Module 04 · Reverse‑prompting & virtual colleagues

Stop asking for answers. Ask for the questions.

You don't know what you haven't specified. The model does, in about four seconds, if you let it ask. Then borrow a colleague's brain to ask from the angle you'd never think of. Five of them are on call, always.

~12 minreverse‑prompt makerpick‑the‑colleague gamequiz · pass ≥ 80%

THE BRUTAL TRUTHModule 02 showed that under‑specifying costs an hour. Here is the uncomfortable part: you cannot list what you forgot. Nobody can. Reverse‑prompting is the only reliable fix, it takes sixty seconds, and almost nobody at THG does it.

Reverse‑prompting: the sixty‑second insurance policy

Instead of "write X", you say: "Before you write anything, ask me the five questions a <role> would need answered to do this properly. Then wait." The model lists what it would otherwise have guessed. You answer in one message. Then it writes, and it writes right.

Forward prompt, straight to output"Write the Q4 roadmap update for operators."

→ 700 words. Wrong quarter boundaries. Mentions a
  feature Legal hasn't cleared. Two rounds to fix.
  You never knew those were open questions.

Reverse prompt, questions first"Before writing the Q4 roadmap update for operators, ask me the 5 questions a senior account manager would need answered. Wait for my answers."

→ Q1: Which markets?  Q2: Which features are
  cleared to mention?  Q3: Fiscal or calendar Q4?
  Q4: Tone vs last update?  Q5: One CTA or none?
  You answer in 90 seconds. Output is right first time.

The questions the model asks are a list of everything you were about to leave to chance. Read them as a diagnosis of your own prompt. It stings once, then you start writing better forward prompts too.

Watch it: reverse‑prompt versus straight‑to‑output

two ways to ask for the same migration plan

press ▶ to replay both

Five virtual colleagues. Five lenses. Always on call.

Reverse‑prompting gets sharper when you tell the model whose questions to ask. THG has five virtual colleagues, each a persona with one signature question that catches a whole category of mistakes. Full profiles live at virtual‑colleagues‑thg.pages.dev. Here is what you need to pick the right one.

Picking the wrong colleague is not harmless. Ask Charlie to review a database migration and you get a beautifully worded disaster. Ask Leo to review a launch post and you get a bulleted list nobody will read. Lens first, then prompt.

Reverse‑prompt maker: pick a colleague, type your task

This builds the exact reverse prompt you should paste into your AI tool. Use it until you can write it from memory.

Build your reverse prompt

Your task, one sentence

Paste this

Game: pick the colleague

A real THG task. Whose lens do you reverse‑prompt with first? One pick. Wrong goes red and shakes. Right goes green, and you get the exact question that colleague would have asked.

Ten scenarios. Aim for 8 or better. Some tasks have a second defensible pick; it counts, but you'll be told who should have gone first.

Whose lens first?

Score 0/0Streak 0

Quiz: prove it

Five questions. Four right to pass.

Module 4 quiz

THE ONE THING TO REMEMBERWhen it matters, ask for the questions before the answer, and name whose questions: John for leverage, Charlie for brand, Maya for the user, Leo for what breaks, Sophie for who it lands on. Sixty seconds. Every time.

← M03 · Anatomy of a good prompt Next: M05 · Subagents & second opinions →