Hci Paper Writing

Most academic writing tools start with the prose. HCI Paper Coach starts with the argument: what the paper claims, what evidence supports it, and where a reviewer may lose confidence.

The skill is intended for authors submitting to CHI, CSCW, DIS, UIST, TOCHI, and related HCI venues. It uses different review lenses for qualitative inquiry, interaction techniques, field deployments, research through design, and other HCI traditions.

[!IMPORTANT] This is an author-side reasoning and revision tool. It does not predict acceptance, generate citations, or replace research judgment.

[!NOTE] We are working on a five-year CHI embedding analysis and a rights-cleared dataset of accepted/rejected submission journeys, reviews, and rebuttals. Benchmark contributors get priority invitations to the private beta.

Three core workflows

The same skill also handles paper planning, section revision, study reporting, interaction figures, LLM-system reporting, privacy checks, and current venue policies.

Highlights in v0.4.0

This is not a grammar wrapper with an HCI prompt pasted on top. The workflow is built around HCI's different contribution types and research traditions.

It also keeps four decisions separate: what the contribution is, which research area it belongs to, how knowledge is produced, and where the paper will be submitted. A human-AI paper is not automatically a systems paper; a CHI paper is not automatically an experiment.

30-second install

Codex and Agent Skills-compatible tools

git clone https://github.com/RobbieRao/hci-paper-writing.git
cd hci-paper-writing
./install.sh codex

Claude Code

git clone https://github.com/RobbieRao/hci-paper-writing.git
cd hci-paper-writing
./install.sh claude

The installer creates a symlink so git pull updates the installed skill. It refuses to overwrite an existing installation.

Then ask your agent:

Use $hci-paper-writing in hci-diagnose mode on my abstract and contributions.

Or:

Use $hci-paper-writing to red-team this draft for CHI 2027.
Separate contribution, method, and reader/venue risks.

For a multi-session paper project:

Use $hci-paper-writing in hci-init mode for this project, then run hci-panel.

Local manuscript preflight

Run the deterministic local scanner before semantic review:

python3 skills/hci-paper-writing/scripts/manuscript_audit.py paper.tex \
  --anonymous --strict

It detects:

• paper structure and missing common headings;
• explicit RQ and contribution markers;
• strong constructs such as trust, understanding, agency, and usefulness;
• common study, ethics, and limitations markers;
• a reverse outline from each section's opening move;
• LaTeX figure/table definitions, references, captions, and orphaned labels;
• LaTeX inputs, graphics, bibliography files, and citation-key resolution;
• common anonymity leaks such as author metadata, email, identifying URLs, and acknowledgments;
• unfinished text such as TODO, TBD, and FIXME.

The scanner uses only the Python standard library. It is read-only and makes zero network requests. Its output is a set of review leads, not a paper quality score. It accepts .docx, .md, .tex, and .txt files. Strict mode fails only on deterministic integrity defects, not on debatable semantic advice.

Add --format json when you want stable, machine-readable output for an agent pipeline or benchmark harness.

# Local Manuscript Preflight

- File: synthetic-paper.md
- Sections detected: 6
- Privacy: local read-only scan; no network requests

## Contribution Candidates
- We introduce TraceLens and use it as a research artifact...

## Review Leads
- Strong-claim term detected: usefulness
- Verify that each construct is operationalized or carefully bounded

Try it on the intentionally synthetic example paper.

What changed across five years of CHI?

The short version: CHI is moving beyond generic “AI can do X” prototypes toward harder questions about agency, professional practice, governance, care, education, identity, verification, and failure.

Our higher-confidence CHI 2027 scenarios are negotiated AI agency, situated deployment, epistemic interaction (provenance, uncertainty, verification), and multi-party consequences. Medium-confidence scenarios include embodied/spatial generation, longitudinal AI relationships, accessibility as method and governance, and AI-mediated research practice.

Read the full bilingual atlas: English · 简体中文

These are proceedings trends, not submission or acceptance predictions. The atlas distinguishes raw paper growth from growth in annual share and says what evidence would falsify each forecast.

Give each paper a memory

Initialize a local workspace inside an existing paper directory:

python3 skills/hci-paper-writing/scripts/project_workspace.py \
  /path/to/paper --manuscript paper.tex

This creates .hci-paper/ with a context file and separate ledgers for claims, sources and endorsement, figures, reviewer comments, revisions, and verified venue policies. It also creates runs/ for comparable audit reports. The initializer has no external dependencies, makes no network requests, and refuses to overwrite an existing workspace.

Use --dry-run to inspect the plan first, or --json to integrate it into another tool.

The workspace carries an ordered lifecycle:

framing -> study-ready -> evidence-frozen -> claim-locked -> drafted -> reviewed
-> response-ready -> submission-ready

Advance one checked gate at a time:

python3 skills/hci-paper-writing/scripts/project_workspace.py /path/to/paper \
  --advance study-ready --note "RQs, ethics, and study plan reviewed"

The script prevents silent stage-skipping. It does not pretend that a state file can certify research quality; the evidence note and underlying artifacts remain reviewable.

Integrate notes without stealing their claims

hci-integrate handles chats, emails, meeting transcripts, reviewer comments, and lab notes. It first records who proposed each idea, whether the author endorsed it, what evidence supports it, and whether it may be used. Tentative or rejected ideas stay out of the manuscript unless the author adopts them.

hci-positioning separately checks whether the paper establishes importance, states a defensible unresolved question, names the smallest honest novelty delta, and chooses a contribution noun whose review contract the evidence can actually satisfy.

From reviewer concern to verified change

hci-panel gathers role-separated readings before synthesis. If the platform cannot run truly independent reviewers, the skill says so instead of pretending that sequential personas are independent evidence.

hci-rebuttal then assigns stable IDs to concerns and tracks this chain:

reviewer concern -> evidence -> response claim -> manuscript change -> verification

The point is not to produce a more confident rebuttal. It is to prevent a good response letter from drifting away from the paper the committee will actually read.

How it works

Contribution first, prose second

"We conducted a user study" describes an activity. The skill asks what the study reveals, validates, enables, or changes for HCI before treating it as a contribution.

Match the method lens to the paper

Qualitative, quantitative, design-research, systems, field/CSCW, and mixed-method papers call for different standards. The skill selects the lens that fits the paper instead of defaulting to an experiment.

Trace criticism back to evidence

Each major criticism must identify its basis in the manuscript, explain the consequence, and suggest a credible repair. Strong author claims need evidence, a citation, or narrower wording.

Keep corpus comparisons traceable

When comparing a manuscript with a declared corpus, the hci-grounded protocol records the corpus, years, query, filters, coverage gaps, and inspected sources. Embeddings can retrieve nearby work, but similarity alone does not establish novelty, quality, or likely acceptance.

Check current venue policy

Venue rules change. The skill requires agents to verify deadlines, length, anonymization, accessibility, ethics, AI-use disclosure, and supplementary material rules from current official sources at run time.

State the privacy boundary

The local scanner never sends a manuscript anywhere. Data handling for semantic review depends on the AI platform running the skill, not this repository. Do not upload a confidential paper you are reviewing without explicit authorization and policy support.

What's inside

skills/hci-paper-writing/
├── SKILL.md                         # Router, guardrails, output contracts
├── agents/openai.yaml               # Discoverable UI metadata
├── assets/                          # Intake and revision templates
├── scripts/
│   ├── manuscript_audit.py          # Local deterministic preflight
│   ├── project_workspace.py         # Safe persistent paper-state initializer
│   ├── release_guard.py             # Block private artifacts from releases
│   └── validate_skill.py            # Zero-dependency package validator
└── references/
    ├── contribution-types.md
    ├── chi-trends-2022-2026.md
    ├── chi-trends-2022-2026.zh-CN.md
    ├── evidence-grounding.md
    ├── novelty-and-positioning.md
    ├── workflows.md
    ├── method-lenses.md
    ├── study-evidence.md
    ├── section-patterns.md
    ├── reviewer-checklist.md
    ├── reviewer-panel.md
    ├── rebuttal-and-revision.md
    ├── project-workspace.md
    ├── source-integration.md
    ├── llm-systems.md
    └── policy-and-privacy.md

Design principles

1. The author retains judgment. The agent structures the critique but does not decide what the research should claim.
2. The skill does not invent citations, participants, policies, statistics, or reviewer consensus.
3. Research strength and venue fit are reported separately. The skill does not assign an acceptance probability.
4. Rigor is evaluated within the paper's research tradition.
5. Feedback includes its evidence, severity, proposed repair, and a way to check the revision.

Roadmap

• Contribution diagnosis and claim-evidence workflow
• Three-lens HCI red-team
• Local Markdown/LaTeX/text preflight
• Qualitative, quantitative, design, systems, field, and mixed-method lenses
• Privacy and live-policy guardrails
• English / Simplified Chinese README
• Persistent per-paper workspace and machine-readable ledgers
• Reverse outline and LaTeX figure/table consistency checks
• Role-separated HCI reviewer panel and meta-review protocol
• Reviewer-comment, rebuttal, revision, and verification loop
• DOCX, anonymity, LaTeX dependency, bibliography, and citation-key checks
• Ordered paper lifecycle with non-skippable stage transitions
• Source ownership, endorsement, permission, and evidence tracking
• Four-axis contribution form, research area, tradition, and venue model
• Bilingual CHI 2022–2026 trend atlas and CHI 2027 scenarios
• Facet-aware threat-set analysis and bilingual search-stability checks
• Public-release privacy guard
• HCI Paper Coach Benchmark: reproducible release of the five-year CHI analysis
• Rights-cleared accepted/rejected submissions, reviews, and rebuttals
• Public benchmark release with versioned splits, annotations, and data cards
• Numeric and terminology consistency across text, figures, and tables
• Subcommunity packs for CSCW, DIS, UIST, accessibility, and health HCI
• Bilingual Chinese/English report templates
• Longitudinal reviewer-feedback comparison across manuscript versions

HCI Paper Coach Benchmark: in progress

The bilingual 2022–2026 trend atlas is available now. We are still preparing the underlying reproducible benchmark package for method-aware HCI paper feedback. That release will report three data sources separately:

1. A five-year CHI corpus for embedding-assisted analysis by contribution, method, topic, and subcommunity. The release will document exact coverage, retrieval settings, source links, and known gaps. We will index or redistribute full text only when the license permits it.
2. Accepted and rejected submission histories contributed by authors. A history may include the manuscript, reviews, rebuttal, decision, and revisions. Each artifact must be cleared for use, allowed by the applicable venue policy, and de-identified where necessary.
3. Synthetic and expert-annotated cases for evaluating contribution diagnosis, claim-evidence alignment, method fit, reviewer-risk recovery, actionability, and false positives.

The benchmark has not been released. The current skill has not been trained or validated on private CHI reviews.

Contributions can include synthetic failure cases, annotation protocols, retrieval or evaluation code, public metadata sources, and authorized submission histories. Contributors will receive priority invitations to the private beta, subject to capacity and data/consent checks. See CONTRIBUTING.md. Do not attach confidential material to a public issue.

Contributing

Useful contributions include method-specific failure cases, synthetic test manuscripts, and current official policy sources. See CONTRIBUTING.md.

If the project helps you catch a serious problem before review, consider starring the repository so other HCI researchers can find it.

Responsible use

• Use the project for formative author feedback, not automated peer-review gatekeeping.
• Verify citations and venue policies yourself.
• Follow your venue's AI-use disclosure policy.
• Do not submit confidential manuscripts belonging to others to third-party models.
• The project is not affiliated with or endorsed by ACM, SIGCHI, CHI, CSCW, DIS, UIST, or any other venue.

License and provenance

MIT licensed. See LICENSE and THIRD_PARTY_NOTICES.md. Release history is recorded in CHANGELOG.md.

The project is independently authored and informed by public HCI submission guidance and open academic-writing workflows. No third-party source code is vendored.