PaperlayerView exact revision on arXiv

Source-linked AI summary

CoWriter: Guiding Methodological and Synergetic Human-LLM Co-Authoring

Shmuel Amar, Benjamin Lieber, Ori Shapira, Debanjana Kar, Sebastian Maximilian Dennerlein, Yufang Hou, Ido Dagan

arXiv:2610.31277v1cs.CLcs.AIcs.HC

TL;DR

Standard LLM writing interfaces can leave human authors in a passive role, and end-to-end co-authoring approaches remain underexplored. CoWriter adapts expert co-authoring methods to guide the workflow while engaging the user. In a user study, its authors reported stronger perspective expression and ownership, and readers rated its texts higher in human-touch qualities than baseline texts.

  • Problem

    Standard LLM writing can sideline human authors, while approaches that actively engage them across the full co-authoring process remain underexplored.

    Source: §1 · p. 1

  • Method

    CoWriter adapts expert human co-authoring methodologies into a flexible agentic workflow that guides writing and keeps the human author engaged and influential.

    Source: §1 · p. 1

  • Results

    Compared with a standard LLM setting, CoWriter authors reported stronger perspective expression and document ownership, and external readers rated its texts higher in human-touch qualities.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The findings indicate that human co-writing methods transfer well to human-LLM schemes while human involvement remains valuable.

    Source: §1 · p. 2, §6 · p. 8

    The user study had a limited population and demography recruited through universities, and covered a moderate number of writing topics.

    Source: Limitations · p. 9

Abstract

from arXiv · show

While LLMs have transformed text generation, standard chat-and-canvas interfaces often reduce human authoring to quite passive supervision, causing review fatigue, over-reliance on AI, and the sidelining of the human role in long-form, source-grounded writing. Prior systems typically address isolated writing phases, leaving comprehensive end-to-end human-LLM collaboration largely unexplored. In this paper, we introduce CoWriter, a proactive agentic scheme that adapts expert human (co-)authoring methodologies from cognitive science to flexibly steer the entire writing workflow while actively encouraging the user to engage. In an exploratory user study, compared to a standard LLM baseline, authors using CoWriter better expressed their perspectives and maintained a higher sense of document ownership, supported quantitatively by $2.2\times$ more successful agent engagements. Furthermore, external reader judges rated the resulting texts higher in "human touch" qualities like storytelling and integrating unique personal experiences. Overall, our results demonstrate that proactively engaging users through established human practices is beneficial for effective LLM-supported co-authoring, providing useful methodologies, techniques and insights for future research on collaborative agents.

1 Introduction

Standard LLM writing tools can sideline human authors, while end-to-end approaches to active co-authoring remain underexplored. CoWriter adapts established co-authoring methods to guide the writing process and engage users; its study reports stronger author ownership and engagement and more human-touch qualities in the resulting texts.

1 Introduction

  • Standard LLM-assisted writing can sideline human authors, contributing to over-reliance, faithfulness-verification challenges, and review fatigue.Prior work has addressed particular phases, but active engagement across the full process—from brainstorming to reviewing—remains underexplored.Source: §1 · p. 1
  • CoWriter adapts expert human co-authoring methodologies to steer long-form, source-grounded writing while keeping the human author engaged and influential.The approach draws on models that formalize writing activities and, for teams, collaborator roles.Source: §1 · pp. 1–2
  • Authors using CoWriter reported stronger engagement, perspective expression, and document ownership, while external readers rated its texts higher in human-touch qualities than baseline texts.The study also found about twice as many successful agent engagements with CoWriter as with the baseline LLM setting.Source: §1 · p. 2
  • The paper contributes an end-to-end agentic harness, a human-LLM co-authoring evaluation scheme, and a user study offering findings and insights for future work.Its overall conclusion is that human co-writing methods transfer well to human-LLM schemes while human involvement remains valuable.Source: §1 · p. 2

2 Related Work

Prior LLM writing assistants often focus on inline suggestions or particular phases of co-authoring. CoWriter instead draws on established models of expert individual and collaborative writing to structure a broader, actively guided workflow.

2.1 Human-LLM Co-Authoring

  • CoWriter targets source-grounded factual authoring, where continued human involvement supports claim verification, perspective shaping, and narrative steering.Unlike work focused on pre-writing exploration, turn-based collaboration, or alternative interfaces, it aims to orchestrate an established methodology while sustaining user engagement.Source: §2.1 · pp. 2–3

2.2 Human (Co-)Authoring Methodologies

  • Expert writing models describe iterative planning, translating, and reviewing under a central monitor, while collaborative-writing research formalizes activities and roles.Lowry et al. (2004) identify activities including brainstorming, outlining, drafting, reviewing, revising, and copyediting, alongside roles such as leaders, writers, and reviewers; CoWriter adapts this taxonomy into an agentic workflow.Source: §2.2 · p. 3

3 COWRITER Co-Authoring Architecture

CoWriter is a source-grounded co-authoring platform that maps collaborative writing roles onto agents and adapts an established workflow to support flexible, active human participation. A LEADER guides the work, a MONITOR tracks progress, and an on-demand REVIEWER provides feedback independent of the authoring history.

3 COWRITER Co-Authoring Architecture

  • CoWriter adapts a collaborative-writing taxonomy into a structured LLM-based workflow guided by specialized agentic components.The platform is designed to lead a methodical co-writing process and assess the benefits of this guided approach.Source: §3 · p. 3, §3.2 · p. 3

3.1 COWRITER Task and Environment

  • The task is long-form writing with explicit structure, grounded in supplied sources that support claims requiring citation.The augmented chat-and-canvas environment includes a source explorer and allows concurrent user and agent document updates.Source: §3.1 · p. 3

3.2 COWRITER Agentic System Architecture

  • The architecture maps collaborative writing roles to the LEADER, MONITOR, and REVIEWER, accommodating iterative writing rather than enforcing a rigid sequence.The LEADER guides the workflow, the MONITOR tracks progress, and the REVIEWER offers on-demand external feedback.Source: §3.2 · pp. 3–4

3.3 LEADER Agent

  • The LEADER reads sources and revises the document while proactively guiding and engaging the user through co-authoring.It acts as process lead while motivating users to participate actively.Source: §3.3 · p. 4, §3.2 · p. 4

3.3.1 Core Functionality

  • The LEADER’s workflow adapts seven co-authoring steps, inviting user participation while allowing ad hoc requests and non-sequential transitions.Step instructions define transitions, goals, actions, and engagement invitations; the agent can also respond proactively to document edits and offer follow-up options.Source: §3.3.1 · pp. 4–5

3.3.2 The MONITOR

  • The MONITOR tracks environment events and workflow progress, using a Writing Log to support visibility and trigger reminders to the LEADER.The log records visited steps, the writing goal, and a timeline; reminders cover process status, outstanding requests, candidate next steps, and divergences from guidelines.Source: §3.3.2 · p. 5

3.4 REVIEWER Agent

  • The REVIEWER reads the current document and writing goal without chat or activity history, providing feedback independent of the authoring process.On demand, it returns strengths, issues, and suggested fixes for the LEADER to share in chat; it can retrieve sources to verify claims or identify missing information.Source: §3.4 · p. 5

4 Experimental Setup

The study compares CoWriter with a standard LLM chat-and-canvas setup on source-grounded blog writing. Its evaluation combines user reports, process measures, and judgments of the resulting documents.

4 Experimental Setup

  • The study tests whether workflow guidance supports collaborative writing and produces source-grounded documents users can confidently approve.Participants wrote 750–1000-word blog posts from provided sources while contributing their own perspectives, experiences, and examples.Source: §4 · p. 5, §4.1 · p. 5

4.1 User Study Design

  • Fourteen graduate students completed within-subject writing sessions with CoWriter and a baseline sharing the same environment and underlying model.The study counterbalanced system order and topic assignment; each session was intended to last up to 75 minutes, with the second session within 48 hours.Source: §4.1 · p. 6

4.2 Evaluation

  • Evaluation combines author questionnaires and quantitative process measures with judgments of document quality and human author expression.Process measures include workflow coverage, engagement rate, successful engagement ratio, time, turns, and generated tokens; document ratings cover fluency, interestingness, logical fluency, and coherence.Source: §4.2 · p. 6

5 Results and Analysis

Compared with a bare-LLM baseline, COWRITER guided sessions more closely through co-writing activities and increased author engagement. Authors and external raters also found its documents more natural and reflective of the author, with ratings higher or similar across assessed dimensions.

5.1 Workflow Guidance via User Engagement

  • COWRITER prompted more goal-setting (+86%), brainstorming (+43%), and outlining (+57%), while sessions were more balanced across activities and elicited equal or greater user participation.The baseline’s agent-turn activity counts were concentrated in localized reviewing and revising (54% vs. 29% for COWRITER).Source: §5.1 · p. 7, Figure 2 · p. 7
  • COWRITER transitions were predominantly between adjacent activities (74%), unlike the baseline’s 44% non-adjacent transitions (p = 0.015).Both systems supported iterative processes, with roughly a third of transitions moving backward to an earlier activity.Source: §5.1 · p. 7
  • 235 vs. 109 successful agent-engagement expressions (2.2×) favored COWRITER, alongside higher subjective and objective evidence of author participation.Engagements were successful 57% vs. 47% of the time; user-initiated engagement rates were comparable across systems.Source: §5.1 · p. 7

5.2 Process Efficiency

  • 11 participants preferred COWRITER for future writing tasks versus one for the baseline (p < 0.001).Sessions averaged 62 vs. 68 minutes, and COWRITER was rated less cognitively demanding (9 vs. 3); it also used more frequent, shorter interactions and revisions, with citation-F1 of 0.84 for both systems.Source: §5.2 · pp. 7–8

5.3 Document Quality

  • Human raters preferred COWRITER documents by 43% for Linguistic Fluency and 30% for Interestingness; ratings across four dimensions were higher or similar overall.LLM-judge evaluations showed similar preferences for storytelling, personal experiences, and authors’ unique ideas or perspectives.Source: §5.3 · p. 8
  • The authors conclude that structured co-writing with greater human engagement yields documents that authors and external raters find more natural and reflective of the author.The reported trends arose despite a modest cohort, with multiple statistically significant results.Source: §5 · p. 7, §5.3 · p. 8

6 Conclusion

COWRITER shifts human-LLM co-authoring from passive text generation to an actively guided, methodical workflow, adapting expert writing practices to elicit user engagement. Its evaluation reports improved co-authoring experiences and documents preferred over a strong control baseline.

6 Conclusion

  • COWRITER documents were preferred by authors, human annotators, and LLM evaluators over a strong control baseline.The evaluation also found improved writing experience, engagement, and ownership.Source: §6 · p. 8
  • COWRITER adapts expert human writing methodologies to lead the full authoring process while actively eliciting user engagement.The framework shifts co-authoring from passive text generation to an actively guided, methodical workflow.Source: §6 · p. 8
  • The authors present COWRITER as a foundation for designing user-centered agents and harnesses for broader collaborative tasks.The conclusion grounds this direction in the value of proactive, methodical guidance.Source: §6 · p. 8

Limitations

The user studies had limited population size and demography and covered a moderate number of writing topics. The implementation used claude-sonnet-4.6, so some observed behaviors may depend on that model family.

Limitations

  • The studies recruited students and graduates through universities and covered a moderate number of writing topics.Broader studies across diverse populations, professional settings, and genres would further validate the findings.Source: Limitations · p. 9
  • COWRITER used claude-sonnet-4.6, and some observed behaviors may depend on characteristics of that model family.The authors primarily study workflows, interaction design, and agentic behavior, and expect the principles to transfer to modern LLMs with comparable abilities.Source: Limitations · p. 9

Ethical Considerations

The appendix describes participant protections and data handling alongside the interface and agent components used in COWRITER. The system combines a shared authoring canvas, a Writing Log, workflow monitoring, and external review.

Ethical Considerations

  • Participants gave informed consent, could withdraw without penalty, and had their recorded session data anonymized; sources lacked sensitive personal information.LLM output was presented as assistive material requiring verification, and AI-assisted writing was manually verified and edited.Source: Ethical Considerations · p. 9

A User Interface

  • The interface combines chat, a co-edited document canvas, source information, and a Writing Log maintained by the MONITOR.The Writing Log opens from the left pane and overlays the sources panel.Source: A User Interface · p. 12

B Main Agent (LEADER) Details

  • On user messages or document updates, LEADER receives the document, chat history, Writing Log, and system reminders, then runs a ReAct loop.The loop interleaves reasoning steps and tool calls before sending a final chat message.Source: B Main Agent (LEADER) Details · p. 12, Figure 6 · p. 14

C Writing Log

  • The Writing Log contains a Writing Activities checklist, Writing Goal, Sections, Key Decisions, and Activity Log.MONITOR updates the checklist from its latest activity reminder and edits the other sections through structured output fields.Source: C Writing Log · p. 15

D MONITOR Details

  • MONITOR decides whether user document revisions trigger LEADER, while user chat messages always trigger it.After LEADER responds, MONITOR updates reminders and the Writing Log in the background.Source: D MONITOR Details · p. 15
  • MONITOR maintains reminder sections through per-field deltas: null preserves a value, empty values clear it, and other values replace it.Its status signals include document word count, progress markers, broken citations, and whether the user has edited since LEADER last read the document.Source: D MONITOR Details · p. 15

E REVIEWER Agent

  • The REVIEWER runs a short ReAct loop over the Writing Goal and document, then returns structured feedback for LEADER and the user.Its review may include a coverage gap and ends with a top fix.Source: E REVIEWER Agent · p. 17

F Writing Task Definition

Participants received the same writing-task instructions across topics and systems, with the topic title instantiated for each session. The task framed them as bloggers writing a source-based post that reflects their perspectives and serves readers.

F Writing Task Definition

  • The writing task used one of four assigned topic titles, while the remaining instructions stayed the same across topics and systems.The task definition was shown at the start of every writing session.Source: F Writing Task Definition · p. 17, Figure 8 · p. 18
  • Participants were asked to write as popular everyday-life bloggers, educating readers with research findings while following their own perspectives.The post was to be based on source facts and the writer’s perspectives on them.Source: F Writing Task Definition · p. 18
  • The task asked writers to select information they considered interesting and potentially useful for readers’ life habits and choices.The example topic shown is the impact of smartphones on family and friend communication.Source: F Writing Task Definition · p. 18

G Writing Topics and Sources

The study used four writing topics supported by 40 research sources, then compared participants’ co-authoring sessions with COWRITER and a baseline using counterbalanced assignments and complementary questionnaires.

G Writing Topics and Sources

  • The study used 40 permissively licensed research papers, ten per topic, and supplied participants with textual versions converted from PDFs to Markdown.The four topics covered family and friends, smartphones and communication, education and play, and young children’s formal education and play.Source: G Writing Topics and Sources · p. 19, Table 4 · p. 21

H User Study Details

  • Fourteen eligible Masters and PhD students completed a baseline and a COWRITER session on different topics, with system order and topic–system pairings counterbalanced.Each topic was assigned to seven participants and appeared in both conditions; sessions were logged, and participants had prior AI-tool experience and had written a multi-page academic document in English.Source: H User Study Details · p. 22, Table 5 · p. 22

I User Study Questionnaires

  • Participants rated each system separately on 37 seven-point statements, then compared systems on 13 five-point questions, with cognitive load and usability also measured.The two questionnaire formats distinguish absolute system ratings from direct preferences; reverse-coded items were inverted so higher scores indicated more positive responses.Source: I User Study Questionnaires · p. 22

J Attention-Check Handling

The analysis retained P6 after a failed attention check in both post-session questionnaires, based on convergent engagement, debrief, and re-rating evidence.

J Attention-Check Handling

  • P6 was retained despite failing the instructional check twice: session evidence and the debrief indicated engagement, while blind re-rating matched the reported direction.Cognitive Load was PROACTIVE = 1.00 vs. BASELINE = 6.25; UMUX was PROACTIVE = 7.00 vs. BASELINE = 6.50.Source: J Attention-Check Handling · pp. 22–23

K Session Annotation Details

The authors used LLM annotations of session messages to quantify writing activities and user engagement, and checked annotation agreement against human labels.

K Session Annotation Details

  • An LLM labeled each session message with its most prominent writing activity from seven predefined activities, or “other” if none applied.It also extracted engagement expressions verbatim and classified their type and whether the collaborator followed them up in the next turn.Source: K Session Annotation Details · p. 23
  • Message-level activity labels, engagement-expression counts, and successful follow-ups were aggregated by system to calculate the process metrics.Source: K Session Annotation Details · p. 23
  • Human–LLM agreement across 118 messages averaged κ = 0.54 across writing activities, engagement expressions, and successful engagements.Two human annotators independently labeled messages from four sessions using the LLM’s guidelines.Source: K Session Annotation Details · p. 23

L Human and LLM-based Document Assessment Details

The document assessment combined human ratings of sampled document pairs with LLM-judge ratings of all pairs, alongside an automated citation-quality assessment.

L Human and LLM-based Document Assessment Details

  • Five annotators rated 30 sampled document pairs; study participants were prevented from reviewing pairs containing their own documents.Two external annotators rated all 30 pairs.Source: L Human and LLM-based Document… · p. 24
  • Human ratings covered 19 comparative questions grouped into interestingness, linguistic fluency, coherence, and logical fluency, using a five-point scale.The reported score per pair and item was the mean across the annotators who rated that pair.Source: L Human and LLM-based Document… · p. 24, Table 8 · p. 27
  • Three LLM judges rated all 48 same-topic document pairs on the same 19 questions and four aspects, with COWRITER documents balanced across A and B positions.Each judge used a five-point scale; the COWRITER document appeared in each position for 24 of the 48 pairs.Source: L Human and LLM-based Document… · p. 24
  • Citation correctness was assessed with LongCite prompts applied to sentences paired with their cited source snippets, reporting citation length and recall.Recall was split between statements that cite a source and statements that do not.Source: L Human and LLM-based Document… · p. 24

M Writer Assessment Detailed Results

The detailed assessment reports that COWRITER spread agent engagement more evenly across writing activities and supported a more methodical workflow. Participant responses also describe benefits from guidance and time savings, alongside concerns about control and the amount of text requiring review.

Assessment measures

  • Questionnaire constructs had Cronbach’s α of 0.65–0.89 per system; comparative dimensions were 0.94 for overall experience and 0.55 for document self-expression.All three statements in the latter dimension had an identical +0.93 mean preference for COWRITER.Source: M Writer Assessment Detailed Results · p. 29

Workflow adherence and engagement

  • COWRITER distributed agent engagements more uniformly across the workflow, while baseline agent turns concentrated in localized reviewing and revising (54% vs. 29%).The figure also reports higher user participation across most activities for COWRITER.Source: Figure 14 · p. 30, Figure 17 · p. 32
  • 74% of COWRITER activity transitions were between adjacent activities, versus 56% for the baseline (p = 0.015).Both systems featured over 30% backward transitions.Source: Figure 15 · p. 31

Writer assessments

  • One participant felt COWRITER took too much control, leaving them more like a reviewer than a creative writer.Other responses mention confusion after system errors and difficulty correcting document structure.Source: M Writer Assessment Detailed Results · pp. 36–37
  • Participants described COWRITER as helping with outlines and ideas while giving them more authority to write or structure ideas themselves.Another participant said the agent's initiation and questions helped overcome prolonged brainstorming.Source: M Writer Assessment Detailed Results · pp. 36–37
  • Participants also valued faster source summarization and drafting, although one reported that a large amount of generated text needed checking.A baseline respondent described manually finding citations as time-consuming and tiring.Source: M Writer Assessment Detailed Results · pp. 36–37

N Significance Testing

This section specifies the source-grounded, user-steered writing workflow and the agent prompts, interaction rules, and citation conventions that support it. It also describes study measurements and paired-test choices, with examples and operational rules for annotating and supporting engagement.

N Significance Testing

  • Across 14 within-subject pairs, paired t-tests covered constructs within normality thresholds; Wilcoxon tests handled Document Publish, goal-setting adherence, and the skewed choice-of-content item (p=0.031).For paired differences, substantial violations were defined as |skewness| > 2 or |excess kurtosis| > 7; the choice-of-content item was the only one of 37 per-question analyses to exceed a threshold.Source: N Significance Testing · p. 39
  • Engagement rates were compared with an exact paired permutation test, while the percentage of transitions to adjacent activities was compared with a paired t-test across 14 participants.The permutation test enumerated all swaps of system labels within each participant’s paired sessions.Source: N Significance Testing · p. 39

O Automatic Measurements

  • Table 11 lists per-session duration, LLM cost, agent and user messages and words, document revisions, net character change per revision, and four COWRITER-only metrics.Cells show averages and standard deviations across 14 sessions per system; significance markers indicate paired tests on proactive-minus-baseline differences, with p < 0.01 marked ** and p < 0.05 marked *.Source: O Automatic Measurements · p. 39, Table 11 · p. 40

Writing Goal

  • The example writing goal was an accessible, source-cited blog for general readers on formal education and play, structured in five sections with personal experience before practical takeouts.The user confirmed the title “Stop Drilling, Start Playing: What Science Says About Raising Smart Kids.”Source: Writing Goal · p. 41, Key Decisions · p. 41

Activity Log

  • The engagement annotation prompt defines questions, choices, action requests, and next-step proposals, excludes plain compliance and acknowledgments, and requires a row for every chat message.A message with no engagement receives a minimal row marking engagement false.Source: Activity Log · p. 45

Follow-up

  • An engagement counts as followed up only when the other party’s next contiguous turn explicitly addresses the same scope; later uptake or a redirect does not count.The criteria vary by engagement type: answering a question, resolving a choice, performing or committing to an action, or proceeding with a proposed next step.Source: Follow-up · p. 45

Output format

  • The annotation output is a JSON array ordered by message ID, with fields for engager, engagement status and text, type, follow-up status and text, and a comment.When engagement is false, only msg_id, engager, and engagement are emitted.Source: Output format · p. 45

Updating reminders

  • The Monitor maintains project reminders without writing or editing the document, tracks the collaborative source-grounded workflow, and surfaces only current open questions, requests, or prioritized pending steps.The reminder rules limit open questions and requests to two each and pending steps to two, prioritizing the lowest-numbered unfinished step.Source: Output format · p. 49, Updating reminders · p. 50

When to trigger (trigger_now)

  • The monitor triggers on actionable, finished document edits, including large additions or deletions, repeated-term renames, uncited paragraphs, restorations, and structural changes.It proposes grounding and integration for additions, downstream cleanup for deletions or structural changes, and respects restored content as a user steering signal.Source: When to trigger (trigger_now) · pp. 50–51

Goal

  • The reviewer checks whether cited source lines support document claims and evaluates the text for a defined audience and motivation.Its review includes an objective summary, one located strength, two to four audience-impact issues with evidence and fixes, and one top fix.Source: When to trigger (trigger_now) · p. 52, Goal · p. 52

Collaboration Model

  • The writing agent leads the process while users retain control of strategic choices, with confirmation required for major changes but not routine local edits.It works collaboratively in incremental turns, proactively proposes steps, and invites users to review, share perspectives, and make decisions.Source: Collaboration Model · p. 54, §4 · p. 54

Turn Mechanics

  • Each agent turn orients to context, selects and performs an activity, then sends one chat message containing a contribution and an invitation.The writing process completes team formation and goal definition before iterative production activities, which can be interleaved or revisited.Source: Turn Mechanics · p. 55, The Writing Process · p. 55

Team Formation

  • Team formation starts with the agent welcoming the user and asking what they want to write about, while spontaneous user requests take priority over the writing flow.After fulfilling a request, the agent offers brief follow-ups to continue that thread or return to writing.Source: Team Formation · p. 55

1 Writing Goal Definition

  • The writing goal is defined with the user across topic, motivation, audience, constraints, and sources, with unknown aspects marked [TBD].The agent asks about missing aspects together and grounds the writing in user-provided sources.Source: §1 · p. 56

1 Writing Goal Definition / 2 Brainstorming

  • Before brainstorming, the agent reflects the writing goal for confirmation, then explores source-grounded directions and updates the goal as the user chooses an angle.Brainstorming can be revisited when new ideas emerge, and the workflow permits skipping steps at the user’s request.Source: §1 · p. 56, §2 · p. 56, Quality Standards · p. 58

3 Outlining

  • After a direction is confirmed, the agent proposes a source-backed outline, checks each section’s support, and iterates with the user before writing it to the shared document.Alternative structures and orderings are offered for the user to consider.Source: §3 · p. 57

4 Drafting

  • Drafting proceeds paragraph by paragraph: the agent proposes source-grounded content points, drafts with citations, and waits for user review before drafting another paragraph.The user helps choose where to start and what to emphasize, add, or skip.Source: §4 · p. 57

5 Reviewing & Revising

  • After each paragraph, the agent and user evaluate it against the writing goal, verify that citations support claims, and revise it before continuing.The user must sign off before the next paragraph begins.Source: §5 · p. 57

6 Copyediting / 7 External Review

  • When the draft is ready, copyediting checks grammar, consistency, style, and citation issues, then an independent reviewer provides a fresh target-audience read.The agent presents review issues and possible next steps to the user; unresolved issues return to drafting, while a clean review leads to completion.Source: §6 · p. 58, §7 · p. 58

Document Production

  • The workflow allows activities to loop or occur in any order, with the agent proposing the next activity and briefly explaining the choice.It also encourages taking multiple turns on important decisions to keep the user engaged.Source: Quality Standards · p. 58

Quality Standards

  • Quality standards require source-grounding and citations for non-trivial claims, while allowing users to skip steps and loop back to activities.The agent should take time on consequential choices and encourage user contributions.Source: Quality Standards · p. 58

Tools & Environment

  • The agent works across source files, an editor, and chat, and must read the shared document state because users can edit it directly.Document changes use XML block-level updates that are checked by rereading the document.Source: Tools & Environment · pp. 58–59

Citation Format

  • Citations should wrap the claim’s own words in clickable links using plain file-line references, never expose file IDs or footnote-style markers as link text.Use granular line ranges and stack refs with semicolons when multiple sources jointly support the same claim.Source: Citation Format · p. 60

Examples

  • The examples distinguish stacked references for jointly supported claims from separate links for sentence spans supported by different sources.Bare references outside links and stacked separate link tags are disallowed.Source: Examples · pp. 60–61

Chat Message Rules

  • Chat messages should be concise and direct, focus on one task, and invite user participation through a choice, question, or review.For bigger, previously undiscussed changes, ask before editing; when drafting, provide one or two alternative phrasings for key sentences.Source: Chat Message Rules · p. 61

Completion

  • The Leader agent’s complete system prompt encodes the seven-step writing workflow, ad-hoc requests, turn mechanics, citation format, and tool-use rules.The workflow steps are listed in Table 1 and discussed in §3.3.Source: Figure 28 · p. 62

Related papers

Loading 2610.31277v1…