1. Extract the transcript.
- Keep only the person's text and the AI's text. Reduce each tool call to a one-line marker, and drop tool results.
- Do not summarise or rewrite. The transcript is the evidence.
- Find the start by its position, or by the person's own turn. A phrase can be quoted back later in the conversation, and a script that looks for it then starts in the wrong place.
- Cut the transcript into threads. A thread is one subject the person kept coming back to. A single message is too small to judge on its own, so the thread is what you score.
- Split a turn that holds an acceptance and something else. "Okay, drop it. So what counts as done?" is two claims. Left as one line, the question decides the score and the acceptance is lost.
2. Pull out the person's own words. A fresh agent reads only the transcript. A fresh agent is a new AI session with no memory of the conversation. It quotes the person and does not interpret. It pulls out:
- what they claimed
- what they asked for
- what they rejected
- how they framed the subject
It tags each claim with who said it: the owner, someone relaying for them, or another participant.
Text the person pasted from elsewhere sits inside their turns, such as another person's messages or another chatbot's answers. Tell the agent to quote only the person.
A dictated line can come out garbled, or phrased as a question. That is not an unsettled thought. The writer in step 4 turns it into a plain statement using the person's key words.
3. Score each claim against a purpose. Write one sentence saying what the output is for. Without it, nothing separates what is relevant from what is not. Then Jev labels each claim against that sentence. Jev is a decision model: it picks one option from a list you give it, and says how sure it is.
- Hand Jev one claim at a time, with the speaker tag written in front of it, and nothing beside it.
- Score an example together with the claim it illustrates. On its own, an example reads as a fragment and scores as not belonging.
- Write the labels for each purpose. An example set is below.
- Jev's label stands. Nobody picks again.
- You assign the ruling label yourself, from the speaker tag, before Jev runs. A ruling is the owner's own words settling something, in any tone: an instruction, an acceptance, a choice. It sits above every score. Left to Jev, a firm instruction from someone relaying outscores the owner's casual yes.
- Jev still scores a ruling, but only to see whether it belongs to this purpose.
- An instruction from someone relaying, or from another participant, is context, however firm.
If Jev is not available, a fresh agent that sees only the claims and the purpose can pick the labels. The three sessions stay separate, but a language model makes the picks.
4. Write from the scored claims. A second fresh agent, which never saw the conversation, writes the output. Give it:
- the scored claims and the person's own framing
- who the output is for
- the things the person rejected, which are the strongest signal
- any position the person agreed to when it was put to them but never said in their own words, labelled as that
- what the person receiving it has already been told
Tell it to leave out jargon, names only your own system uses, and your own interpretations.
5. Check the output. Hold it against this list, and have the writer redo any sentence that fails:
- it answers the question first
- every required section is there
- every sentence traces back to a claim
- nothing the person rejected appears
- no excluded term appears
- the person's own framing leads
- no closing line reads as a slogan
To check clarity, score each sentence on named levels, such as "a teenager could follow this", "one term needs explaining" and "the paragraph needs rewriting". Use levels because a yes or no answer about clarity follows how technical the subject is, whatever the writing.
The labels. Write them for each purpose. This set is for the purpose "what the person designed for how their system should work, in their own words":
belongs: it states a rule, principle, design choice or definition about how the system should work, even if said casually or with we, they or us.too_personal: it is about the person's life, feelings or schedule, or a one-off request, and says nothing about how the system should work.question_not_claim: it only asks or wonders, and states no rule, principle or choice.
The words "even if said casually" are in belongs because without them, a claim said with "we" or "they" scores as too personal.
When not to use it. Skip it when the person said little about the subject, or has not made up their mind.