The AI Tools Editors Use in Compose
Agents

The AI Tools Editors Use in Compose

Skip to main content
Table of Contents
< All Topics
Print

The AI Tools Editors Use in Compose

Last updated: November 20, 2026

What it is

Compose is our authoring console, the private tool editors use to write, check and publish documents. Its AI tools help an editor draft, research, check and illustrate. This article lists what each one does and does not do.

The rule that frames all of them: an editor asks, the AI proposes, and an editor decides. Nothing in this article publishes on its own. For the wider policy, see How Citizen Analyst Uses AI and Editorial Review & Publishing Process.

When it runs

Only when an editor uses it. None of these tools runs on a timer. Each one runs when someone clicks a button, opens a document or sends a chat message.

What it does

  1. Co-writing chat. An editor chats with Claude about the open document. The assistant is told to act as a peer collaborator, to cite a source for every factual claim or flag the claim as needing verification, and to follow our evidence tiers: Documented, Credibly Reported and Alleged. It must never present AI-generated analysis as documented evidence. It is also given a list of banned words and structural habits that make text read as machine-written. When the document is an accountability profile, extra rules switch on: every action needs a cited source, criminal claims need court records, and a short list of loaded labels for people is barred unless a court record supports it.
  2. Grounding sources. The editor picks the documents the chat should rely on. The server assembles them. If the total is too large, the tool truncates and tells the editor which sources were dropped, instead of quietly using less than the editor expects. A manifest shows everything injected into each turn and what it cost.
  3. Research to draft, with citations. The tool searches the web (through Tavily), then asks Claude to write from those sources using Claude’s citations feature. Each inserted claim carries a numbered marker tied to a source and, where available, the exact passage it came from.
  4. Web Update. One button refreshes the open document. It searches through our trusted-source routing, limited to a lookback window anchored on the document’s last-updated date (from one week to one year), and returns a briefing on what is new, corrected or superseded, with citations. It writes nothing. The document itself is sent only as a baseline, never as a source, so a claim in the briefing has to cite web material, not the document it is updating. The editor applies the briefing by inserting it or by a merge that shows a diff for review.
  5. Document Credibility Score (DCS). A score from 0 to 100 for how well a document is sourced. It has four parts, each worth up to 25 points: the mix of sources, attribution, claim support and red flags. The source mix is calculated by code against our Source Credibility Registry. Claude judges attribution and claim support. The score measures sourcing rigor, not political direction. Claude is told not to penalize a point of view or strong language. It highlights passages and asks questions. It never rewrites. The score is internal. It is not saved in the document and readers never see it. See Source Credibility Scoring (Document Credibility Score).
  6. Optional OSINT pre-pass. An editor can switch on a pre-pass that gathers open-source material first and adds a short brief to the chat turn. It runs under token and call budgets. Its search queries come from fixed templates filled with the document’s subject. It does not run searches about a private or unknown person in categories that target people. It can also query some official record sources, such as congressional, federal register, campaign finance and securities filings, under rate limits. A source that cannot run is reported as “not run” with the reason, and the turn carries on.
  7. Generators. Panels produce infographics (PNG), Datawrapper charts, timelines, network graphs, slideshows (video) and AI images. They appear only in Compose. Timelines and network graphs show on the website only and are dropped from the newsletter version. Each generator builds from the document’s content, and the editor decides what goes in. For a Datawrapper chart, for instance, the editor reviews the proposed chart type and data before the chart is created.
  8. AI images. Illustrations come from two outside image models. Each one carries a visible “AI-generated illustration” label in the image itself. How Citizen Analyst Uses AI explains the labeling and what we never do. In particular, no software checks for an AI image of a real person in a scene that did not happen. An editor does.
  9. Suggested next prompts. A small, cheap model offers three or four ideas for what to ask next. They nudge toward sourcing and counter-argument work. If the feature fails, a fixed list appears, and it never blocks the chat.
  10. Glossary and timeline extraction. Glossary terms and timeline events are proposed from an AI-built summary of each document, not written freehand. A glossary term goes through candidate, approved and applied stages. Only applied terms are ever injected, so an unapproved term never reaches the graph.

Skills and models it uses

Rules and gates it obeys

  • Anti-slop and evidence rules in the chat prompt, described above. These are prompt instructions. They shape output but do not guarantee it.
  • Every write to the document store goes through one guarded path. The guard audits the text and reports findings. It warns rather than blocks, because the editor’s buffer is often the only copy. The same guard covers AI output and hand-typed text. A failed save never costs the editor the text in the buffer.
  • The DCS is advisory by default. An editor can turn on an optional publishing floor for the source-mix part of the score. When that is on, a document below the floor is blocked.
  • Publish-time gates still apply to everything. See Publish-Time Gates and Evidence, Sourcing and Language Rules.

What it may never do

  • Publish, or push to the live site, by itself.
  • Write to a document without the editor seeing it. Web Update writes nothing. Merges show a diff.
  • Treat AI output as evidence.
  • Show the Document Credibility Score to readers.
  • Run on a schedule.

Where a human decides

Everything. The editor chooses which tools to use, which sources ground the chat, whether to accept a merge, whether an illustration goes in, whether a glossary term is approved, which persona to consult and when to publish. That includes checking every AI illustration before it is placed.

Factual correction requests: If you believe information in this page is incorrect, please contact factcheck@citizenanalyst.ai with your name (optional), the specific claim, and any supporting documentation. We review all submissions and correct verified errors promptly.

Was this article helpful?
0 out of 5 stars
5 Stars 0%
4 Stars 0%
3 Stars 0%
2 Stars 0%
1 Stars 0%
5
Please Share Your Feedback
How Can We Improve This Article?