Book a free call
TrainingProductsTeamSupportContactLabBlog Book a free call
Lab build · October 31, 2026

AI Explainer Lab: Prompt Anatomy

Most people type a simple chat prompt into AI and hope. We took three everyday jobs, wrote the simple chat prompt most people would type, then added the five pieces of a complete prompt one at a time, in every combination, on four AI models. Snap the pieces in and out and watch what each one fixes. Every reply below is a real, recorded run.

Draft: partial data while the recording finishes. Not for posting.

Improve the prompt

This is the simple chat prompt most people type, and what the AI sent back. Your turn: add the pieces below and watch the failed checks turn green. Every reply is a real recorded run of that exact prompt.

  1. Pick a job (and a model, if you like)
  2. Click a piece to add it to the prompt
  3. Watch the reply and checks change
The five pieces of a complete promptClick each piece to improve the prompt ↓
The prompt that was sent
What came back
0 / 0 checks

Same job, two prompts

The simple chat prompt against all five pieces, on the job and model you picked above, across every recorded run.

What each piece fixes

For each check, how much more often it passes when that piece is in the prompt than when it isn't, averaged over every other combination of pieces. Green helps, red hurts.

Check your own prompt

Paste a prompt you use. This looks for the five pieces and suggests what to add. It's a quick checklist, not a grade, and it runs in your browser: nothing you type leaves this page.

    Get the Prompt Builder, free

    A small AI assistant that interviews you about the job, then writes a complete prompt with all five pieces. It works in the AI tool you already have, with no connectors and nothing for IT to set up.

    • Claude: a skill you add once in Claude's settings, or paste into a Project
    • ChatGPT: instructions for a custom GPT, or paste them into a Project
    • Microsoft 365 Copilot: instructions for a Copilot agent
    • Anything else: a plain-text version you can paste into any chat
    Get more free Lab builds. A new one most weekends, straight to your inbox.

    Let us help your team write prompts that work

    The jobs on this page are made up. The mistakes aren't: invented details, offers nobody approved, risky wording copied straight from someone's notes, the answer buried at the bottom. They happen in real businesses every day, one rushed prompt at a time. We train teams to write prompts and skills that get it right the first time, in the AI tools they already use.

    Hands-on prompt trainingwith your team's real work
    A shared prompt libraryfor the jobs you do every week
    Custom skillsbuilt around how you work
    Guardrails that stickinside IT's rules
    Book a free call

    About this build

    Built with Claude Opus 5.5 on October 31, 2026. We wrote three everyday jobs, each with a simple chat prompt and five prompt pieces (role, context, example, constraints and format), then recorded every combination of the pieces, 32 prompts per job, six times each on four models: Claude Haiku 5.5, Claude Sonnet 5.5 and Claude Opus 5.5 through the Claude API, and gpt-oss 20B, an open model, on our own PC. That is 2,304 recorded replies. From the first pilot run to the last record took about 3 hours and 45 minutes and about 8 million tokens, including two false starts while we tuned the check for made-up claims.

    Each reply is scored with the checks shown on the page, written before the recording. For the customer replies, a separate Claude Sonnet 5.5 pass listed any specific claim the customer would rely on that the facts did not support: a time, a cause, money or a promise. Options left in [brackets] count as not ready to send, not as made up. Across all three jobs, the simple chat prompt passed 64% of checks and all five pieces passed 99%.

    The dealership, the customer, the meeting, the property company and the job are fictional. Every reply on this page is a recorded run of a real model, saved and replayed here; nothing is generated while you watch, and nothing you type leaves the page. Everything shown works in a normal Claude, ChatGPT or Microsoft 365 Copilot account; we used the API only to record and measure.

    More builds in the Lab

    Sunlight shining down through deep blue water

    Find light in the deep waters of AI

    Let's talk! A short call will help you learn more about our training and consulting solutions. We can also discuss whether our existing licensed tools can help, or whether you need a custom build.

    Book a free call