Examine a disputed claim
Weak: “Tell me if this is true.” Better: ask for the precise claim, supplied evidence, alternative explanations, uncertainty and a verification checklist.
Enable JavaScript to browse and load all 30 authored cases.
A free experimental workshop / edition 01
THE PROMPT
TEST KITCHEN.
Better prompts are better specifications. Define the task, provide context, constrain the output, set an evidence policy, and evaluate the result. No secret incantations required.
Private by default. No login, model API, cookies or tracking. The lab checks prompt structure—not the truth or quality of any AI answer.
RECIPE / 001TEST → REVISE → REPEAT
NOT MAGIC ✳
INGREDIENTS ≠ GUARANTEES • EVIDENCE BEFORE ELOQUENCE • TEST BEFORE TRUST
01 / The working bench
Build a prompt.
Inspect the recipe.
Change any ingredient. The transparent checker updates instantly. Its score is an editorial checklist of present fields—not an AI performance metric or prediction.
THE STRUCTURAL CHECK
Five checks, one point per filled ingredient. Not a measure of answer correctness.
5/5
- Task present
- Context present
- Constraints present
- Evidence policy present
- Output specification present
ASSEMBLED PROMPT
Your recipe appears here when JavaScript loads. Without JavaScript, the full method and examples below remain available.
WHAT THIS CANNOT TELL YOU
A complete-looking prompt can produce a false answer. A short prompt can outperform a long one. Real evaluation requires running the prompt with a model, comparing outputs and checking evidence.
02 / Recipe archive
30 real-world prompt makeovers.
Six disciplines. Five cases each. Select a card to load its stronger version into the workbench. These are authored teaching examples, not measured model results.
03 / Failure lab
Where good-looking prompts fail.
Invented sources
Asking for citations does not guarantee real citations. Require verifiable source details and independently open the originals.
Overconstraint
More rules can make a task brittle. Keep the constraints that protect meaning, then compare outputs with and without the rest.
Prompt injection
Instructions hidden in retrieved documents or websites are untrusted data. Do not let them override the user's task or disclose private information.
Private information
Remove unnecessary customer, employee or health details before using external AI tools. This website does not transmit your inputs.
Ambiguous success
“Make it better” gives no testable target. Specify who the result serves and how a human will judge success.
False precision
A structural checklist can help spot omissions, but it cannot predict correctness, fairness or usefulness. Test with real tasks.
04 / Keep learning
Read further. Think wider.
Methods and primary references
OpenAI · Prompt engineering guidance ↗Anthropic · Prompt engineering overview ↗Google · Prompting strategies ↗NIST · AI Risk Management Framework ↗OWASP · LLM application security risks ↗White et al. · A Prompt Pattern Catalog (research paper) ↗These are reference destinations, not proof that any prompt technique always works. Check their current versions before applying advice.
Explore related worlds
Sahil Tayal · The canonical atlas and wider research ↗Sahilverse · Browse connected experiments and writing ↗Vibe Coding · Apply specifications and testing to generated code ↗SYNTH · Explore the wider thinking and synthesis project ↗These links are editorial routes to adjacent ideas; no automatic recommendation engine or hidden tracking.
Send feedback to Sahil ↗05 / Take it with you
The field guide is yours.
Ten pages of prompt design, before-and-after cases, evaluation protocols, failure modes and offline exercises. You can also export the exact prompt you create above.
Download the Field Guide ↓Published 26 October 2026 · Static educational tool · No live AI model · Editorial review recommended when underlying model behavior changes.