academic-writing-toolkit

academic-writing-toolkit

agent-native, local-first workflows for evidence-controlled literature review, thesis writing, citation auditing, reference verification, and release governance.

Academic Writing Toolkit is a public toolkit for researchers who want their local AI agent to work through academic writing tasks with repeatable files, checks, and evidence controls. Install it into a writing project, open that project in Codex, Claude Code, Gemini CLI, Cursor, or another compatible agent host, and use the skills as a structured research workflow.

This is not a SaaS product and it does not host your thesis. Your chapters, PDFs, reading notes, evidence registers, release packets, and exports stay in your clone. Optional reference metadata checks use external APIs only when you explicitly run them with --online.

Product Surfaces

Local agent skills are the full workflow. Use them when an agent can read and write the project files in your clone: chapters, reading notes, evidence registers, release packets, and export outputs.

The Codex plugin packages those same local skills for Codex plugin installation. It is a distribution surface, not a separate workflow.

The ChatGPT App MCP server is narrower. It provides pasted-text checks and template generation through temporary files only; it does not read or write a local thesis project, run the full skill pipeline, or persist user submissions.

10-minute demo

Inspect the runnable demo project:

python3 scripts/verify-refs.py --bib examples/demo-project/references.bib --json
python3 .claude/skills/evidence-review/scripts/check_review_package.py examples/demo-project --strict
python3 .claude/skills/release-governance/scripts/check_release_packet.py examples/demo-project --json

Then open examples/demo-project/ in your agent host and ask it to explain the local workflow.

Common use cases

• Write a literature review
• Audit thesis citations
• Verify references before submission
• Prepare a release-governance packet
• Choose the right product surface

Workflow

/read -> /note -> /map -> /evidence-review -> /integrate -> /audit -> /release-governance -> /style -> /logic-review -> /export
             |                                      |
             v                                      v
        /verify                               /verify-refs
             |
             v
        /progress

/human-eval-handoff-repair supports evaluation-package QC and annotation repair workflows.

Quick Start

Use git clone, not GitHub’s “Download ZIP”. This repository uses symlinks under .agents/skills/ so Codex, Gemini, and other local agent hosts can discover the same skills as Claude Code.

git clone https://github.com/yha9806/academic-writing-toolkit.git my-writing-project
cd my-writing-project
make setup
make doctor

Then open the folder in a local agent runtime and ask what skills are available. You should see the public academic writing skills listed below.

The intended use is the same as local Superpowers-style skills: install the toolkit into a project folder, let the agent discover the skills from local files, and drive the workflow from natural language or slash commands.

Agent Runtime Support

Runtime	Local discovery path
Claude Code	.claude/skills/
Codex	.agents/skills/
Gemini CLI	.agents/skills/
Cursor	.cursor/rules/ baseline rules

Setup guides live in docs/setup-claude-code.md, docs/setup-codex-cli.md, docs/setup-gemini-cli.md, and docs/setup-cursor.md.

Skills

Skill	Purpose
/read	Read academic PDFs page by page with structured output
/note	Record reading notes in the shared notes format
/verify	Fact-check factual claims during reading
/map	Map sources to chapters and coverage gaps
/evidence-review	Build evidence-controlled gap maps, claim registers, citation plans, and overclaim audits
/integrate	Propose and apply approved note-to-chapter integrations
/audit	Check numbers, terminology, cross-references, and citations
/release-governance	Prepare release, rebuttal, artifact, and claim packets with ref-artifact-gate controls
/style	Check and safely fix British English spellings
/logic-review	Review paragraph flow, transitions, and argument continuity
/verify-refs	Validate BibTeX records offline or with explicit metadata checks
/human-eval-handoff-repair	Validate, repair, and map human-evaluation handoff packages and filled annotation CSVs
/progress	Show reading, writing, and coverage progress
/export	Convert Markdown outputs to .docx and ZIP packages

Detailed guides live in docs/skills/.

The local agent skill guides are organized by command name; goal-oriented guides live in docs/use-cases/.

Quality Checks

Agent-facing skills call deterministic scripts so checks are repeatable in CI and easy to run manually:

python3 scripts/audit-citations.py --base-dir . --style harvard --json
python3 scripts/audit-citations.py --base-dir . --style harvard --fix-safe --apply

python3 scripts/audit-british-english.py --base-dir . --json
python3 scripts/audit-british-english.py --base-dir . --fix

python3 scripts/audit-logic.py --base-dir . --json

python3 scripts/verify-refs.py --bib references.bib --json
python3 scripts/verify-refs.py --bib references.bib --json --online
python3 scripts/verify-refs.py --bib references.bib --json --online --metadata-dir path/to/metadata-fixtures

python3 .claude/skills/release-governance/scripts/check_release_packet.py .

Safe fixers are intentionally narrow. Citation fixes only apply conservative formatting changes such as Harvard comma normalisation; British English fixes only apply whole-word spelling replacements from the built-in map. Paragraph logic and reference metadata checks report findings for agent or user review. Release packet checks are also narrow: they validate files, columns, evidence-state values, parseability, local path leakage, and unresolved template markers, but they do not judge scientific validity or venue compliance.

Project Layout

my-writing-project/
├── .claude/skills/          Claude Code skills
├── .agents/skills/          Symlinks for Codex, Gemini, and compatible agents
├── .cursor/rules/           Cursor baseline rules
├── apps/                    ChatGPT App MCP server
├── chapters/                Your chapter drafts
├── literature/
│   └── reading_notes/       One notes file per source
├── plugins/                 Codex plugin package
├── release/                 Optional release governance packets
├── final_output/            Exported documents
├── scripts/                 Deterministic helper checks
├── CLAUDE.md                Canonical project configuration
├── AGENTS.md                Generated local-agent configuration
└── GEMINI.md                Generated Gemini configuration

Configuration

Edit CLAUDE.md for project-specific settings, then run:

make sync

AGENTS.md and GEMINI.md are generated from the shared block in CLAUDE.md; do not edit them directly.

Key settings include:

• chapter, literature, notes, and export directories
• reading page limits
• British English writing policy
• citation style (harvard, apa, chicago-author-date, mla, ieee, vancouver, gb-t-7714-2015)

Testing And Maintenance

make doctor   # read-only environment check
make repair   # repair symlinks and generated configs where possible
make test     # full regression suite
make sync     # regenerate AGENTS.md and GEMINI.md
make plugin-sync   # regenerate the Codex plugin skills from .claude/skills
make plugin-check  # validate plugin metadata, sync state, and bundled helpers
make chatgpt-app-check  # run the ChatGPT App MCP server checks

For Codex plugin release preparation, use docs/plugin-publishing-checklist.md. For ChatGPT App deployment and submission preparation, use docs/chatgpt-app-publishing.md. For v0.3.0 release preparation, use docs/product/v0.3.0-release-readiness.md.

The regression suite covers local skill discovery, config sync, export assumptions, citation auditing, public-content cleanup, British English checks, paragraph-logic checks, human-evaluation handoff repair packaging, and offline plus fixture-backed reference verification.

Reference Verification

/verify-refs uses a deterministic offline core by default:

• parse BibTeX from .bib files or Markdown code fences
• check required fields by entry type
• detect duplicate keys
• validate DOI, URL, and arXiv identifier shape

For explicit metadata verification, run with --online. The script checks CrossRef for DOI records, Semantic Scholar as a secondary DOI/arXiv source, and arXiv for preprint identifiers. Tests use --metadata-dir fixtures so CI stays deterministic; live network calls are only attempted when --online is set and no matching fixture is present.

Project-specific self-citation rules are intentionally excluded from this public toolkit.

License

MIT. See LICENSE.