A handbook on the underrated craft layer. Five components, six anti-patterns, and the prompts professionals actually use. No hacks, no incantations — just the principles that make outputs reliably good.
Prompting is not a vibe. It is the practice of writing a brief that an extremely capable, slightly forgetful intern can execute reliably. Once you see prompts as briefs, everything that follows is just being a good manager — in writing.
If you wouldn't hand this prompt to a thoughtful new hire and expect a good first draft, the model won't give you one either. Fix the brief.
Every reliable prompt has five sections. Skipping one is the single most common cause of "the model didn't do what I wanted." Get the anatomy right and 80% of prompt problems disappear.
01 · Role. Who is the AI playing? "You are a careful copy editor." "You are a senior infrastructure engineer reviewing this for risk." The role sets vocabulary, depth, and stance.
02 · Context. What does the model need to know? The brand voice. The customer. The previous email. Paste it in; don't gesture at it.
03 · Task. One sentence. What you want the model to do. Verbs at the front: summarise, draft, classify, extract.
04 · Format. What should the output look like? Markdown? JSON with these fields? Three bullets? A 200-word paragraph? Be precise.
05 · Examples (optional but powerful). One or two filled-in examples of input → output. Pulls the model toward your taste faster than any adjective.
Same model. Same lead. Wildly different outputs. The model didn't change; the brief did.
Take a prompt you used today. Mark which of the five components you included. Add the missing ones. Run again. The difference is the chapter.
Most operators jump to the task and starve the model of the two things that most determine quality — who it is and what it knows. Spend the first third of the prompt here, even when it feels like throat-clearing.
01 · Specific beats grand. "Senior product designer with 10 years in fintech, careful about accessibility" beats "world-class expert." Specificity narrows vocabulary and stance.
02 · Stance, not just title. "Skeptical reviewer who pushes back on weak claims" gives you a different output than "summariser." Decide what you want the model's posture to be.
03 · One role per prompt. Asking the model to be "a marketer and a lawyer and a designer" produces mush. If you need three perspectives, run three prompts.
Anything the model needs but doesn't have. Most "the AI doesn't get our business" complaints are missing-context complaints.
More context isn't always better. Past ~80% of the context window, models start losing track of details in the middle ("lost in the middle"). For long inputs, summarise non-essential context — keep the essential verbatim.
Before sending, ask: (1) Does the model know who it is? (2) Does it know what it needs to know? (3) Does it know what's NOT in scope? If any answer is no, you'll be re-prompting.
Always write the role and context first; the task last. Reverse the order most beginners use. Output quality climbs noticeably the first week you try it.
A vague task gets a vague output. A precise task with a precise format gets, surprisingly often, exactly what you wanted. The model is more obedient than its critics admit — when the brief is real.
Lead with the verb. Chain dependent steps with "then." Each step is a single, checkable thing.
For reasoning-heavy tasks (math, multi-step logic, classification with judgement), asking "think step by step before answering" measurably improves output. The thinking is exposed, which also helps you debug.
For drafting, summarising, and other language-shaped tasks, it's usually unnecessary noise. Use deliberately.
If a task feels like it needs five things to be true at once, split it. Three short prompts beat one heroic prompt. Each can be reviewed, fixed, and chained (BPA Ch. 11).
Specify the shape with the same care you'd give a database schema. Examples:
{ priority, owner, next_action }."Sometimes the format is what to not do: "No emoji. No marketing speak. Don't start with 'In today's fast-paced world.'" The model takes specific bans seriously when stated explicitly.
Treat format as a contract. The model honours contracts more than it honours pleas.
For one prompt you'll run tomorrow, write its output schema before the task. Five fields, named. Then write a one-row example. Send. Compare to last week's output on the same task.
If the model keeps drifting from what you want, stop describing what you want and start showing it. Two good examples are worth a page of style notes. This is the single highest-leverage technique in the handbook.
Zero-shot — just the task. Use when the model has clearly seen this pattern in training (summarise this article).
One-shot — one input/output example. Use when you have a specific format or voice that's not the default.
Few-shot (2–5) — multiple examples. Use when the task is judgement-y (classification with ambiguous categories, copy in a particular brand voice). Diminishing returns past 5 for most tasks.
The model picks up the format (slash-separated triple), the vocabulary (specific team names), and the implicit rubric — without you ever defining "high" vs "medium." That's the power.
If your examples are inconsistent (different shapes, different tones), the model averages them and you get muddy output. Examples teach by pattern; pattern requires consistency.
For each recurring task, save your best 2–3 examples in a snippet doc. Reuse them. Edit them as the work evolves. Your "prompt library" is really an example library.
If your outputs are disappointing, the answer is almost always in this list. None of these are exotic — they're the same six mistakes everyone makes the first month. The fix in each case is small.
"World-class." "10x." "Genius." These adjectives don't condition the model toward better work; they condition it toward more confident-sounding work — which is worse when wrong.
Fix: specific role, specific stance. Skip the superlatives.
"Be concise. Be thorough. Use active voice. Match this 400-word reference. Don't be too formal. Don't be too casual. Cite sources." The model averages them and you get mush.
Fix: 3–5 constraints max. Rank if conflicts are possible.
Vague modal verbs ("maybe," "perhaps," "try to") read as optional to the model. Polite is fine; soft is not. Be courteous and direct in the same sentence.
Fix: imperative verbs. "Draft. Extract. Classify."
No length, no shape, no audience. You get an essay you didn't want. The model will fill any vacuum you leave; leave fewer.
Fix: name length, audience, and structure. Always.
Long preambles dilute the signal. Models pay disproportionate attention to the first and last parts of a prompt; the middle is more easily ignored.
Fix: state the task at the top and again at the bottom. Yes, repeat it.
You sent one prompt. Output was mediocre. You concluded the model is bad. You stopped. The pros expect 2–4 short iterations to land a tricky task. Beginners send one prompt and despair.
Fix: budget for iteration. "First pass, then I'll edit." Sometimes the edit is the work.
There is no secret prompt. There is a clear brief, given to a capable intern, with two examples and a willingness to iterate twice. That's the whole craft.
Patterns, not magic strings. Each one is a fill-in-the-blanks template that fits the five-component anatomy. Mix and match.
{ sender, ask, urgency, owner }. JSON only."Prompting Like a Pro · The Operator's Library · No. 02. Read No. 03 — Working with Claude — for the surface-level craft of applying these prompts in the actual product.
Write the brief. Ship the work. Iterate twice. That's the loop.
— END · OPERATOR'S LIBRARY NO. 02