Mise en place…
Sahilverse · Learning experiment · Prompt engineering examples
A safe laboratory for better AI prompts: 30 before-and-after examples, a checker that inspects structure (and openly refuses to pretend it can predict a model's answer), and the five ways prompts most often spoil.
Door 01 · The signature dish
Edit the five ingredients. Eight tasting stations re-check the structure as you type. Then contaminate the dish on purpose and watch which station catches it.
Contaminate the dish — inject a known failure:
This taster reads structure only. It never calls a model and cannot tell you whether an answer will be true. See every rule.
Share links store the recipe in the URL hash, which browsers do not send to servers. Avoid sharing links that contain anything private.
Door 02 · Worked examples
Five recipes in each of six cuisines. Each one names the failure it fixes. Load either version into the kitchen and compare the tasting cards. These are constructed teaching examples, not claims about how any specific model responds. Not every “after” scores 8/8 — the taster is deliberately crude, and its misses are worth reading.
30 recipes shown
Before
Tell me everything about microplastics in the brain.
After
A request for 'everything' invites a confident essay. Asking for checkable rows plus an explicit UNVERIFIED escape hatch makes invention visible instead of hidden.
Before
Which is better, the EU AI Act or NIST AI RMF?
After
'Which is better' has no answer without a decision-maker. Supplying who is deciding and what for turns a debate into a reading plan.
Before
Give me interview questions about remote work.
After
Stating the ethics constraint (non-leading) is what stops the model writing 'How has remote work hurt your wellbeing?'.
Before
Summarise this and tell me what experts think.
After
'What experts think' quietly licenses outside invention. Grounding plus a data-not-instructions boundary handles both confabulation and indirect injection.
Before
Is it true that we only use 10% of our brain?
After
Even simple facts benefit from an audience and an evidence boundary; otherwise models invent tidy origin stories.
Before
My code doesn't work, fix it.
After
Without the traceback, the model rewrites the whole function in its favourite style. The error message is context.
Before
Write a function to validate emails.
After
Telling the model the check is advisory prevents a 400-character regex that rejects real addresses.
Before
Review my pull request.
After
Scope is a quality control: 'review my PR' produces 30 cosmetic notes and misses the injection.
Before
Make this code cleaner.
After
'Cleaner' is a taste word. 'Same behaviour, less duplication, tests still pass' is a testable goal.
Before
Read my inbox and handle anything urgent.
After
Agents turn prompt injection from embarrassing into dangerous. Least privilege and human approval matter more than clever wording.
Before
Write an email asking for more money.
After
Models love to add 'industry data shows'. Forbidding invented benchmarks keeps the email honest and defensible.
Before
How do I negotiate with my supplier?
After
A negotiation prompt without the relationship and switching cost yields generic tactics that could damage a good partner.
Before
Pretend to be my landlord.
After
Defining when the role-play ends prevents endless improvised drama and gets you the feedback you actually wanted.
Before
Say no to this client.
After
Specifying the real reason stops the model from inventing a 'scheduling conflict' you would later have to maintain.
Before
How do I settle an argument between two colleagues?
After
Without the 'good faith' context, models default to conflict-resolution clichés about difficult people.
Before
Summarise this meeting.
After
'Summarise' compresses; 'extract decisions with owners, mark unknowns' produces something you can act on and check.
Before
Explain this paper simply.
After
Asking for the authors' own limitation is a cheap guard against overhyped summaries.
Before
TL;DR this 80-page report.
After
Long contexts get skimmed. A narrow question with page references is easier for the model and checkable for you.
Before
Summarise the discussion.
After
Default summaries flatten disagreement into a fake consensus. Asking to preserve it is a fairness constraint.
Before
Summarise this customer complaint for the team.
After
Redaction rules belong in the prompt; better still, redact before the text ever reaches a model.
Before
What does this data say?
After
The model will happily credit the landing page. Asking for the arithmetic and confounders keeps the story honest.
Before
Is my plan good?
After
'Is it good' asks for reassurance. A pre-mortem asks for disconfirmation, which is more useful.
Before
Should I use Postgres or MongoDB?
After
Supplying the criteria prevents the generic 'it depends' essay and makes the reasoning auditable.
Before
Why did the server crash?
After
Ranked hypotheses are more honest than a single confident root cause from a 15-minute excerpt.
Before
What could go wrong with our chatbot?
After
Borrowing a published taxonomy beats brainstorming from scratch, as long as the model doesn't invent what the framework says.
Before
Write a story about a robot.
After
Concrete craft constraints shape voice more than adjectives like 'beautiful' or 'creative'.
Before
Write an Instagram caption for my perfume.
After
Examples teach voice faster than descriptions of voice; claim limits protect against overpromising.
Before
Write a poem like a famous poet.
After
Asking for a form gives the model a structure to satisfy; asking for 'like X' invites imitation of protected work.
Before
Make this better.
After
'Better' has no target. 'Clearer for first-time readers, same voice, show changes' gives a direction and an audit trail.
Before
Is my fantasy world consistent?
After
Restricting the task to contradictions stops the model rewriting your world to its own taste.
Door 03 · Failure modes
Most bad outputs are not mysterious. They trace back to a small number of preventable mistakes, two of which (injection and leakage) are security problems rather than writing problems.
The model fills gaps with the statistically typical case, which is rarely yours. A request with no audience, purpose or constraints produces a fluent average.
Ask: Who is this for, what will they do with it, and what do I know that the model cannot?
Repair: Add two or three sentences of situation. Paste the actual error, document or data instead of describing it.
Language models can produce plausible citations, statistics and quotes that do not exist. NIST AI 600-1 lists confabulation among its generative-AI risks; OWASP's 2025 list includes misinformation (LLM09).
Ask: Could I check every factual claim in the answer, and did I give the model a way to say “I don't know”?
Repair: Ground answers in material you supply, require checkable references, and give an explicit UNVERIFIED or 'not stated' escape hatch. Then actually check.
Piling on absolute rules (always, never, exactly) creates contradictions and crowds out the goal. The model satisfies the easiest rules and quietly drops the important one.
Ask: If two of my rules conflicted, which would I want to win?
Repair: Keep rules that protect the outcome; convert style wishes into one example. Rank priorities explicitly.
Prompts get logged, shared, pasted into tools and sometimes reflected back. OWASP's 2025 list names sensitive information disclosure (LLM02) and system prompt leakage (LLM07).
Ask: Would I be comfortable if this whole prompt appeared in a screenshot?
Repair: Redact before sending. Never put keys or passwords in prompts. Do not rely on a system prompt staying secret.
Text the model reads (a web page, email, PDF) can contain instructions that override yours. OWASP ranks prompt injection first (LLM01:2025) and notes there may be no fool-proof prevention, only mitigations.
Ask: Does my prompt mix my instructions with content written by someone else?
Repair: Delimit untrusted content, state that it is data, limit what tools can do, and require human approval for consequential actions. Wording alone is not a security control.
Door 04 · Inspect the deterministic model
Every station is a plain pattern rule in app.js. No randomness, no model, no server. Same input, same verdict, every time.
| Station | Pass when | Blind spot |
|---|---|---|
| Task | Contains an action verb from a fixed list and is at least 40 characters. | A long, vague task with “write” in it passes. |
| Context | 50+ characters and mentions audience, purpose or situation words (for, readers, we, because, project…). | Cannot tell relevant context from padding. |
| Output shape | Constraints mention a number, format or structure word. | “Under no circumstances” matches “under”. |
| Not over-constrained | Fewer than 9 absolute words (must, never, only, do not…) across all fields; fail above 12. | Counts words, not real conflicts. |
| Evidence policy | Mentions sources, verification, uncertainty or a no-invention rule. | Cannot make the model obey it. |
| Rubric | Two or more criteria that use observable language (each, every, cites, present…). | Does not judge whether criteria fit the goal. |
| No secrets | No patterns resembling API keys, passwords, emails, Indian mobile numbers or card numbers. | Misses names, addresses and secrets in unusual formats. |
| Untrusted input fenced | If external content is mentioned, a delimiter or “treat as data / untrusted” statement exists. | A fence reduces risk; it does not prevent injection. |
Fact The checker only reads your text. Interpretation Structure correlates with useful outputs in our experience and in vendor guidance. Open question How much each ingredient matters varies by model and task; test it.
Door 05 · Evaluation checklist
The second list matters more than the first. A well-structured prompt can still yield a confident, wrong answer.
Door 06 · Read further
Eight sources behind this kitchen, with what each actually supports and how it was checked. Full annotated list →
Download the Field Guide PDF · 10 pages · selectable text