Think first, prompt second.
AI is a thinking partner, not a thinking replacement. Sketch the answer in your head before you ask — you'll spot bad outputs faster.
field guide · 2026 editionUse AI like
Use AI like
a curioushuman.
A friendly, no-fluff guide to getting more out of AI without losing your judgment, your privacy, or your soul. Smart prompts. Honest limits. Habits that age well.
Plain English
Tool agnostic
Updated regularly
prompt.txt
Role: You're a careful editor. Goal: Tighten this paragraph without losing my voice. Constraints: - Keep it under 90 words - No new claims - British English Audience: curious beginners Format: a single paragraph, then 3 bullet edits with a one-line reason each.
v3 · iterated shipped
"good prompts are edited, not written."
humans review · always
be specificverify the factscite your sourcesrespect privacystay curioushuman in the loopcredit your inputsiterate, don't acceptbe specificverify the factscite your sourcesrespect privacystay curioushuman in the loopcredit your inputsiterate, don't accept
the six principles
Six rules. Memorise them.
If you only take one thing from this page, take these. Everything else is just tactics.
Verify like a journalist.
Treat every claim as a draft. Cross-check names, numbers, quotes, and citations against primary sources before you ship them.
Privacy is a default.
Don't paste secrets, client data, or anything you wouldn't email to a stranger. Anonymise inputs. Read the data policy.
Credit and consent.
Disclose AI assistance when it matters. Don't impersonate real people. Don't train on work you don't have rights to.
Keep your taste.
If you can't tell good from mid, AI will average you. Read widely, write often, and stay opinionated about what 'great' means.
Automate the boring.
Use AI for friction, not flair. Summaries, drafts, refactors, translations — the stuff that drains energy from the real work.
the prompting playbook
Five moves that make every prompt better.
Not magic words. Not "ultimate" templates. Just the moves that working pros run on autopilot.
- Step 01
Frame the job.
Tell the model who it is, who it's for, and what 'done' looks like. Vague in, vague out.
- Role
- Audience
- Definition of done
- Step 02
Give it raw material.
Paste the source text, the data, the brand voice notes — context beats cleverness every time.
- Examples
- Constraints
- Reference style
- Step 03
Ask for structure.
Specify the shape of the answer: headings, bullet count, tone, word limit. Models love rails.
- Format
- Length
- Tone
- Step 04
Iterate ruthlessly.
First drafts are negotiations. Tighten one knob at a time: shorter, punchier, more concrete, less hype.
- Critique
- Constrain
- Re-roll
- Step 05
Verify, then ship.
Check facts, names, code, and citations. Have a human sign off on anything with real-world stakes.
- Fact-check
- Test
- Human review
Anatomy of a great prompt
A simple skeleton you can paste at the top of any chat. Fill in the blanks, delete the rest, and you'll out-prompt 90% of users.
RoleGoalInputsConstraintsAudienceFormatExamples
# Role
You are a senior {role} who values clarity over cleverness.
# Goal
Help me {specific outcome}, optimised for {who reads it}.
# Inputs
"""
{paste the raw material here}
"""
# Constraints
- Max {N} words
- Voice: {warm / dry / playful}
- Avoid: jargon, hype, em-dashes
- Cite sources you actually used
# Format
1) A {short / structured} answer
2) A 3-bullet "what I'd change next"
3) Open questions you'd ask me the short list
Do this. Don't do that.
Print it, pin it, send it to the group chat. The boring rules are the ones that save your week.
Do
- Treat AI output as a first draft.
- Anonymise data before pasting.
- Cite the AI when it shaped the work.
- Keep a human in the loop for decisions.
- Save prompts that work — they're assets.
- Compare two models on hard tasks.
Don't
- Paste passwords, keys, or client data.
- Trust citations without checking them.
- Ship code you don't understand.
- Use AI to impersonate real people.
- Hand off judgement on health, money, or law.
- Believe confident tone equals correctness.
notes to self
Stick these on the monitor.
"AI doesn't replace expertise — it amplifies whatever judgement you bring to it."
"If your prompt fits in a tweet, your output will too."
"The model is confident. That is not the same as correct."
common questions
Things people actually ask.
Should I tell people I used AI?+
If it materially shaped the work — yes. A short note ('drafted with AI, edited by me') is usually enough. For journalism, academic work, or client deliverables, follow the stricter rule of the room.
Is it OK to paste my company's data into a chatbot?+
Default to no. Use a tool with a clear data-handling policy, an enterprise plan, or a self-hosted model. Anonymise where you can, and never paste credentials, customer PII, or unreleased financials.
How do I keep my own thinking sharp?+
Write the first draft yourself, then ask AI to critique it. Set 'AI-free' hours for hard problems. Read primary sources. The skill you'll miss most is the one you stopped practising.
Are AI-generated citations safe to trust?+
No. Models can confidently invent papers, court cases, and quotes that look real. Always click through and confirm the source exists and says what you're claiming it says.
What about creative work — is using AI cheating?+
Tools aren't cheating; misrepresentation is. Be honest about your process, respect the rights of artists and writers whose work you draw on, and keep your taste in the loop.
How do I know which model to use?+
Pick by task, not brand. Try two models on the same prompt; keep the one whose answer you'd actually ship. Re-evaluate every few months — the leaderboard shifts fast.
one last thingThe best prompt is the one
The best prompt is the one
you keep editing.
Save this page. Send it to a teammate. Come back when AI does something weird — it will, and you'll want the rules close.