Conversation review: write the conclusion after the work is done

In user language: for people who get continuous work done with the same work agent. After a stretch of work is done, you can write a review together: what was finished, key decisions, your requirements, what is still open, and practices you can reuse later. By default it does not nudge you; when you need it, say one sentence, or turn on prompts in the agent settings.

For the user-facing product notes, see [docs/产品介绍.md](/docs/product-intro); for the mechanism source of truth, see [docs/core-mechanisms/对话主题与复盘.md](/docs/mech-topic-review); for how to change the topic, see Conversation topics.

Date: 2026-08-26

The takeaway (read this first)

One conversation will run many rounds. The topic covers “what we are doing now”. Review covers something else: this stretch is done — what is the conclusion.

Review is not written into long-term memory automatically, and it is not compressing the conversation history again. It is a five-column conclusion the assistant drafts, you revise, and you nod before it is kept. Without confirm, nothing is written.

By default it will not pop a button or send an invite right after you finish, or after you pause for a while. Most of the time you are just getting work done; not every stretch needs a review. When you want to keep it, tell the assistant “review this”; if you want this assistant to remind you later, turn on Prompt for review in the agent settings.

In one sentence: review is a shared conclusion after the work is done; by default it happens when you speak, and it only nudges after you turn prompts on.

flowchart TD
  A["A stretch of work is done"] --> B{"Leave a conclusion?"}
  B -->|"say “review this”"| C["Assistant drafts the five columns"]
  B -->|"Prompt for review is on in settings"| D["A button appears after it is done, or an invite after a pause"]
  B -->|"say “skip review this time”"| E["This stretch is not nudged again"]
  D --> C
  C --> F["After you edit, say “confirm the review”"]
  F --> G["The conclusion is kept: later similar topics can refer to it"]
  E --> H["Keep working or change topic"]
  G --> H

Who this is for, and what it solves

It only appears in work-agent, main-site conversation. The help assistant, visitors in the site widget, and scheduled conversations do not do review.

Typical situations:

• After reconciling attendance, you want to leave “which columns we matched, which days count as exceptions, what still needs filling next time”;
• After writing a chapter outline, you want to note “the voice is already humorous in a Jin Yong pastiche, no more characters, the next chapter starts with the duel”;
• The same assistant will do similar work later, and you want it to refer to last time’s conclusion, not treat this round’s details as rules that always hold.

For people getting work done: say review and it drafts, say confirm and it is kept, say skip review this time and it stops nudging this stretch. For people who do not want to be interrupted: quiet by default; prompts are an option, not required.

What review is and is not

Five columns (draft and confirm both follow these):

1. What was finished
2. Key decisions and why
3. Your explicit requirements / corrections
4. What is still open or continues next time
5. Reusable practices (for example skills used, databases already approved)

The topic writes “right now we are reconciling attendance”. The review writes “May is already matched against clock-in and leave; exception days follow the leave form; next month still needs overtime notes”. Both hang on this conversation; they are not one field.

What you can say to the assistant

Say what you want done, and it acts; ask “how / how do I review” and it explains.

You can also tick Prompt for review under the agent Persona → Memory. It is off by default. After it is on:

• At least two rounds, and the assistant judges the work in the current topic is done: Review appears in the composer or menu (it does not start writing on its own);
• This conversation has been idle more than 2 hours and less than 2 days: the assistant invites once in the same conversation.

Inside the window it invites only once. If you speak again before the time, already reviewed, or said skip review this time, it cancels. After 2 days it does not nudge. It counts by duration, not by calendar day.

After a successful topic change, the next stretch may be reviewed. The old review stays; a new topic does not wipe it.

How to see and do it in the UI

1. You speak (default): in this conversation say “review this”. After the assistant drafts, edit then say “confirm the review”, or confirm on the card.
2. Prompts are on: Review appears next to the top-bar topic strip or in the conversation menu. Opening it is the same: draft → you edit → confirm. After confirm the button becomes “View review”.
3. Reading mode: the nudge button can be hidden so reading is not interrupted.
4. Phone: after prompts are on, “Review” appears in the composer; if they are off, it does not take space. Drafts or confirmed reviews can still be viewed.

The invite is an ordinary assistant message, with “start review / skip review this time” as next-step suggestions; it does not open a new conversation. The scheduled-task path that “creates a new conversation each time” does not use review.

What it is for after confirm

Unconfirmed drafts are not used as conclusions. Rolling compression of conversation history is not review; the two are separate: one lets the model keep talking, one leaves you a conclusion you can edit.

When not to force a review

• Only two sentences in, and there is no conclusion yet — write the topic clearly first, or keep going.
• You only rephrased, still circling the same work — change the topic; you do not need another review.
• Preferences you want every future conversation to follow — say “please remember”; do not only write them into this review.
• Asking the help assistant how to use the product — that conversation has no topic and does not do review.
• You do not want to be interrupted — keep the default off; say “review this” when you need it.

Purpose goes in the topic, the conclusion after the work is done goes in review, habits across conversations go in memory, team metric goes in the knowledge base.

Related reading

• Conversation topics — the topic covers “what we are doing now”; one sentence can change it
• Conversation topics and review — the full convention for draft, invite, confirm, and the next stretch after a topic change
• Product introduction — where Messages and work agents sit in the product
• Agent memory — how “remember” splits from “your requirements” in a review
• Memory forging — how archiving carries topic and review
• Conversation history and context compression — a rolling summary is not a review