Personalization layer for AI products

AI that learns how each user wants to be answered.

Your users already tell you how they want to be written for: they rewrite your drafts, they say “too long”. Prepon turns that into a rule, proves it against their past complaints, and applies it to every reply after.

prepon — draft

Scripted walkthrough in five beats. A writer asks for a draft, the model returns a long hedged one, the writer rewrites the opening sentence and says "too long, verdict first", and the agent turns that edit into the rule "give the verdict first", replays it against fourteen past complaints before it takes effect, and applies it to the next reply.

Scripted walkthrough. No live model calls.

How it works

Pushback, rule, replay, live, decay.

One edit, followed all the way through. Each step carries its own scripted frame.

  1. You rewrite a sentence, or say “too long”. Prepon keeps the before and the after.

    Both versions stay on the draft, so whatever comes next can point at the exact sentence it came from.

Draft — reply to Priyaedited just now

Happy to help! I can walk you through the options if that would be useful. There are three options.

You, rewriting it

There are three options.

Both versions kept, side by side.

Scripted scene. No account, no live data.

Why this is not a prompt wrapper

Two rules about the rules.

Rules have to earn their place.

A rule does not go live because you made an edit. It is replayed against the complaints you already made — same prompts, same rewrites — and what comes back is a count, not a claim.

Replayed against 14, fixed 11. That fraction is the only reason it gets near your next reply, and all 14 stay listed so you can open one and disagree.

Replay — rule 07before promotion
11/ 14

complaints it would have fixed

Passes. All 14 stay listed, and open to the draft they came from.

Scripted: the replay readout a rule has to pass.

Rules expire.

Every live rule carries three counts: how often it applied, how often you kept the result, how often you reverted it. A few reverts and it is under the line within days.

Only 60 rules stay live. A new rule gets in by being healthier than the weakest one there, which retires and stays readable, with the edit that created it.

Rule 22 — nine days live60 / 60
applied
41
kept
28
reverted
13
health78 → 12
below the lineRetired

Scripted: a rule losing health and retiring.

What would make us wrong

The number that would kill this.

The baseline is not a stripped-down model. It is the same frontier model handed a style guide the writer wrote themselves, in one sitting, then frozen.

Same prompts, same model, two sets of instructions: one written by hand, one learned from edits. Both answers go into a blind pair and the writer picks. If Prepon is not winning more than 60% of those pairs after a hundred captured edits, the rules cost more attention than they return, and we say so.

Win rate against blind pairs as captured edits accumulateSample data. Win rate starts at 47 percent with no captured edits and rises to 67 percent at one hundred captured edits, crossing the 60 percent kill line at roughly seventy.80%40%60% — kill line050100

Edits captured

Sample data, not a measurement. It shows the shape of the test: win rate in blind pairs crossing the 60% line somewhere around seventy captured edits — or the rules do not earn their keep.

No users yet. No revenue. The test above is the one we intend to run, not one we have run.

Questions

The parts people push on.

A system prompt is one block of text you maintain by hand and never find out whether it worked. Prepon writes each rule from a specific edit, replays it against complaints you already filed, and only then lets it touch a reply. The prompt is the output; the replay gate and the expiry are the product.

Rules are plain English sentences, not weights or a fine-tune, so they move to whatever model you point them at. On a model change the replay runs again, and a rule that no longer changes the output drops out of the live set. What you keep is the record of what you asked for, not an artefact tied to one version.

You do. Your sentences, readable as a list, exportable as text, deletable one at a time or all at once. Nothing is pooled: a rule learned from your edits never touches anybody else's replies.

Memory stores facts about you and retrieves them later — your dog's name, your stack. Prepon stores instructions about delivery: sentence length, hedging, what order the answer arrives in. They need different gates: a wrong fact is a bug, a stale style rule is only irritating, so one is retrieved and the other has to survive a replay first.

Every rule shows the edit it came from and the complaints it was replayed against, so you can read why it exists. Delete it and it stops applying immediately. Or keep reverting what it produces — its own counts drop it below the line and it expires on its own.

No. Prepon changes how an answer is said, never what it claims is true. If the model is wrong about a fact, Prepon will make it wrong in your voice. It is a writing instrument, not a fact checker.

Watch the demo one more time.

There is no product to sign up for yet. The scripted scene above is the whole claim.