Prompt craft — a working cheatsheet
A short, opinionated reference for writing prompts that actually do what you want. Skim it, steal from it, evolve your own house rules.
The anatomy of a strong prompt
Most great prompts contain seven ingredients. Not all are required, but missing two or more is usually the problem.
- Role — "You are a senior B2B copywriter." Sets vocabulary, tone, and judgment.
- Task — "Rewrite the section below." Use a precise verb. Avoid vague ones like "improve" or "make better".
- Context — Who, what, why. Paste the source material, link the brand, name the audience.
- Constraints — Length, tone, reading level, language, what NOT to include.
- Format — Markdown, JSON, table, bulleted list, plain text. Show the shape you want.
- Examples — One or two demonstrations of input → ideal output. The single highest-leverage move you can make.
- Evaluation — How will you judge it? "Reject if it uses marketing clichés" or "Pass only if every claim is sourced".
If your prompt is failing, walk these seven items in order. The missing one is almost always obvious in hindsight.
Patterns that work
- Few-shot examples — Show 1–3 input/output pairs before the real task. Models pattern-match aggressively.
- Chain of thought — "Think step by step before answering." Costs tokens, buys reliability on multi-step problems.
- Self-critique — "After your draft, list 3 things that could be wrong, then revise."
- Role + audience — "You are a financial advisor explaining to a curious 14-year-old."
- Output schemas — Define the shape:
Return JSON with keys: title (string), bullets (string[]), tone (one of: warm, neutral, formal). - Delimiters — Wrap long context in
"""..."""or<context>...</context>so the model can tell where instructions end and material begins. - Negative space — Tell it what NOT to do. "Do not invent statistics. If you don't know, say so."
Common failure modes & fixes
| Symptom | Fix |
|---|---|
| Vague output | Add evaluation criteria + a one-line example |
| Off-tone | Add 2–3 example sentences in the target voice |
| Hallucinated facts | Add "If you are not certain, write 'UNKNOWN' instead" |
| Wrong format | Show the format with a literal example, not a description |
| Ignores instructions | Move the most important rule to the END (recency bias) |
| Too long | Add a hard limit: "Reply in ≤120 words." |
| Persona drifts mid-answer | Add "Stay in role for the entire response." |
The refinement loop
Draft → Run → Diagnose → Tweak ONE variable → Re-run.
Change one thing at a time so you know what helped. Scrivenry's Versions feature (coming soon) and Notes field are designed for this. Capture what you tried and why.
When to reach for which model
| You need… | Try |
|---|---|
| Hardest reasoning, long context, multimodal | openai/gpt-5 |
| Balanced reasoning at lower cost | openai/gpt-5-mini or google/gemini-2.5-flash |
| Big visual + text + huge context | google/gemini-2.5-pro |
| Cheap, fast classification or summarization | google/gemini-2.5-flash-lite |
Capture the model in Scrivenry's Model field — it's a real variable in how the prompt behaves.
Variables done right
- Name them like product fields, not algebra:
{{customer_name}}, not{{x}} - Group related variables next to each other in the prompt body
- Use Fill & copy in Scrivenry to test combinations without copy-pasting
- Default unfilled tokens render as
{{name}}so you can spot misses
Sequences vs single prompts
Use a single prompt when the task fits in one model turn and a single context window. Use a sequence when:
- The output of one step is the input to the next (research → outline → draft)
- Different steps need different roles or models
- You want to checkpoint and rate intermediate results
- Different teammates own different stages
Rule of thumb: if your prompt has more than three distinct phases or exceeds ~800 words, it probably wants to be a sequence.
A starter template — Role · Task · Format
You are [ROLE].
Your task is to [TASK] for [AUDIENCE].
Context:
"""
[PASTE CONTEXT HERE]
"""
Constraints:
- [CONSTRAINT 1]
- [CONSTRAINT 2]
Format:
- [DESCRIBE OUTPUT SHAPE]
Example output:
[ONE CONCRETE EXAMPLE]
Use the Critique with AI button on any prompt to get a score, weaknesses, and a rewrite — the fastest way to apply this guide to a real prompt.