one question at a time until you have enough to write my CV properly" turns a blank-page nightmare into a conversation.

Forget "act as a world-class copywriter"; costume jewellery. Framing works when it changes the job, not the badge.

• Set the stance: "be my editor, not my cheerleader; the draft survives only if it deserves to". Claude is agreeable by default, and this is the off switch.
• Set the adversary: "you're the procurement director reading this proposal, looking for reasons to say no. List them in order of danger."
• Set the bar: "would this survive a partner review at a good consultancy? Mark it against that standard and tell me the gaps."
• Argue both sides: "make the strongest case for, then the strongest case against, then tell me which you find more honest."

Draft one is a probe, not a product. The skill is in what you do next, and there are two moves most people never use.

• Steer with contrast, not adjectives. "Paragraph two is the good one; make the rest sound like that" beats "make it better" every time.
• Edit your message, don't pile on corrections. When a conversation has drifted through four rounds of "no, more like…", stop patching. Go back, edit your original message to include everything you've learned, and let it regenerate clean. One strong brief beats five corrections layered on a weak one.
• Fork on purpose. Same brief, then "now a completely different approach". Two genuinely distinct directions teach you what you actually want faster than ten refinements of one.
• Know when the chat is spent. If it keeps reaching back to earlier, abandoned instructions, the context is polluted. Start fresh, carry over only the good final brief.

Hard problems deserve two things easy problems don't: a thinking model, and permission to work before answering.

• Split plan from execution. "First give me the outline and the argument. Don't write it yet." Approve the skeleton, then let it write. Ten times cheaper than fixing structure after the fact.
• Ask for reasoning on judgement calls. "Recommendation first, then the three factors that would change your mind." The factors are often worth more than the recommendation.
• Use extended thinking for genuinely hard work: intricate analysis, tricky trade-offs, anything multi-step. And match the model to the stakes (guide 01, section 03).
• Don't pay the tax on easy work. Thinking time on a subject line is waste. Reserve the heavyweight treatment for problems that push back.

Deep Research is prompting with the stakes raised: one brief, many minutes of autonomous work, and a report you'll be tempted to trust because it looks thorough. Brief it like you'd brief an analyst, and check it like you'd check one.

The brief

• Scope in, scope out. "UK market only, last three years, exclude anything about the US regulatory fight" saves the run from drowning you in adjacent material.
• Name the decision it feeds. "This is to decide whether we enter this market" makes the report argue towards something instead of surveying everything.
• State source preferences. "Prefer regulator publications, filings and trade press; treat vendor blogs as claims, not facts."
• Fix the output shape like any artifact: sections you want, a comparison table, an executive summary with the three load-bearing numbers.

Research [question]. Scope: [markets/period]; exclude
[out of scope]. This feeds a decision about [the
decision]. Prefer [source types]; treat [weak source
types] as claims needing corroboration. Output:
1) one-page summary with the key numbers,
2) the evidence by theme with citations,
3) a section titled 'What would change this conclusion',
4) a list of claims that rest on a single source.

Checking without re-doing the work

• Spot-check only the load-bearing claims: the numbers the conclusion rests on, direct quotes, and anything surprising. Click those citations; ignore the rest.
• The single-source list is your risk register. That's why the brief demands it: anything important resting on one source gets a second look or a second run.
• Follow up narrower, not wider. A good first run's job is to teach you the right second question. "Go deeper on finding 3; the rest is settled" beats re-running the whole thing.

Ten seconds, one table. Find the symptom, add the missing ingredient, go again. Re-rolling the same prompt and hoping is not a strategy.

Six workhorses. Copy, adapt, keep the ones that earn a place in your muscle memory.

The critique ladder

Review this draft in three passes:
1) What works, keep it short.
2) What fails: logic, evidence, structure, in order
   of damage.
3) The single change with the biggest payoff.
Then, and only then, show me the improved version.

Extract to a table

From the attached, extract every [commitment / date /
cost / decision] into a table: item, who, when, source
quote. If something is implied but not stated, include
it and mark it 'implied'.

The audience translator

Rewrite this for [the board / a new starter / an angry
customer]. Keep every fact. Change what's emphasised,
what's explained, and what's left out, based on what
this reader needs and fears.

Meeting notes to actions

From these notes:
1) decisions made,
2) actions as a table (owner, task, deadline, blocked-by),
3) open questions nobody answered,
4) anything that sounded like a commitment someone
   will forget.
Flag items where the owner is ambiguous rather than
guessing.

The options spread

Give me three versions:
1) the safe one I'm probably imagining,
2) a bolder one you think is actually better,
3) a wildcard.
One line each on the trade-off.

The pre-mortem

It's six months from now and this plan failed. Write
the honest post-mortem: what went wrong, in order of
likelihood. Then tell me which two failures I could
cheaply insure against today.

When a pattern earns its keep, stop retyping it: save it into preferences, a Project's instructions, or a Skill so it's always on tap. That's guide 03's territory.