Most people treat AI like a chat toy. Then they blame the model when results seem, well, childish. This is a usage problem, not an intelligence problem. Below is a concise playbook to fix it.

Play in the Playground

For expert level users, using a prompt playground or API-backed IDE will give you the best flexibility. While you don’t see it, applications like ChatGPT add in some extras with each prompt to make your prompt “better”. In the playground you control system prompts, temperature, models, and can run side-by-side tests. This turns prompting into engineering, not vibes. Visit ChatGPT’s playground.

  1. Sign in to the OpenAI developer platform and open Playground.
  2. Pick a model (start with a general one like GPT-5 class).
  3. In the left panel, set:
  • System: rules and role (e.g., “You are a precise writing coach. Return only JSON.”)
  • User: the actual task and data.
  • (Optional but beneficial.) Examples: one tiny example of ideal input → output.
  1. Temperature near 0.2–0.5 for accuracy; leave other defaults as is for now.
  2. Click Run. If you like the setup, Save it as a preset so you can reuse and A/B test later.

Tip: Playground is meant for fast experimenting without code. When you are ready to automate, switch to the API (https://daniellefavreau.com/2025/09/11/ai-and-playing-with-the-pros/).

Keep Prompts Short (~200 words)

Short prompts reduce ambiguity and cut context noise. Start tight, then add only what matters.

Verbose (what not to do)

“Hi ChatGPT! I’m experimenting with AI summaries for my newsletter and would love a super helpful, creative-yet-concise overview of whatever article I paste below. Please aim for a friendly but professional voice that feels like The Economist meets Wired (but less stiff), and also try to include a tiny emoji or two to keep things light. I need five key takeaways, but if you think six or seven would make it better, feel free! Also please add a narrative paragraph at the end that ties everything together for busy executives who only have 30 seconds. Keep it short but thorough , don’t omit anything important. Include quotes if they’re punchy, and add sources if you can infer them. Output should be JSON so my tool can parse it… but also include the bullets as markdown for human readers. Oh, and avoid negativity unless critical, yet identify at least one risk. Try not to be repetitive, except when repetition helps memory. Please stay under 120 words but also be comprehensive. Here’s the article text (paste below). If it’s medical or legal, add a disclaimer; if it’s marketing, add a CTA. Thanks! 🙏”

Concise (≤200 words, structured)

You are a precise summarizer. Follow rules. Be brief.

Summarize the article text I paste as five bullets (≤12 words each).
Also provide one concrete risk and an overall sentiment: positive | neutral | negative.
Here is the article text:
{PASTE_TEXT_HERE}

Return ONLY valid JSON, example:
{
“bullets”: [string, string, string, string, string],
“risk”: string,
“sentiment”: “positive” | “neutral” | “negative”
}

Rules:
– No text outside JSON.
– Don’t invent facts; if unsure, write “unknown” and set sentiment “neutral”.
– Keep total under 120 words.

Prompt Purposefully

  • System: sets role, guardrails, rules.
  • User: supplies task and data.
  • Assistant: shows format or a worked example.

Separation prevents instruction collisions.

An Example Increases Accuracy

If you have an example of what you want as a result, provide it, just one is fine, adding multiple is only helpful if there may be differences between them (e.g. chunks of data in one that may not be in others). Example:

Return ONLY valid JSON, example:
{
“bullets”: [string, string, string, string, string],
“risk”: string,
“sentiment”: “positive” | “neutral” | “negative”
}

Write Unambiguous Language

No “roughly,” “kind of,” or “be creative.” Use explicit verbs, constraints, and acceptance criteria.

You are a precise summarizer. Follow rules. Be brief.

Remove Conflicting Instructions

Conflicts cause drift. Keep one source of truth for rules, one for task, one for example: “super helpful, creative-yet-concise overview” Not helpful. Should it be creative or concise? Super helpful or an overview?

Keep It Simple

Plain, direct, testable. If a requirement can’t be tested, rewrite it.

AI works best when it’s provided with incremental instructions. It can also become overwhelmed with too many instructions (very long conversations). If you need to, start a new conversation, branch, and seed with what you’ve established as a core pattern.

Iterate with Data, not Feelings

A/B prompts. Freeze seeds where possible. Track accuracy, latency, cost, and violation rates. Promote only when a variant wins on your metric. Think of it like those March Madness maps, only one can win. Keeping notes of what works and what doesn’t will help you refine (and reduce costs).

Define the Output Format Explicitly

If you ultimately need JSON then ask for JSON, need an Excel sheet? Ask for it. Need a PDF formatted with certain titles? Give it the format and ask for it. Save yourself work by having AI do the heavy (and light) lifting for you.

Use the Right Model

Large language models are skilled guessers: they generate the next word based on patterns from lots of text. That skill makes them fluent, but it can also produce mistakes or invented details.

When you connect the model to real information, like a well trained LLM, your product’s knowledge base, or a web search, and ask it to cite where it found things, accuracy goes up and “made-up” answers go down. Training helps, but access to up-to-date sources is what turns fluent output into trustworthy answers.

A Reliable Prompt Scaffold

Use this four-part template, it’s simple and hard to break.

[System | Context]
You are a {role}. Follow the rules. Refuse unsafe requests.
[User | Instructions]
Task: {single clear objective}. Constraints: {limits}. Inputs: {data here}.
[Assistant | Output format]
Return ONLY valid JSON that matches this schema:
{
  "result": string,
  "confidence": number,   // 0..1
  "citations": [string]
}
[System | Rules]
- No extra text outside JSON.
- If unsure, set "confidence" < 0.6 and list "citations" as [].

When in Doubt, use AI to Generate Examples for AI

If you’re anything like me, you may not be as good as making concise instructions as the system would prefer. So, ask AI to help. First ask the model to draft 10 candidate examples (good and bad). Manually cull for accuracy. Feed the curated few back as shots. This is fast and raises floor quality. Iterate and refine.

How to Operationalize

  1. Set up a playground with versioning and side-by-side runs. Start a metrics sheet.
  2. Model choice. Start smart, then optimize for cost/latency. Back facts with retrieval.
  3. Shrink your prompts: remove adjectives, metaphors, and repeats. Measure before and after.
  4. Split messages into system/user/assistant. Move rules to system, task to user, schema and examples to assistant.
  5. Add one example: if accuracy lifts, consider a second if necessary.
  6. Request a specific format (JSON/XML/CSV). Add a validator. Fail fast on bad output.
  7. A/B variants for a single metric. Keep a changelog. Promote the winner.

Using AI is like making a recipe. While you could make a cake in the forest with some stones and a fire, you’ll get a much better result in a nice convection oven. The best meals are simple, few high-quality ingredients, only what’s absolutely necessary. That’s your prompt, keep it clean, concise, and direct. I know, you want to be polite and conversational with AI (it’s so polite and nice!) but not only does that waste cycles it can also cause inferior results.

AI is a tool there to make your life easier, help it help you.