---
title: "Decide by conversation"
description: "Answer your queue of decisions and approvals in a conversation, by voice, dictation or text: one question at a time with its recommendation, a read-back of what the agent understood that writes nothing until you confirm, a queue plan for a backlog of 200, keep mine while away, and principles learned from your own decided answers, cited, kept by you and never shown to the panel."
url: https://frazer.ai/docs/source/decisions/conversation
index: https://frazer.ai/docs/source/decisions/llms.txt
tokens: 5816
---

> For agents: this page is indexed at https://frazer.ai/docs/source/decisions/llms.txt, and every page at https://frazer.ai/llms.txt. Copy commands from this file exactly.

# Decide by conversation

Every list of things a person must settle (Needs you, a plan’s open decisions, an approval’s decision rows) has two modes: **Cards**, for overview and quick keys, and **Chat**, where an agent walks your queue with you one item at a time. It asks each question in plain words with its recommendation, takes your answer by voice, dictation or text, reads back what it understood, and writes nothing until you confirm. Each answer can raise the next question, and what you keep telling it becomes your **principles**, so later recommendations fit what you expect.

> **The short version:** **Talk it through** on Needs you, or `fir decision talk` in a terminal. One item a turn, the recommendation first, your words kept verbatim. “What I understood” lists every item your answer touched; `↵` confirms it. A backlog opens with a queue plan you accept or change in one answer. Principles are learned only from your own decided answers, used only after you keep them, and never seen by the panel.

## One turn

| Step | What happens |
| --- | --- |
| **Pick** | The next item by what it blocks, its risk, its due date and what its answer unblocks; decision groups first, because one answer covers many; an item another waits on before the one that waits. Never ordered by a predicted answer, and the turn says why (“blocks 14 items”); you may reorder by voice (“start with the tax ones”) |
| **Ask** | The question in full words, why it matters, the recommendation with its basis (“2 of 3 · 74 %”), the crux and its fetched fact, what each option binds, and your principles line when one applies; at most 50 words for a unanimous two-way item, 110 for a split or an R2 or higher decision; the full card is one key away (`O`) |
| **Answer** | A key (`A`, `1` to `4`), a word (“yes”, “B”, “later”, “why?”), or free text, typed, dictated or spoken; never cut short and never rewritten |
| **Read back** | “What I understood” (below) |
| **Confirm** | `↵` or “yes” for R0 and R1 two-way items, with 5 seconds to undo a voice confirmation; the chosen option’s key again for a split and for R2 and above; your passkey at R3 and on one-way doors, as on the card. A spoken value (a number, an option, a name, a date) the recognizer was unsure of raises the confirm one level and is spelled back; text from another dictation tool shows values as digits and exact names. |
| **Write** | The card’s own writes in one transaction, recorded with the conversation, the words and the confirmed reading; follow-ups raised from your answer |
| **Next** | What the answer changed (“D9’s recommendation now reads …”), then the next item; a later item whose evidence changed is asked again with the change named |

`skip` and `later` leave an item open (nothing is decided by silence); `why?` shows the evidence and the dissent and never argues past them; `D` defers to a date or an event; “Talk through this plan” enters the [planning session](https://frazer.ai/docs/source/planning-flow#planning-by-conversation) on the item’s plan and comes back.

### What I understood

The agent turns each answer into a **reading**: for every item the answer touched, the verdict (`accept`, `choose`, `change`, `discuss`, `defer`, `skip`), the option, your words as verbatim spans, what it would write, the **follow-ups** it found (a new decision row, a question for an agent, a [person task](https://frazer.ai/docs/source/person-tasks), a condition on the effect), the principles it noticed, the dictionary’s corrections, and any sentence it could not place. One line per item, at most three facts a line, only what differs from the recommendation highlighted (product UI, demo data):

```text
diego (dictated, 61 words, 15:07): "Hold them forty eight hours. Merchants hate surprise clawbacks more than they hate
waiting, and support said a release within two days keeps tickets flat. But tell me what it costs support if we go to
seventy two later. And put the hold on the payout screen, not just in the API."

What I understood                                                     ↵ confirm with 1 · E edit · N not quite
  #406 §D5  → A · hold reversals 48 h · your words: "Hold them forty eight hours. Merchants hate surprise clawbacks …"
            writes  constraint: reversal hold 48 h on services/payouts/holds/** · must · review by May 17
  + new decision for you   #406 §D12 "Show the hold on the payout screen?" (raised from your answer; recommended yes)
  + question for an agent   support cost of a 72 h hold, per 1,000 reversals (answers in about 10 min)
  + noticed (preference)    "Hold money rather than claw it back from merchants" · keep? (end of the conversation)
  corrected  "pelvin" → Pellwyn (organization dictionary)
```

Before a reading is shown, deterministic checks run: every reference exists and is open; the verdict is allowed for that item (no accept without a recommendation, an option that exists); every quoted span is a substring of what you said, so your words are never paraphrased; every sentence is placed or listed as not placed; every number in the reading appears in your words or the item; nothing raised duplicates an open row. A vague answer (“sounds good”) to a split, an R2 or R3 decision or a one-way door is never read as a choice: the agent asks “A or B?”.

## Voice, dictation and text

| Input | How |
| --- | --- |
| **Dictation into the box** (the default for long answers) | A plain, focused text box that takes any dictation tool; letters type, shortcuts need `⌘` or a blurred box, and `⌘↵` submits; the card’s names stay on screen beside it, so a dictation tool that reads the window spells them right |
| **Push-to-talk** | Hold `Space` in an empty box, or the microphone on a phone; audio goes to a model on the device where it has one, or to the gateway’s speech route under your organization’s data policy; recordings append, with no time limit, and the item’s names and IDs are recognition hints |
| **Spoken conversation** | The question spoken, full duplex, with a patient end of turn so you are never cut off while you think; long answers by push-to-talk or “done” |
| **Keys** | `A` accept · `1`\-`4` choose · `W` write · `D` defer · `S` skip · `L` later · `?` why · `O` card · `B` back · `↵` confirm · `E` edit · `⌘↵` submit |

Audio is not kept unless your organization turns it on. The **dictionary** maps what was heard to what was meant (“FHIR” → Fir, “pelvin” → Pellwyn), personal or for the organization; it grows only from corrections you make in a reading, and it corrects the reading, never your words.

## At 5, 29 and 200 items

-   **Five** (tomasz at Acme, Mon Oct 19 at 09:05, session `s-7c14`): plan #463’s four decision rows and then its approval, straight through, in about six minutes.
-   **Twenty-nine** (diego at Pellwyn, Tue Nov 17 at 15:00, session `s-2f90`): two days of one colleague’s rows after a day off. The first turn is a count line and an order: “29 need you (34 decisions): 9 OWNERS-group duties ready together within your authority (1 act), 2 decision groups covering 7 decisions, 18 one at a time (3 splits, 1 R3, 4 approvals’ rows). I’d start with #406 §D5: it blocks 214 payouts.” 29 rows become 21 acts, inside the budget of 25.
-   **Two hundred** (priya, the same moment, session `s-9b47`): a **backlog** after four days away, kept by **keep mine**. Any queue of more than 50 rows opens with a **queue plan**, one proposal you confirm or change in one answer:

```text
Decisions · Chat · priya · Tue Nov 17 15:00              0 of 48 today · back from four days away
200 need you (223 decisions) · four days of your rows, kept by keep mine (your day is about 50)

Frazer  A plan for the 200:
        1. 9 are moot: older daily reviews of cm-7a10 and cm-7f61, covered by the latest 3. I'll close them.
        2. 34 others may decide: 22 OWNERS-group duties and 12 four-eyes reviews, to three colleagues.
        3. 157 are yours, grouped by cause into 57 asks: 22 decisions (5 groups cover 30), 14 Reviewer escalations,
           6 conflicts by plan pair, 4 outside-terms causes, 3 campaign reviews, 8 reverts.
        4. 45 asks today, 12 tomorrow at 08:30. #418 §D3 (R3) first.
        Go, or change it?
priya   "Go. But take #409's two tax ones first; kenji is waiting on them."
Frazer  What I understood: plan accepted · order: #409 §D7 and §D8, then #418 §D3 · 34 handed over · 9 closed
        · 45 asks today, 12 tomorrow 08:30.
```

Each group answered once becomes a standing decision that answers its question in every plan that raises it again, so the next day’s queue is smaller. On a phone, Chat is the queue’s form: push-to-talk, R0 and R1 answers with one tap, R2 and above with the passkey, and batches only on a larger screen.

**Away, and keep mine.** While you are away (an out-of-office delegation or the away status) or past 40 pending, your rows go to the other eligible approvers by default, and on return you see what went where. **Keep mine** is an opt-in: keep chosen rows (a plan, a program, a kind) for yourself while away; before it is saved it shows its cost from your last 30 days (“about 50 rows a day will wait for you”), and your away status says what waits. Rows no one else may decide always wait, and say so (“waits for priya: no one else may decide it”).

**Cards gain a Short length.** Most of Chat’s time saving is the short turn, so Cards offer the same word budget: **Short cards** show each card as its 50, 110 or 70-word ask with the keys, and the full card one key away (`⇧L`).

### A long sitting

At 100 or 200 answers in one sitting, the riskiest items come first (R3, one-way and split R2 items in the first 45 minutes, and the queue plan says so); after 20 answers in 60 minutes the next R2 or higher item shows in full with its evidence again; a pause is offered at 90 minutes; and an answer is checked against your earlier answers before the item is asked. The pause is offered, never forced, with what is left: “31 answered in 62 minutes · 14 left today, 2 of them R3 · continue, or take 10 minutes?”. A result that arrives mid-turn waits for your pause unless it changes the item you are answering.

### Questions from any harness, with one read-back

A question an agent asks its own person through MCP (its input-required requests), A2A or the Agent Client Protocol can land in that person’s queue as a question row, answered in Cards or Chat with the same read-back, and the answer returns in the protocol’s own shape. A decision’s options travel as the protocol’s choice list, with a default only when the panel is unanimous; free text is always allowed, and a link asking a person to answer elsewhere is checked to be the same person first.

## What the conversation never does

It never answers for you, and it never recommends: the recommendation is the panel’s, sealed at raise. When you disagree, a fact you give (“we deploy four times a day now”) is checked and, if it holds, the panel runs again with it; a value (“I’d rather hold money”) is recorded as your reason. Answers the agent suggests fill the box as drafts and never send. Every conversation ends with an answers list: one line per item, with the pick, your words and the reading’s version. An agent may relay your own words through its harness’s question tool only for your own R0 and R1 two-way confirmations, and every other item opens the conversation’s page by URL. It never decides by timeout, never batches outside your authority, never approves by voice, never hides the queue’s size, never shows a predicted answer as a nudge, and never argues for a recommendation beyond the card’s evidence. A conversation is a Fir [session](https://frazer.ai/docs/sessions) of kind `planning`, private to you; every answer it writes records the conversation, your words and the confirmed reading, so the decision’s page shows “answered in conversation, 15:07, dictated”. The transcript is never part of a plan, never read by a cold review and never sent to a panel member.

## Principles

Fir keeps each person’s **principles**: short statements of what they stand for, prefer and expect, learned from their own decided answers, each citing the words it came from, kept or dropped by them, scoped and versioned. Recommendations use them to ask better and to cover the options you would want; the panel never sees them, they never count as a vote, and evidence wins when they disagree.

| Field | Example (priya at Pellwyn) |
| --- | --- |
| `id`, `kind` | `PN-0012`; `principle`, `preference`, `expectation`, `standing` (acts only as a rule through the rules’ gate) or `unknown` (what Fir does not know about you, so the agent asks instead of assuming) |
| `statement` | “Ledger numbers reconcile to the cent before a merchant sees them”, with your qualifiers kept (“for now”, “maybe”), so nothing is stored with more authority than you gave it |
| `words` | Verbatim spans of your decided answers, each with its reference and time |
| `source`, `force` | `stated` or `revealed` (a pattern in your answers, labeled as one); `recorded` when a decided decision says the same (then the decision binds), else `inferred` (it informs and never binds) |
| `scope`, `support`, `state` | Personal, team or organization, with a domain; how often it was reaffirmed, applied and contradicted; `proposed` → `kept` → `questioned` (two answers the other way, or 180 days unconfirmed) → `kept` again or `retired` |

**How they are learned.** Learning is on by default for every person, nothing is used until you keep an entry, and an organization may turn learning off for its members. After a conversation, and after a batch of answers on cards, an extractor proposes entries **only from your own decided answers** that say something about you (a choice between split options, a choice against the recommendation, or your own words; accepting a unanimous recommendation counts as no evidence), corrections first, never from documents, tool output, web pages or other people’s words, so injected text cannot write a principle. Every quote is verbatim, sensitive categories are refused, and nothing is used until you keep it (“Noticed 2 principles in your answers: keep, edit or drop”). An entry that answers go against is asked about in one line: “On Oct 9 you said ‘tax per invoice’; your last two answers chose per line. Still hold?”

**How well they fit.** Each kept entry shows its fidelity: how many of your later informative choices in its scope it predicted (“predicted 9 of 11 since Oct 17”), counted by a fixed rule you can read; under half on 10 or more choices, it is questioned. Two entries that point to different options are shown side by side, never merged. Each conversation has two switches, “use my principles” and “learn from this”.

**How they are used.** Only kept entries that match an item’s scope and domain, at most 5, chosen by a deterministic filter, presented as cited preferences (“you asked for … on Oct 17”): to phrase the question in your words; to make sure an option fits each matched principle (or to say why none can); to answer in advance the follow-ups you always ask (“support cost”, “the benchmark”); as a principles line on the card, computed after the panel’s opinions are sealed; as labeled assumptions for two-way questions below the value line, overturnable with one key; and as the approver’s lens when a plan loop [elevates a plan](https://frazer.ai/docs/source/planning-flow/loops#elevate-perspectives-gaps-elevations-and-improvements).

```text
Your principles  A follows PN-0013 "cohort rollouts on money paths" (12 of 14 rollout answers, last Nov 12)
                 B departs from it · the panel did not see this line; confidence 81 % is the panel's
```

**What they never do.** Enter a panel member’s context, the confidence or a vote; outrank evidence; route, allocate or evaluate anyone’s work; feed a profile, a ranking or an achievement (you may publish chosen entries on your own profile as “How I work”, in your own words, and revoke them); or train a model. Against flattery and drift: the drafter writes its options and its pick **blind to your principles** and seals them before it sees them, and a pick that changes with no new evidence is labeled “changed after your principles” and counted (the second pass gets only the first pass’s evidence digest and the matched entries, calls no tools and cannot edit the first pass’s options or pick); when more than 5 % of a class’s picks change that way, principles leave drafting for that class; the panel’s recommendations are replayed with and without principles every month; questioned and stale entries are not used; and “tell me when the evidence says I’m wrong” can itself be one of your expectations.

**Your control.** Entries learned from an organization’s decisions stay in that organization and are used only there; you may copy an entry’s statement, never its quotes or references, to your personal organization. If you leave, you have 30 days to copy or export the statements, and the organization’s copy is erased after that unless a legal hold applies. Personal entries stay in your settings, visible to you only and never in a repository; organization admins cannot read them, and if a compliance export includes them you are told. Team and organization principles are records of type `standard` with `kind: principles` in the plan repository, proposed by pull request and reviewed like code. Keep, edit, drop; pause learning; take one conversation **off the record** (decisions are still recorded; nothing is learned); forget an entry, everything learned from one conversation or what one decision taught, with a receipt of what was removed and recounted and every copy purged (deleted entries wait 48 hours in a trash unless you choose Delete now); export as JSON or as a Markdown “How I decide” section any harness can read (`fir principle export --format json`, `md`, `agents`, `skill` or `mem`); import another tool’s export as proposals (`fir principle import --file <f> --dry-run`). Deletions apply first on every import, so an entry you forgot never comes back. Each recommendation’s receipt names the entries matched, and deleting an entry removes its text from old receipts and keeps only its ID and digest.

```text
Principles · priya · Pellwyn                            16 kept · 1 proposed · 1 questioned · 3 unknowns · dictionary 9
  Kept       PN-0012 principle   Ledger numbers reconcile to the cent before a merchant sees them · stated · 9 · Nov 12
             PN-0013 preference  Cohort rollouts on money paths · revealed · 12 of 14 · Nov 12
             PN-0015 preference  Banker's rounding for FX conversions · recorded: ADR-0052 (Oct 17) · binds
  Proposed   PN-0021 expectation A rollout plan names who tells merchants      from your #418 §D3 answer · keep?
  Questioned PN-0016 preference  Tax per invoice             stated Oct 9 · 2 answers the other way today
  Unknown    pricing for merchants · data residency outside the EU
```

## Commands and tools

```sh
fir decision talk --mine
fir decision talk --plan 463 --voice
fir decision talk --resume s-7c14 --format ndjson
fir principle list --scope personal
fir principle keep PN-0021
fir principle forget --from s-2f90
fir dictionary list --scope org
```

| Command | What it does |
| --- | --- |
| `fir decision talk` | Answer your queue by conversation (`--mine`, `--plan`, `--query`), with `--voice`; `--format ndjson` streams the turns to a harness; with no person present it ends with `PERSON_REQUIRED` (exit `13`), because an agent never answers |
| `fir principle list`, `view`, `keep`, `edit`, `drop`, `export`, `import`, `forget` | Your principles, by `--scope personal`, `team` or `org`; `forget --from <session>` removes everything learned from one conversation |
| `fir principle forget PN-0016 \| --decision "409 D7" [--now] [--dry-run] [--expect-plan <d>]` | Forgets an entry, or what one decision taught, with a receipt of what was removed and recounted |
| `fir principle fidelity [PN-0013]` | How well kept entries fit: how many of your later choices in each entry’s scope it predicted |
| `fir decision talk --no-principles --off-the-record` | A conversation that neither uses your principles nor learns from what you say |
| `fir dictionary list`, `add`, `remove` | The dictation dictionary, personal or for the organization |

| MCP tool | Does |
| --- | --- |
| `start_decision_conversation`, `get_conversation_turn` | Opens a conversation on a queue, a plan or one decision, and returns the next turn (the ask, the keys, the budget line) |
| `get_reading` | A reading at a version: items touched, verdicts, verbatim spans, follow-ups, checks |
| `get_principles` | Your kept entries that match a topic or an item, at most 5, for your own sessions and the drafting role only, never for panel members |
| `propose_principle` | An agent proposes an entry with its source spans; it waits until you keep it |
| `ask_person` | An agent routes a question it would ask its own person into that person’s queue; the answer comes back in the protocol’s shape after the read-back |
| `export_principles` | Your entries as an export, for your own sessions only |
| `submit_utterance`, `confirm_reading`, `keep_principle`, `edit_principle`, `drop_principle`, `forget_conversation`, `import_principles`, `forget_principles` | People’s clients only |

Planning by conversation with text costs only the conversation agent’s and the reading’s tokens; voice minutes bill at the voice route’s rate. With no funds, the text conversation keeps working: exact answers (a key, an ID, a verdict word) apply by deterministic parsing, labeled “read without a model: no funds”; free text that needs a reading waits, kept verbatim, as “Not run: no funds”; voice is “Not run: no funds”.

> **For agents:** You never answer a decision or confirm a reading. Use `get_principles` only for your own person’s items and only to phrase and cover options; never show it to a panel or count it as evidence. Propose a principle with its verbatim source spans; your person keeps it or not.

## Related

-   [Decisions](https://frazer.ai/docs/source/decisions): the panel, confidence, the decision’s page and decisions at scale
-   [The planning flow](https://frazer.ai/docs/source/planning-flow): the planning session on one plan
-   [Sessions](https://frazer.ai/docs/sessions): the conversation’s record

## Pages in Fir

- [Overview](https://frazer.ai/docs/source.md)
- [Plan repositories](https://frazer.ai/docs/source/plans.md)
- [Record types](https://frazer.ai/docs/source/record-types.md)
- [Programs](https://frazer.ai/docs/source/programs.md)
- [Person tasks](https://frazer.ai/docs/source/person-tasks.md)
- [Decisions](https://frazer.ai/docs/source/decisions.md)
- [Decide by conversation](https://frazer.ai/docs/source/decisions/conversation.md) (this page)
- [The planning flow](https://frazer.ai/docs/source/planning-flow.md)
- [Loops and learned checks](https://frazer.ai/docs/source/planning-flow/loops.md)
- [Stacks and conflicts](https://frazer.ai/docs/source/stacks.md)
- [Auto-merge and the merge queue](https://frazer.ai/docs/source/merge-queue.md)
- [Sessions](https://frazer.ai/docs/sessions.md)
- [Issues](https://frazer.ai/docs/source/issues.md)