Prompting patterns that actually work
Not "magic prompts." Reusable structures you can apply to any model, any task. Learn these six and you'll rarely need to copy someone else's prompt again.
Original writing, informed by hands-on use and formal AI coursework. The goal is to teach the underlying pattern, not hand you a fish.
1. Role + Goal + Context + Format (the workhorse)
The backbone of a good prompt has four parts:
- Role — who the model should act as ("You are a meticulous copy editor").
- Goal — the actual outcome you want, stated as a result, not a topic.
- Context — the material and constraints the task needs.
- Format — the exact shape of the output you'll use.
Miss any one and quality drops. Most bad output is a missing format or a vague goal.
2. One good example beats a paragraph of description (few-shot)
Instead of describing the style you want, show one example of input → output. Models pattern-match far better than they follow adjectives. One worked example is worth three sentences of "make it professional but friendly."
3. Ask for the thinking, then the answer
For anything with reasoning — analysis, debugging, math, planning — tell the model to work through it step by step before giving the final answer. You get better answers and a visible chain you can check. For a clean final deliverable, ask it to think first, then put the final result in its own clearly-marked section.
4. Separate instructions from data
When your prompt contains both instructions and content to act on, mark the boundary so the model never confuses the two:
<instructions>Summarize the report below in 5 bullets.</instructions>
<report>
...paste the report...
</report>This is the single most reliable fix for "it answered my example instead of doing the task." (It's also a basic prompt-injection defense — untrusted content stays in the data slot.)
5. Critique-and-revise, not re-roll
Don't just regenerate and hope. Tell the model what's wrong and ask for a specific revision: "The intro is too long and the tone is stiff — cut the intro to two sentences and make it conversational." You converge on what you want in two or three passes instead of gambling on a fresh roll.
6. Give it an out
Tell the model what to do when it doesn't know: "If the answer isn't in the provided text, say 'not found' rather than guessing." This is the cheapest way to cut confident-but-wrong answers (hallucinations).
Putting it together
A strong prompt is usually: a role, a concrete goal, the context marked off from instructions, one example of the output, and a clear format — plus permission to say "I don't know." Everything else is refinement.
The AUTOMAT framework (a reusable skeleton)
When you want a checklist instead of intuition, AUTOMAT covers the parts a strong prompt almost always needs. Walk the letters and you rarely miss one:
- A — Act as a role. Give the model a persona with the right expertise ("Act as a home-network security reviewer").
- U — User persona. Say who the output is for ("explaining to a non-technical homeowner"). Same facts, very different answer.
- T — Targeted action. The specific verb and deliverable ("audit these firewall rules and list risks"), not a vague topic.
- O — Output definition. The exact shape you'll use — table, JSON, 5 bullets, a one-page brief.
- M — Mode / tone. Formal, blunt, friendly, technical. Set it explicitly.
- A — Atypical cases. What to do at the edges: "if a rule is ambiguous, flag it rather than guess."
- T — Topic whitelisting. Fence the scope and sources: "only use the config I pasted; don't invent rules that aren't there."
AUTOMAT and the four-part pattern up top are the same idea at two resolutions — use the quick version daily, reach for the full checklist when the output really has to land.
Going further: reasoning & verification strategies
For high-stakes or multi-step work, layer these on:
- Chain of Thought (CoT). Ask the model to reason step by step before answering — better results on anything with logic, math, or planning.
- Tree of Thought (ToT). For open problems, ask it to sketch a few distinct approaches, compare them, then commit to the best. Slower, but it escapes the first-idea trap.
- Source-fact verification & rating. When accuracy matters (research summaries, briefs), require the model to tie each claim to a source and rate its confidence (e.g., 1–5). It surfaces the shaky claims instead of burying them.
- Iterative refinement (double-pass). Have the model draft, then critique its own draft against your criteria, then produce a v2. One extra pass catches a surprising amount.
- Dataset-grounded anchoring. Paste the actual data/context and instruct the model to answer only from it. The single best defense against confident fabrication.
These sections are informed by the author's AI & prompt-engineering coursework, rewritten as original explanation — the frameworks (AUTOMAT, CoT/ToT) are community/industry terms, not proprietary course content.
Part of HomeForge. See also: Managing Claude & ChatGPT.
Get the free Starter Kit →