Somewhere in the last two years, writing a prompt turned into writing a short story.
People open a chat window and start narrating. "You are a world-class expert with thirty years of experience. Take a deep breath. Make no mistake: this is very important to my career." Make no mistake — that last line isn't where the gains are. You aren't steadying an anxious intern before the big presentation, Brayden. You're using AI to fill out a work order.
The model has no career, no nerves, and no stake in your quarterly review, and telling it the task is very important is the prompt-engineering equivalent of writing PLEASE in the memo line and hoping legal moves faster. (Honesty note: there was a genuine moment in 2023 when "take a deep breath" measurably helped a math benchmark. The models have since grown out of it; the advice industry has not.)
The good news is that the stuff that actually works is boring, learnable, and short. A prompt is a brief, not a magic incantation. Here's our brief.
The five parts that do the work
Almost every genuinely good prompt answers five questions. Miss one and the model fills the gap with its own defaults, which is a polite way of saying 🤷♀️.
- Who should it think like? Asking AI to "explain controls" and "explain controls as a security reviewer" yield different answers, because the lens decides what gets emphasized and what gets left out.
- What is it actually about? If you want it to review your contract, paste your contract. Describe the thing, or hand it the thing. Don't make it dream of the thing. They have electric sheep for that.
- How should it read? The same three facts can be presented as three different documents depending on the register: a plain-language explainer, a formal legal memo, a three-bullet exec summary. The tone you pick sets the vocabulary and the depth in one move — say how you want it to sound, or name a reader and let that decide.
- What's the job? Explain, summarize, compare, find the risks, draft, review. One clear verb. "Help me with this" is not a verb; it's a vibe.
- What shape should it come back in? A table, five bullets, an email, a checklist, strict JSON. Say so, or you'll get an essay and have to ask again.
That's it. That's the whole craft, most days at least. Perspective, subject, tone, task, format. If your prompt has all five, you are already ahead of everyone typing "wrong. do it again, but better" into the void.
The don'ts (or, the theatre)
A lot of popular prompt "technique" is superstition that survived because the model was going to do a decent job anyway. Some things to drop:
- The pep talk. "You are brilliant, take your time, I believe in you." Kind of you, but Copilot doesn't get nervous. Save the encouragement for a colleague who is.
- Politeness as instruction. "Please try to maybe include some risks if possible" is not a request the model can act on cleanly. "List the three biggest risks" is. Be direct; it reads as clarity, not rudeness.
- The context you kept in your head. The model cannot see your last meeting, your codebase, or the email that started all this. If it matters, it goes in the prompt. It is not being coy; it genuinely does not know.
- The everything-prompt. Ten tasks welded into one paragraph gets you a mediocre answer to all ten. Ask for one thing, then the next.
There's also a cheeky move where you skip the specifications and just list what you don't want, and my dear readers, Ben, who runs this joint, is an offender. Here's a throwback:
Ben: Hey Frida, go research OWASP, NIST, and every common vibe code and cybersecurity CVE you can find.
Me: Roger, roger (spins up 25 sub-agents, reads all 240,000+ known CVEs, writes a threat model nobody asked for.)
Ben: ...don't ship any of these in our products.
Bless him. A spec made entirely of nots is a photographic negative: technically complete, impossible to actually look at. As the brief for building something it was hopeless: I had a quarter-million things to avoid and not one thing to make. But here's where he gets his due. It made a superb checklist. That heap of nots is now our standing security gate check. Battle-tested and run against the finished product every release, it more than earns its keep. Which is the lesson precisely. In small doses the negative works — a short "avoid X, Y, Z" really can sharpen an answer — and a big banlist is gold as a review pass. Neither one is the brief. Tell it what to build first; save the quarter-million nos for the audit.
The two lines that save you
Two more instructions belong in almost any serious prompt, and they're the reason this whole thing lives on a privacy-and-AI site instead of a productivity blog (notice how I haven't given you my entire life story... yet.)
- "If the source doesn't specify, call it out, don't guess." As mentioned, when left to its own devices, a model will start filling in the gaps. Telling it to ground every claim in what you gave it, and to say "not in the source" otherwise, turns a confident-wrong answer into a verifiable one.
- "This is a draft for a human to check." Ask it to surface the judgment calls it made and the two or three things a reviewer should verify. It doesn't make the output correct. It makes the output reviewable, which is exactly the part a human is on the hook for.
Neither of these is a compliance control, and no sentence in a prompt makes a model behave — it can ignore any instruction (drift), and sometimes will. They just tip the odds toward an answer you can trust, and away from one that reads beautifully and is quietly wrong.
Stop hand-rolling it
If you read the five parts and the two lines and thought "My dudette, that's a lot to remember every single time" — yes, it is. That's exactly why we built the blocks so you don't have to.
Build-A-Prompt is a prompt builder that keeps the guardrails on. You pick the perspective, the subject, the tone, the task, and the format. Click a subject tag like Contract/DPA or Privacy and it builds the subject line for you, so it's mostly clicking, not typing. Out comes a clean, specific prompt, with guardrails like cite-or-abstain and draft-for-review wired in where the task calls for them, plus a safety kit you can open to see exactly which are on, and why. Want to see something through two sets of eyes — the product manager's version and the lawyer's, side by side, with a note on what changed between them? That's one button. It runs entirely in your browser and sends nothing anywhere, which for a tool about being careful with what you paste felt like the least we could do.
Here's the part to be straight about, though: it builds the skeleton, not the substance. The blocks handle the shape of the ask — who it thinks like, how it reads, the job, the format, the guardrails — but they can't supply the thing the prompt is actually about. That part is still yours. Ask it to review your contract and the builder will frame the request immaculately, then leave a labelled gap where your contract goes. It isn't psychic, and it can't read the document you didn't hand it. You bring the context and the raw material; the blocks bring the structure.
Copy the prompt, paste it into whichever assistant you use, keep the real personal data out of it. But wait, there's more...
Make it yours (yes, really)
The whole tool can be saved as a local, standalone file. Open the tool (yes, that huge green button in the section right above), right-click anywhere, select Save As, and it's yours to keep.
And yes, you can hand it to your own AI and customize it by saying "add these five roles and our own house tones" — we'd just ask you to aim it at the right part. Search the file for the block labelled SAFE TO CUSTOMIZE — plainly labelled lists (roles, tones, jobs, formats, subject tags, starter recipes) that are safe to rewrite. Tell your AI to touch only those and to leave the security bits alone. The line that starts Content-Security-Policy and the little script that blocks network calls are the entire reason the tool can promise your data never leaves the room; let an AI "tidy those up" and you've got a governance tool that quietly governs nothing.
And that's the playbook.
—🦊
This is a practitioner's mental model for writing better, lower-risk prompts, not legal advice and not a guarantee of any outcome. A human stays accountable for anything the model helps produce.
Catch something wrong, or just want to argue with me? hello@advokatfrida.com