Start with the result, not a magic phrase
A good prompt is a useful brief. It tells the AI what you need, gives it the information that changes the answer and explains what an acceptable result looks like. You do not need a secret sentence, an elaborate persona or a claim that the model is the world’s best expert. You need to remove the guesses that matter for your task.
For a creator, that might mean the difference between “write a video” and a script you can actually record. For a Roblox builder, it might mean naming the script location and a test instead of asking for an entire game. Start by writing one sentence describing the deliverable. Then add only the details that help the model produce that deliverable.
1. Role: choose a useful perspective
A role describes the lens the AI should use. “An editor for beginner science videos” is more useful than “a genius,” because it identifies a kind of work and an audience. It can guide vocabulary, priorities and tone, but it does not create genuine credentials or make the response factually correct.
Choose a role connected to the task. A code reviewer should inspect evidence and risks; a scriptwriter should care about spoken rhythm and filmable scenes. If the task is simply sorting a list or fixing punctuation, skip the persona and give direct instructions. Structure should help the work, not become a ritual you repeat automatically.
You edit short software tutorials for beginners. Prefer plain language, visible actions and a practical payoff over feature lists.
2. Context: explain the situation that changes the answer
Context includes who the result is for, what you already have and which constraints are real. A plan for someone with a camera crew is different from a plan for someone with a phone and two hours. Tell the model about your actual resources, existing material and the viewer or customer’s starting point.
When you provide source text, separate it from instructions. Explain whether it is evidence, a draft to edit or just a tone sample. Do not paste private client files without permission, and never include passwords or access tokens. More background is not always better: remove material that distracts from the decision the model needs to make.
My audience has never used Roblox Studio. I have one working checkpoint script and can test with two players. Use the hierarchy and source I supply; ask about missing paths.
3. Task: name an observable deliverable
Use a concrete verb: compare, rewrite, diagnose, outline or draft. “Help me with my channel” gives no stopping point. “Compare three science Short formats I can produce with diagrams” creates a result you can review. If you need several stages, say which stage should happen first and what must be checked before the next one.
For code, distinguish investigating from editing. For research, distinguish collecting sources from writing narration. For an image, describe the visible scene rather than asking for a marketing strategy. A large project becomes easier to prompt when you ask for the next useful milestone instead of every eventual feature in one response.
Inspect why my checkpoint resets after death. Explain the cause with source references, then propose the smallest fix and a two-player test. Do not add new game systems.
4. Format: make the answer easy to use
Choose the shape of the answer before you get it. A tutorial might need narration, on-screen text and a shot list in separate columns. A decision might need criteria, assumptions and a recommendation. An editing task might need the revised paragraph plus a short list of substantive changes. Clear formats save you from reorganizing a long answer afterward.
Use realistic length constraints and a small example when necessary. “Six narration beats” is easier to inspect than “make it engaging.” However, do not force every answer into a table. Code, lyrics and visual prompts have their own useful shapes. Choose the format that supports the next action rather than the format that looks most elaborate.
Return six short narration beats. For each, include one filmable visual in brackets. Put factual verification notes in a separate section, not in the spoken script.
5. Guardrails: protect what must stay true
Guardrails define boundaries: do not invent statistics, do not change supplied prices, do not expose secrets, do not edit unrelated files. They work best when tied to a real risk. A newsletter rewrite needs exact dates preserved; a game economy needs server-side validation; a side-hustle plan needs honest assumptions about demand and costs.
Also say what to do when information is missing. “Ask for the relevant script” or “mark this claim for verification” is more useful than demanding confidence. Prompts cannot enforce every rule. Permissions, access policies, tests and human review still protect important systems. Treat a careful instruction as guidance, not a substitute for technical or editorial safeguards.
Use only the claims in my notes. Flag gaps instead of filling them with guesses. Do not promise views, earnings or time savings. Preserve the stated eligibility conditions exactly.
Adapt the brief to the AI you actually use
Claude: use clear task boundaries and distinguish source material from instructions. Labeled sections or XML-style tags can help organize a long brief; they are not magic. For a script, keep narration separate from facts to verify. For Claude Code, include repository context and a check it can run, then review the diff and the observed result.
ChatGPT: direct instructions, sufficient context and a clear output format are a practical starting point. A short example can show the tone or structure you want. After reading the response, identify a specific problem and revise that part of the brief. Do not rely on an “expert” role to make unsupported statements accurate.
Gemini: describe the task, inputs, constraints and expected response. When working with supplied media, say which parts it should analyze and what it should return. If you request current research, verify that the selected product can access current sources and inspect its citations. Do not assume every Gemini workflow browses automatically.
Midjourney: turn the brief into a visual description: subject, action, composition, light and style. Its own Prompt Basics recommends short, clear phrases rather than long conversational instructions. Keep the most important scene details and use supported controls for things such as aspect ratio. For thumbnails, leave deliberate title space and add lettering in an editor.
Suno: translate your intention into musical choices: genre, energy, instruments, vocals and song shape. Keep sound direction separate from lyrics where the interface supports it. Listen to the output rather than assuming tempo or section labels will be followed exactly. Check the current account terms before using generated music commercially.
Before and after: make the missing decisions visible
Before: “Give me ideas for a faceless channel.” After: “Compare three original science formats I can make with my own voice and diagrams in six hours a week. For each, list three topics, research risks and a first-episode plan. Rank by feasibility, not projected income.” The change adds resources, criteria and a useful output without claiming the niche is easy.
Before: “Make a gaming thumbnail.” After: “One blocky survivor on the left escaping a distant volcano, clear silhouette, orange lava against cyan sky and empty right-hand space for a title, widescreen composition.” This change specifies visible decisions. It does not use a writing persona where a visual description is needed.
Before: “Fix my app.” After: “Reproduce the saved-record ordering bug, inspect the query and make the smallest correction. Add a test with two timestamps showing the newer record first. Preserve ownership checks.” Here the best improvement is not more style: it is a way to prove the work is correct.
Common mistakes that make prompts less useful
One mistake is asking for too much at once: a full game, business and marketing plan in a single response. Break the work into milestones. Another is contradictory instructions, such as a detailed technical explanation that must also fit one short sentence. Decide which constraint matters most and allow the rest to flex.
A third mistake is optimizing for confident wording rather than evidence. Asking for a “viral” script or “profitable” niche does not establish likely performance. A fourth is copying someone else’s prompt without replacing their audience and resources. A fifth is sending the same vague revision repeatedly. Explain what was wrong: missing source, too much jargon, no filmable action or a failed test.
Review, revise and keep what works
Run the brief with real input, then check the answer against the original goal. Read scripts aloud, inspect images small, listen to music in context and test code with realistic users. Change one important instruction at a time when you can; that makes it easier to learn which detail improved the output.
Keep a working prompt together with notes about when it succeeds and what it still cannot do. Our prompt engine uses role, context, task, format and guardrails as a practical framework. The examples here are ready to copy; Generate writes a new prompt and Tweak saves a revised version. Use the engine brief to start clearly, and use your own review to finish responsibly.
Meta prompts: ask the AI to help design a better brief
A meta prompt is a prompt about making or improving another prompt. Instead of asking for the finished script immediately, you ask for a clearer brief based on your goal and constraints. This is useful when you know the outcome but have not identified the information the model needs. It is not a secret control phrase or a way to make a model more truthful. The resulting prompt still needs your review.
For example, ask: “Help me write a prompt for a five-minute spreadsheet tutorial. First ask up to three questions about the audience, source material and recording setup. Then return a brief with Role, Context, Task, Format and Guardrails. Use placeholders for unknowns and do not write the script yet.” This separates designing the request from executing it, and makes the missing decisions visible before a long answer is drafted.
When you already have a prompt, request an audit rather than a complete replacement. Ask the model to identify ambiguity, contradictory requirements and missing input, then revise only the affected instructions. Supply the result you actually wanted and one example of how the previous output missed it. “The narration repeats the setup and has no visual directions” gives a prompt optimizer something concrete to change. “Make this perfect” does not.
Keep a meta prompt bounded. An assistant should not invent audience demographics, research findings or a production budget to make its rewritten brief look complete. Require questions or labeled placeholders for those details. Preserve factual constraints exactly, including prices, dates, allowed files and the scope of edits. If the proposed prompt adds an assumption, decide whether it is true before using it. More confident wording is not evidence.
What2Prompt applies a similar practical idea: describe what you want to make, choose the target tool and receive a structured prompt to review and paste there. Your saved context supplies details you have chosen to share. Tweak can revise the last brief and Rate can suggest strengths and fixes. Those steps prepare instructions; they do not verify every fact, run another tool or guarantee its final output.
Help me improve the prompt below. Identify 3 ambiguities, ask about essential missing input, then revise it while preserving my source facts and constraints. Return the complete brief, not the finished deliverable. Mark assumptions explicitly.
Per-model tips: text, images, video and code need different inputs
For Claude, keep a long task’s instructions visibly separate from source material. If you are asking for a science script, give source notes in their own block and request narration and verification notes separately. When iterating, name the exact part to change. A clear task boundary is more useful than asking the model to be exceptionally intelligent. For Claude Code, provide the relevant project access and request observed checks, not a hypothetical assurance that code works.
For ChatGPT, begin with the deliverable and relevant context. A rewrite benefits from your own tone sample and exact facts to preserve. A comparison benefits from explicit criteria and an output table. If the task requires current information, confirm that the mode you use can access it and review its sources. Do not tell a non-browsing workflow to invent citations, and do not mistake fluent language for verified expertise.
For Gemini, name each supplied input’s job. A screenshot might be evidence to inspect, a document might be the source of facts and a paragraph might only demonstrate tone. Tell the model which one governs the answer when they conflict. Ask it to mark unreadable or missing details rather than invent them. Current research needs the same checks as any other model: access, source quality and a clear separation between evidence and interpretation.
At the time of writing (November 2026), image tools include Nano Banana (Gemini), GPT Image (ChatGPT), Midjourney, FLUX and Ideogram. Start with what should be visible: subject, setting, composition, light and intentional empty space. Adapt parameters to the tool rather than copying Midjourney syntax into another product. For reference edits, identify what should stay unchanged and what may change. Inspect small details and lettering before publishing.
For video, describe one shot’s opening, subject movement, camera movement and final frame. At the time of writing (November 2026), tools to consider include Google Veo 3.1, Kling, Seedance, Runway, Wan and Grok Imagine. Check each selected mode’s duration, references and audio support. For image-to-video, explain the motion you want from the starting image. Split crowded sequences into shots and review continuity instead of expecting text alone to preserve every detail.
For Suno, keep musical style separate from lyrics where supported. For Blender and Python, state the actual environment and request executable checks with concrete expected results. These tasks require different kinds of evidence: listen to a song, preview a timeline and test a script on disposable data. A common briefing framework helps organize them, but it does not make every model obey the same syntax or output controls.
Evaluate prompts with a small, repeatable test
Pick an ordinary real task rather than a showcase designed to make any answer look impressive. Use the same input for two versions of the brief and define what you will inspect before running them. A script might need an answered opening question, filmable visuals and no unsupported facts. A CSV utility might need exact totals and preserved source files. An image might need one readable subject and usable title space.
Write down failures as observable details: a missing column, an invented source, a runtime longer than your recording slot or a camera pointing away from the subject. Revise the relevant instruction and rerun the same check. Changing the audience, task and model at once makes it difficult to know what helped. Keep notes about the environment and input, because a prompt that works with one example may fail on a different case.
For higher-risk work, a prompt test is only one layer. Technical access policies protect user data; a sandbox limits unfamiliar code; expert review handles consequential legal, medical or financial decisions. Do not use a model’s self-assigned score as proof of safety. A good completion report distinguishes what was checked from what remains uncertain, so you can decide the next step without relying on confident presentation.
