Prompt engineering for beginners: a framework that works

Why prompts fail

Prompt engineering, for a beginner, comes down to one skill: giving an AI model the job, the evidence, the boundaries, and the output shape it needs to produce something useful. Most bad prompts fail because they ask for help without defining what "good" means. The model fills the gaps with likely patterns, and those patterns sound polished while missing your real goal.

Almost every disappointing answer traces back to one of four things:

1. The task is unclear. "Make this better" can mean shorter, warmer, more persuasive, more accurate, less technical, or easier to scan. The model picks one and you get the wrong one.
2. The context is missing. Without the audience, the source material, the decision criteria, or the goal, the model optimises for generic quality, which is nobody's quality.
3. The constraints are weak. No length, no tone, no sources, no privacy rules, no list of things to avoid, so the answer drifts toward the average of everything it has seen.
4. There is no review loop. The first answer is a draft. Treating it as the finished product is the most common mistake of all.

A strong beginner workflow treats prompting as a short brief plus a quality check. You define the task, include the context, set constraints, request a format, then ask the model to expose assumptions or uncertainty. One honest caveat before the framework: a two-line question deserves a two-line prompt. The full brief below is for output you will reuse, share, or be judged on, and using it on everything is how prompting starts to feel like paperwork.

Framework: instruction, context, constraints, and format

Use this framework for almost every serious prompt: Instruction, Context, Constraints, Format. It works on every major model, and each part changes the output in a way you can see.

Part	What it answers	Weak version	Strong version
Instruction	What should it do?	"Help with market research"	"Create a competitor comparison table for five products in this category"
Context	What does it need to know?	Nothing supplied	Audience, goal, source text, decision criteria, examples
Constraints	What are the rules?	None	Under 400 words, no jargon, mark anything uncertain, do not invent numbers
Format	What shape should it come back in?	Unspecified	A table with columns for claim, evidence, and confidence

Instruction is the action. Start with a verb and a deliverable: summarize, compare, rewrite, extract, critique, classify, outline, debug, or plan. Instead of "Help with market research," write "Create a competitor comparison table for five products in this category." The clearer the instruction, the easier it is to judge the output.

Context is the material the model needs to avoid guessing. Include audience, goal, background, source text, examples, decision criteria, and definitions. For long inputs, structure matters: Anthropic's long-context tips (linked below) say to put long documents near the top of the prompt, label each one, and place the instructions after the material rather than before it.

Constraints are the rules. Add length, tone, scope, excluded claims, required evidence, privacy boundaries, reading level, and what to do when information is missing. Constraints are especially useful when you need prompts that improve accuracy: tell the model to mark unknowns instead of inventing details, separate facts from assumptions, and cite the source section when possible.

Format is the output container. Ask for a table, checklist, memo, outline, JSON-style fields, email, scoring rubric, or step-by-step plan. OpenAI and Gemini both ship dedicated structured-output modes for exactly this reason: an explicit shape is what makes an answer easy to parse, compare, and reuse, and stating the shape in plain words gets you most of that benefit in a chat window.

Copy-paste prompt template: "Instruction: [specific action and deliverable]. Context: [audience, goal, source material, examples, definitions, decision criteria]. Constraints: [length, tone, must include, must avoid, evidence rules, privacy rules, uncertainty rules]. Format: [table/checklist/memo/email/outline/JSON-style fields]. Before finalizing, list assumptions, missing information, and one follow-up question that would improve the result."

Five templates cover most work tasks

The best prompts for work are reusable patterns, not one-off magic lines. Use the examples below as starting points, then save the ones that consistently produce useful output.

Writing prompt: "Instruction: Rewrite this draft for clarity and specificity. Context: The audience is [audience], the goal is [goal], and the draft is [paste draft]. Constraints: Keep all facts unchanged, avoid hype, use plain language, and make the next step obvious. Format: Return a revised draft, then a table with original issue, edit made, and reason." This works because it gives the model a job, a reader, a standard, and a reviewable edit trail.

Research prompt: "Instruction: Turn these sources into a research brief. Context: I am evaluating [question] for [decision]. Sources: [paste source excerpts or links/notes]. Constraints: Use only the provided material, separate evidence from inference, mark weak evidence, and do not invent statistics. Format: Return key findings, evidence table, open questions, and recommended next sources." Compare that to "Research this topic," which invites the model to invent whatever the brief did not pin down.

Planning prompt: "Instruction: Create a two-week execution plan. Context: Goal is [goal], team is [team], deadline is [date], constraints are [constraints]. Constraints: Prioritize reversible steps, identify blockers early, and keep each task owner-friendly. Format: Table with day, task, owner, output, risk, and success check." A team can run that table.

Analysis prompt: "Instruction: Compare these options and recommend one. Context: Options are [options]. Decision criteria are [criteria]. Constraints: Include tradeoffs, strongest counterargument, and what would change the recommendation. Format: Scorecard table plus a short decision memo." The counterargument line is the load-bearing part: without it, every scorecard recommends whatever you already preferred.

Review prompt: "Instruction: Critique this output before I use it. Context: Intended use is [use case]. Output to review: [paste]. Constraints: Check accuracy, unsupported claims, missing context, tone, risk, and places needing human review. Format: Return critical issues first, then suggested edits, then a final confidence rating." Run it on anything going to a client.

The iteration loop fixes one thing per pass

Good prompt engineering is iterative. You should expect the first response to reveal what the prompt forgot. Instead of starting over, use a loop: evaluate, diagnose, revise, compare, save.

Evaluate the answer against five questions. Did it complete the exact task? Did it use the provided context? Did it obey constraints? Is the format easy to use? Are claims, assumptions, and uncertainties visible? If the answer fails, diagnose the prompt before blaming the model. A vague output often means the task was vague. A generic output often means the context was thin. A risky output often means constraints were missing.

Then revise with one precise instruction at a time. Say "make the recommendation more evidence-based and mark unknowns" instead of "try again." Say "convert this into a table with columns for claim, evidence, confidence, and next check" instead of "be more structured."

When the output matters, run the revised prompt across two models and compare them on four things: accuracy, completeness, tone control, and how much editing each one still needs. The second run is cheap. By the Whizi model cost index (list prices fetched 2026-08-20), a standard answer of 1,000 input and 500 output tokens costs $0.007 on Claude Sonnet 5 and $0.02 on GPT-5.5, so a second opinion on important work costs pennies while a wrong answer does not.

Use this iteration checklist: 1. Name the failure in one sentence. 2. Add the missing context or constraint. 3. Tighten the output format. 4. Ask for assumptions and uncertainty. 5. Run a second model when the task is important. 6. Save the improved prompt only after it works on a real task, not a toy example.

Build your reusable prompt library

A reusable prompt library is where beginner prompting becomes a daily workflow. Instead of keeping random prompts in chat history, organize templates by job: writing, research, planning, coding, customer analysis, document review, and decision support. Each template should include the same four fields: instruction, context, constraints, and format.

Add a short usage note to every saved prompt: when to use it, what inputs it needs, which model performed best, and how to verify the output. For example, a research prompt might require source excerpts and a claim-evidence table. A writing prompt might require audience, voice sample, and must-keep facts. A coding review prompt might require the diff, expected behavior, and test context.

Whizi fits this workflow because you can run the same template across GPT-5.5, Claude Sonnet 5, Gemini, and DeepSeek V3.2 in one workspace instead of committing to one model forever, and route routine runs to a cheap model: DeepSeek V3.2 answers for about $0.0005 at list rates, roughly 40 times less than GPT-5.5. Start with three templates this week: a writing rewrite, a research brief, and a decision scorecard. Run each on a real task, compare outputs, revise, and keep the version that produces the most useful result with the least cleanup.

For a broader beginner guide, read How to use AI. To turn your best prompts into standing routines, the AI workflow templates guide gives you seven ready-made ones. For research-heavy work, try the Founder Research Stack. When you are ready to turn prompting into a repeatable workspace, create your account at register.