AI Prompt Generator

What is a prompt?

A prompt is the instruction sent to an LLM that tells the model what to produce. Structured prompts compose four explicit pieces: a role, a task, constraints, and an output format. Unstructured prompts skip three of the four and hope the model guesses correctly. We rarely find the guessing pays off.

Compare two versions of the same request. The vague version reads “write me an email about a refund” with no role, no audience, no length cap, and no format. The structured version reads: role, customer-service writer; task, draft a refund email for an order shipped late; constraints, under 100 words, polite tone, no jargon, no apology theater past the first line; format, markdown body only, no subject line. The vague version returns whatever the model defaults to that day. The structured version returns the same shape on GPT-5, Claude Sonnet 4.6, and Gemini 2.5 Flash, run after run.

How structured prompts work

The role-task-format pattern is the smallest viable scaffold. Role pins the model into a perspective (“senior backend engineer” reads code differently than “junior developer”). Task names the deliverable in one sentence. Format names the shape: bullet list, JSON object, markdown table, single paragraph. Three pieces, three lines, and the model already has more signal than most production prompts ship with.

The RISEN framework extends the same idea into five pieces: Role, Instructions, Steps, End goal, Narrowing. Role and Instructions cover the top of the funnel. Steps decompose the instructions into numbered sub-tasks the model executes in order, which materially reduces hallucination on multi-step reasoning. End goal names the success criterion in measurable terms. Narrowing constrains the output: length, tone, allowed vocabulary, forbidden vocabulary.

A worked code-review prompt: Role, senior backend engineer with 10 years of Python experience. Instructions, identify performance issues in the supplied function. Steps, (1) read the function end to end, (2) flag any O(n^2) or worse complexity, (3) flag any unnecessary I/O inside loops, (4) flag any allocations in hot paths. End goal, a triage list a reviewer can act on in 5 minutes. Narrowing, Python 3.11 only, no third-party libraries assumed, output as a bullet list with each finding tagged severity (low / medium / high). Same model, same code, that prompt produces a structured triage list every time.

Common pitfalls

• Vague tone with no constraints. “Make it sound professional” gives the model no measurable target. “Professional” could mean a corporate memo, a legal brief, or a technical RFC. We replace tone adjectives with concrete bounds: “under 150 words, second-person prose, no rhetorical questions, no exclamation marks.”
• Embedding examples without delimiters. When few-shot examples are pasted inline with the instructions, the model cannot reliably tell where the example ends and the actual input begins. We wrap examples in clear delimiters (XML tags for Claude, markdown code fences for GPT) so the boundary is unambiguous.
• Mixing user content with instructions in one block. Always separate user-variable input from static scaffolding. The system prompt holds the role, constraints, and format. The user message holds only the variable input. Concatenating the two invites prompt-injection attacks and makes iteration painful, since changing one line of the scaffold means re-pasting the entire conversation.

When to use this tool

We built the generator for three concrete workflows. The first is a team standardizing prompts across multiple agents. When five different engineers ship five different prompts for “summarize this support ticket,” the outputs diverge and so does the eval surface. A shared generator pins everyone to the same role-task-constraints-format scaffold, so the only thing that varies is the variable input. The second is an engineer iterating on a prompt before production deployment. Cutting three versions of the same prompt, running each against a fixed eval set on a cheap model, and promoting the winner to the production model is a 30-minute loop with the generator and a half-day loop without. The third is teaching prompt engineering to new hires by showing the pieces explicitly. Trainees who see the role / task / constraints / format slots laid out as separate fields internalize the structure faster than trainees handed a finished prompt and told to imitate it.

Frequently asked

What makes a good AI prompt?

Five components, every time: a concrete role, a specific task, explicit constraints, a named output format, and (when format matters) one or two few-shot examples. Vague prompts produce vague outputs. Compare 'write an email about a refund' (no role, no constraints, no format) against 'role: customer-service writer; task: draft refund email under 100 words; format: markdown body only.' The second prompt scores reliably across every model we test.

Should I use few-shot examples?

Yes when output structure matters: data extraction, classification, JSON generation, structured rewrites. Two or three examples typically beat a long rule list at fewer total tokens. Skip few-shot for open-ended creative writing, where examples lock the model into one style and hurt variety. We also skip few-shot when the task is genuinely novel and an example would mislead the model toward a familiar pattern that does not apply.

How long should a prompt be?

Long enough to specify role, task, constraints, and format, and not a token longer. Most production prompts land between 100 and 500 tokens. Below 50 tokens you almost always lose precision. Above 2,000 tokens you should split into a system prompt (cached) plus a short user message (uncached). We measure every prompt in our Token Counter before shipping so we know exactly what each call costs.

Do prompts work across different models?

The structural pieces (role, task, constraints, format) port well. Tone and formatting conventions need light tuning per family. Claude reads XML tags cleanly (`<task>...</task>`). GPT prefers markdown headings and numbered lists. Gemini tends to follow plain instructional prose. DeepSeek and Llama mirror whatever structure you give them. We keep one canonical prompt per task and ship two or three short adapter variants per provider rather than one monster prompt.

How do I iterate on a prompt that's not working?

Change one variable at a time: role, then constraints, then format. Changing all three at once tells you nothing about which fix worked. Test against three to five representative inputs, not the one that happened to fail. Run iterations on a cheap model (GPT-5 Nano, Gemini 2.0 Flash-Lite, DeepSeek V4 Flash) until the structure stabilizes, then promote to your production model and confirm.

Is the RISEN framework worth using?

Yes for anyone new to prompt engineering. RISEN (Role, Instructions, Steps, End goal, Narrowing) gives you a checklist that prevents the most common failure mode: forgetting to specify one of the five pieces. Experienced prompt engineers cover the same components without naming the framework. We treat RISEN as scaffolding: useful while learning, optional once the components are second nature, but the components themselves are non-negotiable.