Under the hood: technical specifications
For the curious and the technical. You do not need any of it to use the product, but understanding the machinery tends to spark the best ideas for using it well.
The shape of the system
The Intent Amplifier is two halves that talk over HTTP.
- •A Chrome extension (Manifest V3, vanilla JavaScript) owns the button, the questions panel, the reveal, and the undo. It holds no business logic and no secrets. A mouth and a pair of hands.
- •A backend service (FastAPI, mostly Python standard library) is the brain. All prompt engineering, gating and orchestration live here, along with the keys.
Keeping all the intelligence server side is what lets the same experience work identically across every supported site, and lets the product improve without anyone updating their extension.
The engine
The model doing the generation is Llama 4 Scout, with the small side decisions pinned to a sharper Llama 3.3 70B, served over an OpenAI compatible interface on an always-on server, so there are no cold starts. The whole setup is deliberately swappable. Very little of what the product does well comes from the model. Most of it comes from what the surrounding code refuses to let the model do.
The library, also called the vault
This was once the heart of the product, and it is now the part we are most careful about. It is a structured library of expert knowledge kept as a set of interconnected notes, holding 59 distinct experts across an enormous range of domains:
Each expert is broken into small judgment units: a named move, the situation it applies to, and the reasoning behind it. Across the library there are 1,224 of these.
None of this reaches your prompt. In 108 blind comparisons, prompts carrying a unit from this library were preferred 28.7% of the time; a random stranger's personal context scored 26.2%. Injection has been switched off since then, behind a flag that can turn it back on if the evidence ever changes. The routing half still runs. Full detail here.
The library also holds a layer of tension notes. Real experts navigate tradeoffs, and often two of their own principles pull in opposite directions: charge a premium versus stay accessible, ship fast versus build to last. A tension note captures both sides and the condition that decides between them. Adding or improving the library is a matter of editing notes, not writing code, which is why it is cheap to keep while its usefulness is still an open question.
The pipeline
When you tap Amplify, your request flows through a small number of stages.
- 1.Clarify. A deterministic check decides whether to ask you anything, with a grounding rule that keeps questions tied to what you actually wrote.
- 2.Understand and route, in one pass. The model reads your request, returns a structured picture of your intent, and picks the closest specialism from a lexical shortlist, all in a single call. Doing these together, rather than as two round trips, is a large part of why the engine is faster now. The shortlist acts as a domain prior that stops the model drifting to an adjacent but wrong one.
- 3.Preview. Before anything is written, a short read names what you are actually trying to get and the shape of the request about to be made on your behalf. You can correct it, and a correction is folded into what both later stages see, so it binds rather than merely being noted.
- 4.Author, then carry. The model writes the final prompt under the invisibility rule and a firewall that forbids it from answering the task itself. Afterwards, a set of deterministic carries paste your own message and your own answers back in verbatim, with no model call at all. Two model rewrites of the same idea were tried first and both landed in the noise: asking a model for one clean prompt is asking it to compress, so preservation has to be a copy.
- 5.Show improvements. Alongside the rewrite, it produces the plain note of what it sharpened, so the note always reflects the real changes.
A guiding principle runs through all of it: branch in code, generate with the model. Any real yes or no decision is made by deterministic Python, not by asking the model to decide. The model is used for what it is uniquely good at, understanding language and writing well, and kept out of the binary calls where reliability matters more than flair.
Privacy and data
This is a design commitment, not a footnote.
- •The extension holds no account and no keys. You sign in only on the web, and the extension links itself afterward.
- •Storing what you type is off by default. A fresh install keeps nothing on our servers. Switch on data sharing, or sign in so you can see your own history, and a record of each amplify is kept, which does include the text. You opt in, not out, and either switch reverses.
- •Nothing you type is trained on, ever, and nothing about you is inferred and asserted back at you without you seeing it first.
- •Anything the product remembers about you lives in your own browser first, and you can read, edit or wipe all of it from the dashboard.
The short version: the product was built without needing your data at all, which falls straight out of the one idea. What we do hold, and exactly when, is set out item by item in the Privacy Policy.