// the toolkit
Small browser tools for people who run AI in production. Each one takes something that is usually invisible — what a prompt change did, what a request costs, whether a model's confident answer holds up, what's in a wall of logs, how strong a secret really is — and puts it on one screen. They open in a tab and work immediately.
// prompt operations
Writing a good prompt once is engineering. Keeping it good as it changes, and knowing what it costs, is operations. These two are for the second part.
Paste the prompt you run today and the edit you want to ship. A word-level diff shows what moved and a summary says what kind of change it is. Add a few test inputs and pass/fail rules, run both prompts against a real model, and anything that passed before and fails now is flagged as a regression before it reaches production.
A real request is four or five pieces — system prompt, retrieved documents, chat history, the new message — sent and billed together. Paste each piece, pick a model, and watch them fill the context window with the cost per call and at a busy day's volume. Drag the window smaller and see which piece is sacrificed first.
Together: run the candidate through Promptwatch to see if it still passes, then through Token Window to see if it quietly doubled the bill.
// trusting the answer
A single model answers everything in the same tone, right or wrong. The signal you actually want is whether independent checks agree.
A live pilot of one idea from the Nexus proposal: route a claim to two independent models that never see each other's reasoning, then have a third adjudicate. When they disagree, you get the one thing a lone model can't give you — a flag that the answer is shaky, before you rely on it.
// security & detection
The same instinct as the hardening and red-team write-ups, pointed at the two places people still get hurt: the secrets they type, and the logs they don't read until something's on fire.
Paste raw log lines — nginx, JSON, syslog, logfmt — and it detects the format, splits every line into fields and shows the top values with counts. Click a value to filter; the other fields re-count. The filter you build is written out as grep, jq, LogQL and awk, ready for a terminal or Grafana.
Most meters check whether you used a capital letter. This one measures what an attacker measures: guesses to crack, once dictionary words, names, years, keyboard walks and leet-speak are charged at what they really cost. Naive bits and real bits side by side, with time-to-crack across four attacker tiers. Recognises real API-key formats and tells you to rotate them.
// for everyone
Two that need no background at all. One uses AI for the thing only AI can do — language. The other deliberately uses none.
Paste a confusing letter, bill, lease clause, insurance decision or terms email. Get back what it says, what it means for you, the questions to ask whoever sent it, anything worth a second look, and who to talk to next. A translator, not an adviser. A button strips your personal details before anything is sent.
Pick what you want to do outdoors and your town. It reads the week ahead hour by hour and marks the best windows for that activity, with the reason, plus the whole week as a colour grid you can overrule. Free public forecast data, no key, no account.
Each of these could have been a service with a sign-up form, a database and a monthly plan. They aren't, on purpose. A tool you'd use at 3 a.m. during an incident, or with a client's prompt on your screen, has to be something you can trust by reading it — and a single HTML file with no backend is the most readable thing there is.
They are also deliberately deterministic. It would be easy to put an "ask the model what's wrong" box on every one. The point is the opposite: these are the parts of AI work where you want exact counts, legible rules and a diff you can argue about in a code review. The model's turn comes afterwards.