AI prompting techniques, one technique per page
The twenty-five asks that take you from 'why is it ignoring me' to answers you can use.
A diagram of the mechanism, the classic mistake, and one thing to go try. That's a page.
AI prompting techniques, one technique per page
The twenty-five asks that take you from 'why is it ignoring me' to answers you can use.
Set in Space Grotesk, Inter and JetBrains Mono (SIL Open Font License).
Techniques checked against the official prompting guides: OpenAI's prompt engineering guide (platform.openai.com/docs/guides/prompt-engineering), Anthropic's prompt engineering documentation (docs.anthropic.com), and Google's Gemini prompt design strategies (ai.google.dev). The name 'chain of thought' is cited historically from Wei et al., Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, arXiv:2201.11903 (2022). No benchmark numbers are quoted; every technique is presented with the reasoning behind it, not as magic.
General craft guidance only. The assistant can be wrong; verify anything that matters.
Your purchase is for personal use only. You do not have redistribution rights: please do not share, resell, or republish this book or its pages.
© 2026 Steve Hodgkiss. All rights reserved. Personal use only; no redistribution rights.
Edition 1.0 · stevehodgkiss.net
Contents
The conversation
Before any technique: what the assistant sees when you type. Five pages on the context window as a desk, the two kinds of message, why it only knows what you give it, what a fresh chat forgets, and what really happens when the context overflows.
- 01The context window is a desk
- 02System message and user message
The context window is everything the assistant can see of your conversation: your messages, its replies, your pasted documents. It reads all of it, and only it, every turn.
The context window is a desk
Let's say you imagine it remembers your whole relationship. It reads one surface: the context window, every turn.
The mistake is treating it like an old colleague who was there for all of it.
Anthropic's docs picture it as a brilliant new employee: it knows only this conversation.
It only knows what's on the desk.
Open your longest active chat and scroll up until you find a rule you stated early that the newest replies have drifted from. That drift is the desk.
A system message (custom instructions, in chat tools) is read before every turn and carries more authority than an ordinary message. Your everyday asks sit below it, in the flow.
System message and user message
Let's say your rules get ignored three messages in. Pin them in the system message: read every turn, and it outranks the chat.
Buried in an ordinary message, a rule competes with everything after it.
Every major chat tool has the box: custom instructions, or the system field.
Rules pinned above the desk, not left lying on it.
Find your tool's custom-instructions box and pin your three standing rules: tone, spelling, and the one thing it always gets wrong.
Getting answers
The five moves that fix most mediocre answers. One clear ask, a role, a filled example, constraints stated up front, and options instead of a single take.
- 01The single clear ask
- 02Show a filled example: few-shot
- 03Constraints up front: format, length, audience
One ask carries one verb, one target and one shape. Vague asks get an average of everything the assistant guesses you might have meant.
The single clear ask
Let's say you type "help with this" and get mush back. The assistant averaged every reading of your ask.
"Help with this" could mean four jobs. It picked one at random.
"Cut the attached to 3 bullets for my manager" has verb, material, shape, reader.
One ask, one verb, one shape.
Take today's mushiest ask and rewrite it as verb plus material plus shape plus reader. Send both versions, keep the better one.
A filled example (few-shot) shows the exact pattern: the fields, the labels, the casing. The assistant copies the shape it saw far more reliably than a shape you described.
Show a filled example: few-shot
Let's say you describe a format and still don't get it. Show one filled example instead: in, out, exactly as you want it.
Describing a format in prose leaves the details, casing and column order, to guesswork.
Two or three examples, varied, beat a paragraph. All three vendors say so.
Don't describe the form. Hand over a filled one.
Take any recurring task and write two filled examples of input and output, then hand the assistant a third input. Keep the examples in a note for reuse.
Format, length and audience are boundaries, set before the work starts. Stated up front, the first draft arrives inside the lines.
Constraints up front: format, length, audience
Let's say every answer needs the same trim afterwards. You set the boundaries after; set them before.
Format guessed, length open-ended, reader imagined: that's three rewrites you do yourself.
One line: table, five rows, for the board. The first draft fits.
Boundaries before work, not apologies after.
Add a standing constraint line to your three most frequent asks: exact format, hard length limit, named reader. Watch the trim work disappear.
Better thinking
Answers get better when the thinking shows. Five pages on asking for the working, self-critique, iteration, decomposition, and defining good before you ask.
- 01Show your working: chain of thought
- 02Iterate: the first answer is draft one
Ask for the steps before the conclusion. Visible working turns a wrong answer into a findable slip, and a right one into a checkable one.
Show your working: chain of thought
Let's say it answers 41 and you can't tell if that's right. Ask for the working before the answer: steps, then total.
A bare number you must trust is a number you cannot correct.
Named in a 2022 paper, now built in: your tool likely has a thinking mode. For big sums, still say "show the steps".
Work on the page, where you can check it.
Take one real estimate you'd normally accept blind and ask for it twice: once bare, once with numbered steps shown before the answer. Compare.
The first answer is draft one, not the verdict. Each revision of the ask, informed by reading the draft, climbs the same chat.
Iterate: the first answer is draft one
Let's say one shot misses and you conclude the tool can't do it. It rarely can, first pass: read the miss, revise the ask.
Retyping the same ask louder, or giving up after one try.
Second asks are cheap: "shorter, second person, lead with the deadline".
The first answer is draft one.
Take one answer you were disappointed by this week and give it two honest revision passes instead. Note what each pass actually fixed.
The assistant can produce a fluent sentence about something that is not true. Verify what carries weight: numbers, names, quotes, anything with consequences.
Verify facts: trust but check
Let's say a number in its draft looked right, so it stayed. Fluency is free; correctness isn't. Check the load-bearing lines.
Reading it over and calling that verification: plausibility is not proof.
Paste the source in, ask "from this text only", and request a source per claim.
Trust the draft. Check the facts.
Take one assistant-produced number going into real work this week. Trace it to a source you'd cite. If you can't, cut it or fix it.