Vai al contenuto
ErmisAI

Composing and refining a story

Generate the first article, then use chat, quick completion and per-message apply to get it publishable.

The story workspace is where a story stops being source material and becomes your article. You open it by clicking a story in the feed.

Read this first: a story never arrives with an article already written. Every story lands as a placeholder draft. Until you run compose once, there is nothing to refine — the chat will refuse to answer and the review handoff will refuse to submit. Generating the first article is a deliberate human action, and it is the only way to get copy.

Who can open it

Every member of the newsroom can open the story workspace and edit stories. Anyone without a newsroom role gets a screen reading "Story access is restricted" and "Story editing is limited to newsroom writing and editorial roles."

One limit applies to the member role: members can only compose, save, submit or annotate drafts they own, or drafts nobody owns yet. Picking up an unowned draft claims it — the first member who composes, saves or submits it becomes its owner.

If the newsroom's subscription is not in good standing, a red panel sits above everything:

AI generation is paused for this workspace The subscription is not in good standing, so compose and refinement requests are blocked. Settle the open invoice to restore them immediately.

Every generation control is disabled until it is resolved. The panel links to "Open billing settings", which only an owner can open.

The two modes

The header carries a small card labelled "Workspace mode" with two buttons:

ModeHeading you seeWhat it is for
Refinement (default)"Refine article"Working on an article that already exists
Compose"Compose article"Generating the first article, or replacing it

In Compose mode with an article already saved, the heading changes to "Review article". The mode you are in is kept in the page address, so a link you copy reopens in the same mode.

Switching modes also switches the conversation: refinement and compose each keep their own separate chat transcript for the story.

Generate the first article

Open the story. In Refinement mode with no article yet, the article panel is empty and shows a button, "Compose the first article". Click it — or switch to Compose in the Workspace mode card.

Generation starts on its own. Switching into Compose mode on a story that has never been composed fires the first generation immediately. You do not need to press anything else.

Wait. An animated panel replaces the article area with a rotating status line and the note "The article will appear automatically once it is ready." Nothing is lost if you look away.

Read what came back. A toast reads "Article generated". The headline replaces the placeholder at the top of the page, topic tags appear beneath it, and an estimated read time joins the header. The article body fills the panel headed "Saved article (English)" — the language in brackets is your newsroom's article language, not your interface language.

If generation fails, the reason appears in a red panel directly under the article area. When you triggered it yourself you also get a toast; when it started automatically you only get the panel, so check there if a story looks stuck.

Compose settings

In Compose mode, a collapsible panel labelled "Compose settings" holds three drop-downs. There are no free-text fields — no voice notes, no style guide, no extra instructions.

ControlOptionsDefault
FormatQuick brief, Full story, Broadcast script, Digest blockFull story
ToneNeutral, Concise, NewsroomNewsroom
LengthShort, Medium, LongMedium

Length is a real word budget: Short targets 80–160 words, Medium 180–320, Long 320–700.

These three settings are not saved. They reset to the defaults every time you open a story. They also do not reach the chat — a request you type in Compose mode is always built as Full story, Neutral, Medium regardless of what the drop-downs say. They do apply to compose runs and to Quick completion.

Reading the saved article

The article body is read-only here. You cannot click into it and type. That is deliberate: the workspace changes copy by regenerating it, or by applying a block of assistant text over it. Direct sentence-level editing happens on the review desk, not here.

A saved article always shows three controls in its top-right corner:

  • MD — copies the article as Markdown, with the headline as a heading. Toast: "Article copied as Markdown".
  • FMT — copies it as formatted text, ready to paste into a rich editor. Toast: "Article copied as formatted text".
  • A chevron that collapses the article to a short masked preview, with "Show full article" to expand it again.

Copying works on any saved article in any state, including after publication.

The header also carries four read-only chips: the story category, "Signal: …" with the confidence score out of 100, "Stage: …" for the pipeline phase, and "Editorial: …" for the editorial state. The four signal descriptors — Verified signal, Strong signal, Developing signal and Low signal — are always in English, whatever your interface language. See Confidence, sources and conflicts for what the score is built from, and Pipeline phases and editorial states for the two status systems.

Three ways to change the article

1. Ask in the chat, then apply the reply

The chat panel below the article is titled "Article workspace". A pill in its header reads "Ready", or "Generating" while a response streams.

Type what you want changed. In Refinement mode the placeholder reads "Ask for a rewrite, headline options, or wording cleanup."; in Compose mode, "Describe the exact change you want in the saved article." Press Enter to send, Shift+Enter for a new line. Before your first message, a collapsible "Quick prompts" row offers three ready-made requests, written in your article language.

A reply changes nothing on its own. Under every finished reply are up to three buttons:

  • Copy — always available.
  • Append to article — adds the reply to the end of the saved article.
  • Replace article — swaps the whole body for the reply.

Append and Replace only appear once an article has been saved, and both are unavailable while a response is still streaming or another change is being applied. Applied text is stripped of Markdown formatting first — headings, bold, links and tables are flattened to plain text. If the reply contains no usable text you get "The latest assistant reply does not include article text."

Two more things worth knowing about the chat:

  • Your conversation is yours. Transcripts are stored per person. A colleague opening the same story sees an empty chat, not yours. The header has a "Retry last response" action and a download button that saves the conversation as a Markdown file.
  • The attach button is switched off in practice. The composer shows a paperclip for JPEG and PNG images, but the workspace runs on a reasoning model by default, and the composer states: "Attachments are disabled for reasoning models in this workspace."

2. Quick completion

Between the article and the chat sits a one-line surface headed "Quick completion". Type a single instruction — the placeholder suggests "e.g. Add a closing paragraph on what to watch next" — and press Enter or click Generate. The result streams into a preview box with three actions: Append to article, Replace article, Discard. Instructions are capped at 2,000 characters, and the result is a short passage, not a full article.

The on-screen hint says "One instruction, one source-grounded continuation for the saved article." Only the first half is accurate. Quick completion is given the source story — the headline, event summary and source list — and not the text of your saved article. It cannot continue a paragraph it has never read. Use the chat when the request depends on what the article currently says.

3. Regenerate

In Compose mode, the button beside the article heading reads "Regenerate article" once an article exists. It throws away the saved article and composes a new one from the reporting brief, showing "Replacing the saved article with a new version from the reporting brief…" while it runs.

You can only do this while the story is still in Draft state. After it has been submitted, the settings panel shows: "Regeneration is locked because the saved article has already entered editorial workflow." Getting it back requires a reviewer to use "Send back to compose" on the review desk.

The reporting brief

The right-hand column opens on the Brief tab, which holds two panels.

Story status shows where the story sits in the pipeline — a four-step ladder from Source intake through Article synthesis and Editorial review to Published — plus a live connection pill reading Connecting, Live, Polling or Offline, and a "Last updated" time. The phase updates by itself while you work.

Reporting brief is the source material the article was built from. Its own subheading is the warning worth heeding: "Use this as source context for the article, not as publication copy." It contains:

  • The source story headline, which is the wire headline, not your article headline. These are two different things and they will diverge as soon as you compose.
  • Count chips: how many sources, how many angles, and — in amber — how many open issues.
  • The event summary.
  • Angles to preserve, a chip list from the synthesis step.
  • Open issues, an amber block listing the points where the sources reported differently.
  • A collapsed Source attribution list of outlet names.

The one-line description under each source name in Source attribution is generic filler assigned by position in the list — the first source always reads "Primary incident framing and location confirmation", the second "Response details and institutional statements", the rest "Contextual angles and corroborating details". It is not an analysis of what that outlet actually contributed. Read the source names; ignore the sentences.

Claim-level provenance is not on this screen. It appears only on the editorial review desk, and only while a story is in review.

The rest of the sidebar

The right column has up to four tabs.

  • Brief — the reporting brief and pipeline status, above.
  • Editorial — the handoff controls, story notes, review notes and revision history.
  • Desk — a snapshot of the wider feed: Delivered and Queued counters, an "Open feed" link, and up to six nearby stories with their category, source count and whether an article has been saved.
  • Queue — your personal editor queue, if your newsroom has it enabled.

Story notes, the Queue tab, "Nudge reviewer" and "Submit and open next" are each switched on or off per newsroom. If you do not see one, your newsroom does not have it.

Hand the article to review

Open the Editorial tab. The "Editorial handoff" panel shows the current State and Priority, then two drop-downs and a row of buttons.

  • Owner — who is responsible for the article. Reassigning it to someone else requires an owner or admin; a member who tries sees "Owner assignment is available to editorial assignment roles." A member submitting a draft nobody owns is assigned as owner automatically.
  • Priority — Standard, High or Urgent. This drives the order of the editorial board.
  • Submit for review — routes the article to the review queue. It is disabled once the story is already in review. Toast: "Article routed to review queue".
  • Open review — appears only if you can review and the story is in review.
  • Save snapshot — writes an entry to Revision history and a "Manual snapshot" line in Review notes, without changing the article. Useful before a risky regeneration.
  • Nudge reviewer — sends a reminder. It appears only while the story is in review, and never in a one-person newsroom.

If you are the only active person in the newsroom, a note reads: "You're the only active member of this newsroom, so you review and approve your own drafts." Self-approval is allowed and is not a workaround.

Try to submit before composing and the button short-circuits with "Generate the article before sending it to review." (The longer wording "Generate the first article draft before submitting for review." only exists as the API's 409 response and is never shown in the workspace.)

Below the handoff panel: Story notes (short handoff context, which notifies the owner and, once the story is in review, the reviewers), Review notes (feedback left by reviewers, badged Feedback, Approved or Rejected) and a collapsed Revision history. Revision entries are written by the system in English regardless of your interface language, and the list keeps the most recent 50.

What is locked, and when

Editorial stateWhat you can still do here
DraftEverything: generate, regenerate, append, replace, note, submit
In reviewNotes only. Headline and body changes are rejected by the server, not just discouraged
RejectedAppend, replace, note and resubmit. You cannot regenerate from scratch
PublishedRead and copy only. Every change is refused

Two more limits to plan around:

  • The article body is capped at 25,000 characters. Appending long assistant replies repeatedly will eventually push it over, and the save then fails with a blunt "Invalid draft payload". Replace rather than append when the article is already long.
  • Two people cannot save at once. If a colleague changed the story while you had it open, your save is refused with "This story was updated by someone else. Refresh and try again." and the workspace reloads the current version. Your unapplied chat reply is still in the transcript — apply it again.

Article language

The language of the article is your newsroom's article language setting, never the language of the sources. It is fixed onto the story the moment you first compose it. Changing the newsroom setting afterwards does not re-language a story that has already been composed, and there is no per-story language control in the workspace.

Some strings stay in English in every interface language: the four signal descriptors, the assistant's name on chat bubbles, and the system-written revision titles. See Interface language and article language.

Keyboard shortcuts

Available when your newsroom has the personal editor queue enabled, and ignored while your cursor is in a text box:

ShortcutAction
Alt + [Previous story in your queue
Alt + ]Next story in your queue
Alt + Shift + QAdd or remove the current story from your queue
Alt + Shift + EnterSubmit for review and open the next queued story

Messages you may hit

MessageWhat it means
"Article generation is temporarily rate limited by the AI provider. Please retry shortly."Too many generations too quickly. Wait and retry
"Article generation could not run because the story context exceeded the model limit."The source material was too large for one pass
"Article generation was blocked by the AI safety system for this story context."The provider refused this material. Regeneration will not help
"Article generation timed out while waiting for the AI provider. Please retry."No response in time. Retry
"Generate the article before applying assistant changes."Compose has not run yet on this story
"Generate the article before sending it to review."Same cause, on the submit path
"Regeneration is locked because the saved article has already entered editorial workflow."The story has left Draft state. A reviewer must send it back to compose
"Story unavailable"The story could not be loaded — often because your newsroom is over its daily delivery cap

More in Troubleshooting.

In questa pagina