--- title: "A prompting framework that travels across tools" source: https://www.taim.io/ai-productivity/ai-prompting-framework published: Tue Apr 14 2026 14:50:01 GMT+0000 (Coordinated Universal Time) updated: Wed Jun 10 2026 06:28:43 GMT+0000 (Coordinated Universal Time) description: "A compact, portable AI prompting framework you can use across tools, with a concrete first task, quality checks, and a simple loop to improve your results." --- # A prompting framework that travels across tools Most people talk to AI like a slightly smarter search box. That works until you want consistent results. A simple prompting framework turns scattered questions into repeatable workflows you can carry across tools, models, and new releases. Most people talk to AI like a slightly smarter search box. That works until you want consistent results. A simple prompting framework turns scattered questions into repeatable workflows you can carry across tools, models, and new releases. ## What you’ll be able to do - Use a 4-part prompting framework (ROLE → GOAL → INPUT → RULES) in any mainstream AI tool. - Run a concrete first task today and tell, with specifics, whether the AI output is good or bad. - Debug your own prompts and retry intelligently instead of copy‑pasting new “magic” prompts every week. ## 1. A small prompting framework that actually travels Most prompt advice is glued to a specific tool screenshot. That breaks as soon as the UI moves or the model changes. Instead, you want a **small set of patterns** you can reuse whether you're in Claude Opus 4.7, GPT-5.1, Gemini, or a niche in‑house assistant. Think of these patterns like coding style, not copy‑paste snippets. The framework in this article fits into four words: > **Role → Goal → Input → Rules.** Most useful prompts are just these four ideas made explicit. Write them down in that order, keep them short, and you’ll outperform 90% of “10x productivity” prompt lists. Models change; this structure doesn’t. We’ll spend the rest of the article doing three things: 1. Finding your starting point. 2. Using the framework on a real task you can run today. 3. Learning how to read the output, then tighten your next attempt. ## 2. Find your starting level Before we touch a prompt, decide how you currently use AI. That decides how fast you move. Use this quick self‑check: Level 0 – Curious, but barely using it You’ve opened ChatGPT, Claude, or a similar tool a few times. Mostly to ask “what is X?” or to play. No recurring use at work or study. Level 1 – Casual user You paste text in to summarize, translate, or rephrase emails. You’ve never written a multi‑line prompt on purpose. You tweak wording randomly when the output is bad. Level 2 – Emerging workflow You have at least one regular use case: drafting emails, summarizing calls, planning lessons, basic coding help. You’ve noticed that structure and context matter, but you don’t have a consistent pattern yet. If you’re Level 0–1, follow the article as written. If you’re Level 2, skim the explanations and focus on the debugging and reuse sections later on. ## 3. The core pattern: ROLE → GOAL → INPUT → RULES Models from Anthropic, OpenAI, and Google all share a basic behavior: they respond better when you **make structure and constraints explicit**. The 4‑part pattern is a lightweight way to do that. Here’s what each part means in practice: ROLE – who is the AI pretending to be? Short description, tied to the task: “You are a concise business analyst” or “You are a patient CS tutor for a beginner.” This steers tone and level of detail. GOAL – what is the one main outcome? One sentence. Not three. For example: “Create a one‑page brief a busy manager can read in 3 minutes.” If you can’t summarize the goal, the model can’t either. INPUT – what should it look at? Point clearly: “Use ONLY the article between the triple quotes,” or “Use the meeting notes below” and then paste them. If there are multiple pieces, label them. RULES – how should the answer look? Length, structure, style, and any hard constraints: bullet sections, word limits, “no invented facts,” audience level, forbidden topics, etc. The template version looks like this: Text `ROLE: You are a … GOAL: Your task is to … INPUT: Use ONLY this material: """ [paste text] """ RULES: 1. … 2. … 3. … ` You’ll see this run end‑to‑end in the next section. ## 4. Your first real attempt: a manager brief from an article Let’s do something concrete you could plausibly use at work: turn a long article into a one‑page manager brief. You’ll need: - Any modern AI chat tool (Claude, ChatGPT, Gemini, etc.). - A 1–2 page article or internal document. A news article, a blog post, or a short report is fine. ### Step 1: Grab your input Open an article of about 800–1500 words. Avoid opinion pieces with no clear facts for this first run. Copy the main text only. Skip sidebars, unrelated comments, or navigation junk. ### Step 2: Paste this base prompt Open your AI tool and paste this, but **don’t run it yet**: Text `ROLE: You are a concise business analyst who writes for busy non-technical managers. GOAL: Your task is to create a one-page brief that a manager can read in under 3 minutes to understand the key points and decisions from the article. INPUT: Use ONLY the article between the triple quotes below. Do not use outside sources. """ [PASTE ARTICLE HERE] """ RULES: 1. Start with 2–3 bullet points under the heading "Why this matters". 2. Then add short sections for: "Key facts", "Risks", and "Decisions or options". 3. Use clear, plain language. Imagine the reader has not seen the article. 4. Do not add any facts that are not clearly supported by the input. 5. Aim for 400–600 words total. ` Now paste your article where indicated and run it. ## 5. Reading the output: what good vs poor looks like Treat the output like you would treat a junior colleague’s draft. You’re not asking “do I like it?” You’re asking “does it meet the spec?” Use this quick comparison as you read: Aspect Good result Poor result Length Roughly 400–600 words, clearly one page of text Very short (few lines) or sprawling essay with no control Structure Has all sections: Why this matters, Key facts, Risks, Decisions/options Missing sections, headings wrong, or just a generic paragraph Faithfulness Every claim traceable to the article New numbers, sources, or confident claims not in input Audience fit Plain language, no unexplained jargon Overly technical or fluff with buzzwords Emphasis Highlights what a manager must know to act Gets lost in trivia or background details Spend two minutes marking, mentally or on paper: - What worked? (e.g., “structure mostly right, language clear.”) - What clearly missed? (e.g., “left out risks, invented a timeline.”) That gap between your instructions and the output is where the framework earns its keep. ## 6. Debugging your prompt: adjust and retry If the first attempt is perfect, you got lucky. Most runs will be *okay but off* in one or two ways. Instead of starting over with a brand‑new prompt, change just one or two parts. Here are the usual failure modes and how to fix them. **1. Output is too vague or fluffy** Likely cause: GOAL or RULES are fuzzy. Edit the GOAL to be sharper: “focus on decisions and risks only, not background history.” Or tighten RULES: “Every sentence in ‘Key facts’ must contain a concrete number, date, or named entity from the article.” **2. Missing important parts of the article** Likely cause: INPUT unclear or too large. Make the input boundaries obvious: “Focus on the section between the headings X and Y.” If your article is very long, try a shorter one or summarize in two passes (sections first, then combine). **3. Made‑up details (hallucinations)** Likely cause: lack of constraints in RULES. Strengthen them: “If the input does not contain a specific fact, say ‘not specified in the article’ instead of guessing.” On a second run, explicitly ask: “List any statements that might not be directly supported by the text.” **4. Tone doesn’t match your audience** Likely cause: ROLE too generic. Be specific: “You are a chief of staff writing for a time‑poor VP who hates jargon and buzzwords.” If needed, paste a short example paragraph and say: “Match this style.” Run the **second attempt** with just those edits. Compare again to the table from the previous section. You should see at least one dimension clearly improved. ## 7. Making it a repeatable habit across tools Once you’ve done this once, the real value is in reusing the pattern. The key is to treat your prompts like tiny pieces of source code: - **Version them**: keep a note with “Brief v1”, “Brief v2 – stronger rules about facts,” and what changed. - **Name the pattern, not the tool**: “Manager brief from long text” works in Claude, ChatGPT, or an internal bot. - **Port carefully**: when you move from, say, GPT‑4.8 to GPT‑5.1, start with the same prompt and see what changed before tweaking. Most tools now support system prompts or “instructions” (see Anthropic’s and OpenAI’s docs). You can set part of ROLE and RULES there, then keep GOAL and INPUT in each conversation. Over time you’ll build a small library: a few prompts for briefs, a few for code review, a few for lesson planning. Same framework, different inputs. ### Cheatsheet: portable prompting in the field #### ⚡ 4-part prompt skeleton Use this bare-bones version in any tool: Text `ROLE: You are a [type of expert] for [audience]. GOAL: Your task is to [clear one-sentence outcome] that [time/length constraint]. INPUT: Use ONLY this material: """[paste or describe]""". RULES: 1. Structure: [headings, bullets, order]. 2. Limits: [word range, no outside sources, etc.]. 3. Style: [tone, reading level]. ` Aim for 5–10 lines total. If it’s longer, cut or postpone details until the second message. #### 🔧 Debugging checklist for bad outputs When an output misses, scan this in order before blaming the model: 1. **Goal** – Can you point to a single sentence that states the outcome? If not, rewrite it. 2. **Input** – Is the source text clearly delimited and reasonably sized (under ~3–5k words for a beginner run)? 3. **Rules** – Did you specify structure, length, and “no inventions”? If not, add them. 4. **Example** – If tone/format is still off, add a short example and say “match this style.” Change a maximum of 2 things per retry so you can see what helped. #### 📋 Length and structure rules of thumb For typical chat models: - Summaries of 1–2 page inputs: ask for **150–600 words** depending on detail needed. - Emails: most real-world drafts land between **80–250 words** for clarity. - If you want distinct sections, always **name them** in RULES; don’t just say “structured.” - When in doubt, request: *“Use short paragraphs of 2–3 sentences.”* That avoids walls of text and makes editing easier. #### 🎯 When to add examples (few-shot) Add a tiny example when: - The model keeps getting tone or format wrong **even after** you’ve specified ROLE and RULES. - You need a specific genre (e.g., product release note, lab report abstract). Keep examples short: 1–2 paragraphs or 3–5 bullets. Label clearly: “EXAMPLE OUTPUT (imitate structure and tone, not content).” Stop using examples once the pattern sticks; they consume context window space. #### ⏱️ A 15-minute weekly practice loop Once a week, run this cycle: 1. Pick one real document from your work or study (~2 pages). 2. Apply the 4-part framework to create a brief or reformatted version. 3. Spend 3 minutes comparing to your spec (structure, length, faithfulness). 4. Edit GOAL or RULES to fix the biggest gap and rerun. 5. Save the before/after prompts and outputs in a note. Do this for 4 weeks and you’ll have a small personal library of prompts plus a feel for how different models behave. ### FAQ: getting practical with an AI prompting framework #### ❓ How detailed should my prompts be for everyday tasks? Aim for “compact but explicit.” In most day-to-day tasks, 5–10 lines following ROLE → GOAL → INPUT → RULES is enough. If you find yourself writing multiple screens of instructions, you’re probably mixing several goals into one prompt. Split them into stages instead: for example, first ask the model to extract key facts, then in a second message ask it to write the brief. On the flip side, if your prompt is a single vague sentence (“summarize this”), expect highly variable quality — the model is forced to guess what matters to you. #### 🤔 Do I really need a framework for simple questions? For truly simple questions — “what is photosynthesis?” — you don’t. A framework shines when you care about output shape, audience, or reliability. Once money, reputation, or team workflows touch the output, structure stops being optional. The habit of writing ROLE, GOAL, INPUT, and RULES also trains you to clarify your own thinking. Even if the question feels simple, try the framework on a few everyday tasks; you’ll quickly see where vague instructions were hiding. #### ⚠️ How do I avoid the model making things up (hallucinating)? You can’t eliminate hallucinations completely, but you can reduce them and make them obvious. First, **anchor the INPUT** and say “Use ONLY this material” or “If something is missing, say ‘not specified’ instead of guessing.” Second, in RULES, forbid new sources: “Do not add external facts or data.” Third, ask for a separate section like “Uncertain or not specified” so the model has a place to park missing information instead of fabricating it. For high‑stakes work, always verify key claims against the source or an authoritative reference — models are fast, not authoritative. #### 💡 When should I change temperature vs changing the prompt? Treat temperature as a **fine‑tuning knob**, not a primary fix. If the structure, content, or faithfulness are wrong, fix the prompt first. Once the output reliably hits your spec, adjust temperature to taste: lower values (0.0–0.3) make responses more deterministic and repetitive; higher values (0.7–1.0) make them more creative and varied. For summaries, briefs, or anything procedural, stay low. For brainstorming names or story ideas, go higher. If a change in temperature seems to wildly change quality, that’s usually a sign your GOAL or RULES are still too vague. #### 🎯 How do I reuse a good prompt across different tools? Start by separating your **framework** from tool-specific features. Keep ROLE, GOAL, INPUT, and RULES as plain text that you can paste anywhere. When you move from, say, Claude in the Anthropic console to ChatGPT in OpenAI’s interface, paste the same prompt and run a test on a known input. Note any differences: maybe one model is more verbose, or one needs stricter length rules. Capture those as small edits, like “add explicit word range for GPT” in your notes. Over time, keep a single “canonical” version of each prompt pattern and then short per-tool deltas, so you’re not maintaining five independent prompts that slowly drift apart. ### Bringing it together The useful part of “prompt engineering” isn’t clever phrasing. It’s having a small mental model of how these systems respond to structure, then using that model consistently. ROLE → GOAL → INPUT → RULES is small enough to remember and powerful enough to survive model upgrades and tool redesigns. You saw it handle a concrete task — a manager brief — and you saw how to judge the output against specifics instead of vibes. From here, the skill is repetition. Any time you catch yourself typing a one‑line, fuzzy request into an AI tool, pause and add at least GOAL, INPUT, and one RULE. Treat each prompt like source code: version it, debug it, and keep what works. The tools will keep changing; the habit will keep paying off. ### Next steps: build your own portable prompts - Pick a second real task from your work or study (e.g., drafting an email, outlining a lesson, summarizing meeting notes) and rewrite it using ROLE → GOAL → INPUT → RULES. - Run the prompt in two different tools (for example, Claude and ChatGPT) with the same input and compare outputs against the table in Section 5. - Choose one of those tools and do a second iteration: fix the biggest gap (length, structure, faithfulness, or tone), rerun, and save both the old and new prompts. - Create a simple note titled “Prompt patterns” and store your manager brief prompt plus one other. Add a one-line comment under each with when to use it. - Set a reminder one week from now to repeat the 15-minute practice loop from the cheatsheet with a fresh document. Treat it as light training, not a big project.