Field Notes
The
Developer's
Prompting
Handbook
How I make LLM output predictable enough to put in production.
Kateryna Skoryna
Software Developer · 2026
Software Developer · 2026
Turn the page →
I argued with an AI for an hour.
Turns out we were both wrong, but it was more confident.
— inside cover —Turns out we were both wrong, but it was more confident.
01 Foundations
The anatomy of a good prompt
Before any of my own tricks, this is the foundation nearly every guide agrees on. A strong prompt is built from a few named parts. You don't need all of them — the task is the only non‑negotiable — but the more you make explicit, the less the model has to guess.
PersonaTell it who to be. Sets vocabulary, depth, tone. On frontier models this shifts tone & format more than raw accuracy.
TaskThe one thing it must do. The only required part.
ContextPurpose, audience, constraints. Where most weak prompts fail.
FormatExactly how you want output: JSON, bullets, a table, a length.
ExamplesOne or two input→output samples. Showing beats describing.
ReasoningFor anything complex, ask it to think step by step before answering.
Ordering — no single correct order, but put the context/data first and the instruction last, so the model acts after it has read everything.
✦Worked example
Persona: Senior full‑stack engineer (React + Express), performance‑focused.
Context: Page loads slowly — the index bundle got bloated with libs not needed on first paint.
Task: Refactor imports in
<code> to trim the initial bundle (lazy‑load / code‑split), without changing public props.Format: Return only corrected code in
<answer>; one‑line comment noting what moved out.02 My rules
My rules — and the failures that taught them
The fundamentals get a decent prompt; these habits make output reliable enough to build on. My mindset: a prompt is production code — versioned, tested, revised against real failures.
What I do
Set it up right
- Align on the task first. I clarify until the model’s understanding matches mine.
- Persist stable rules in a system prompt / Gem / Project — and tell it who you are.
- Reuse — build entities. Gems ≈ Projects ≈ custom GPTs; a meta‑prompt drafts new ones.
Be precise
- Exact length ("3 bullets, ≤12 words"), never "short".
- Name things by name or number — it doesn't see your screen.
- Isolate inputs — data ≠ commands.
Work it like code
- Temperature = variety, not quality.
- Fix the prompt & regenerate — don't argue with the output.
- Small rule set, no contradictions.
Where I learned it
→ Temp
1.2 invented a false fact → dropped to 0.9 plus an explicit FACTUAL ACCURACY rule. (defense in depth)→ A free‑text field let users inject anything → a narrow CONTENT MODERATION override plus a permanent regression test.
→ "Advanced" meant nothing concrete → one example question per level. (a one‑shot anchor beats an adjective)
None of these came from a guide. I read real output, traced each failure to a gap, closed it, and wrote a test so it stays closed. That's the whole skill.
03 Modalities
The same thinking, across three kinds of output
Prompting isn't just chat. Here's where I apply all of this — and the one tip that matters most for each.
CodeClaude / Gemini + Genkit
Use case: Generating & grading dev quizzes in quizdom — typed JSON (Zod‑validated) plus a second "judge" model that scores the first.
Tip Force structured output against a schema, so a mismatch is a caught error at the call site — not a crash three components downstream.
PresentationsClaude
Use case: Engineering design reviews and technical architecture explainers.
Tip Make it interview me first — end with "ask me any clarifying questions before you generate anything." Then outline‑first, one idea per slide, an exact count ("8 slides").
InfographicsNotebookLM (Gemini)
Use case: Turning a spec or doc into a clean visual for non‑technical people.
Tip It's source‑grounded (clean sources in = good visual out) & renders via Nano Banana Pro. Write the image prompt like a director's brief — purpose → subject → style → composition → lighting → palette → format — and say what you want, not what you don't ("an empty street," never "no cars").
04 My Gems
The specialists I built once and reuse
Gem Architect
Purpose — writes first‑draft instructions for other Gems (meta‑prompting, no blank page).
"You are a prompt engineer. When I describe an assistant, ask clarifying questions, then output a full Gem set: role, rules, tone, output format, two example interactions."
Benefit — "I want a Gem that does X" → ready‑to‑paste config in one step.
Personal Nutritionist
Purpose — personalized nutrition without restating my situation every time.
"You are my nutritionist. My goal is to keep myself fit. Practical, specific suggestions; ask before assuming; realistic for a busy schedule."
Benefit — consistent, tailored answers — the context lives in the Gem.
Senior Full‑Stack Architect
Purpose — a sounding board for design decisions without re‑explaining my stack.
"…ask clarifying questions first, then propose 2–3 approaches with tradeoffs before recommending one. Flag scaling risks & over‑engineering. Stack: React/Next + Node.js."
Benefit — a second opinion that catches blind spots & over‑engineering early.
The pattern: role + rules + context + output format — drafted fast with the Architect, then refined as you use it.
05 Cheat sheet
The whole handbook on one page
Choose the model for the task. I first decide whether I need a quick fix or a complex solution. Then I select the model accordingly: fast and lightweight for simple work, stronger reasoning for ambiguity, trade-offs, or multiple steps.
Covered the basics — persona, task, context, format?
Gave context and the why, not just the task?
Stable rules live in a system prompt / Gem / Project?
Every length an exact number, never "short"?
Pointed to things by explicit name or number?
Untrusted data isolated in its own tag, marked data‑not‑commands?
Temperature matched to the task (hot = variety, cold = repeatable)?
Output schema‑validated where it feeds code?
When wrong, editing the prompt & regenerating — not arguing?
Each rule came from a real failure — and no rules contradict?
"Every failure teaches me twice: I teach the model, and while teaching it, I teach myself. It’s a continuous loop — integrate, deliver, learn, repeat."
I asked AI to write clean code.
It gave me a blank file.
It gave me a blank file.
Get in touch
Wondering which AI tool fits which job? Read my journal post →
↳ Say hi. No prompt engineering required.
Prompting Handbook