01 · RESEARCH
STAR
Systematic Toolchain for AI Research
Runs one research project from idea through reproducible experiments to paper-ready evidence.
STAGE
Every STAR needs a STAGE
The academic-writing companion to STAR — it turns a research project into a submitted paper without losing the chain of custody: every number traces to a fingerprinted evidence file or is visibly marked as missing, and every claim is tracked from proposal to verification in a single ledger.
15Workflow Skills
4AI Agents Supported
2Languages · EN 中文
MITOpen Source
The STAR family
One research life cycle.
Three focused toolchains.
Each works independently. Connect them when you want upstream work to flow downstream as fingerprinted, read-only evidence.
02 · PAPER
Current
STAGE
Systematic Toolchain for Authoring, Guiding, and Editing
Turns one research contribution into a traceable paper, review cycle, and submission package.
03 · THESIS
STORY
Systematic Toolchain for Organizing Research over Years
Shapes graduate research into a defensible master's thesis or doctoral dissertation, defense, and deposit.
Why STAGE
Every number accounted for.
Every claim on file.
Papers rot when numbers live in chat history and claims live in someone's head. STAGE moves the whole chain of custody — evidence, claims, reviews, submissions — into the repository.
A consistent manuscript layout
Sections, figures with their sources, tables, bibliography, and venue styles each have one predictable home under manus/ — and it compiles out of the box.
A read-only evidence layer
STAR artifacts or hand-registered drops are snapshotted into mates/ with a fingerprint per file in MANIFEST.md. Evidence flows one way — never edited in place.
The claim ledger as the hub
notes/claims.md links every claim's statements ⇄ evidence ⇄ status — proposed → drafted → verified — from the story seed to the submission.
The fabrication boundary
Every number in manus/ either traces to a fingerprinted mates/ entry or is written as \todo{...} — no third state. Assertions about cited papers must be checkable against a reading note; venue rules enter only as user-confirmed facts. A visible \todo is a state; a plausible invented number is a defect.
Checks in scripts, judgment in skills
bash execs/run.sh builds out-of-tree with latexmk; lint.sh mechanically fails on undefined references, \todo markers, page-limit overruns, and anonymity leaks. Everything that needs judgment stays with the sixteen skills.
Submission cycles as data
Each venue attempt lives in cycls/<venue>_<year>/ with a user-confirmed venue.yml, received and simulated reviews, the response, and a frozen submission record — rejections included, so the next attempt starts where the last one ended.
AI-agent native, STAR-paired
The same sixteen skills ship for Claude Code, Codex, Cursor, DSH, Kimi Code, Pi, and Qwen Code — in English and Chinese, with shared skill files under .agents/skills/ and one neutral request router under .agents/commands/. Pair with a STAR repo for one-command evidence import, or work standalone with hand-registered drops.
Project structure
Eight directories.
One chain of custody.
The abbreviated names follow STAR's convention: short to type, unambiguous to search, and identical across every paper you will ever write on STAGE.
STAGE/
├── manus/ # the manuscript
│ ├── main.tex # entry — always compiles as the preprint
│ ├── secs/ # sections: <n>_<slug>.tex
│ ├── figs/ # rendered PDFs; figs/srcs/ their sources
│ ├── tabs/ # tables, generated from evidence
│ ├── bibs/ # reference.bib
│ └── stys/ # stage.cls (the look) + stage.sty (macros)
├── mates/ # imported evidence — read-only
│ └── MANIFEST.md # the fingerprint ledger
├── notes/ # story, claims, outline, notation
│ └── refs/ # reading notes per paper
├── cycls/ # submission cycles: <venue>_<year>/
├── tasks/ # revision scratch and promise lists
├── wkdrs/ # builds & reports (gitignored)
├── execs/
│ ├── run.sh # ← the single build entrypoint
│ ├── update.sh # sync skills; --adopt installs the skeleton
│ └── scpts/ # import.sh · lint.sh
├── .env # machine-local config (never committed)
└── AGENTS.md # shared human + agent instructions
manus/Manuscripts — the paper's LaTeX sourcesmates/Materials — fingerprinted evidence snapshots, read-onlynotes/Writing metadata — story, outline, and the claim ledgercycls/Cycles — one directory per submission attempttasks/Tasks — revision scratch and promise checkboxeswkdrs/Work directories — builds and ephemeral reports, never committedscpts/Scripts — evidence import and deterministic checksWriting workflow
Sixteen skills.
One chain of custody.
From imported evidence to a frozen submission — each skill writes its state into project files, so the whole paper is auditable and resumable at any point.
✦
Set upWire the repo and bring in fingerprinted evidence.
/stage-proj-adopt †Wire a new or existing paper repo into STAGE — numbers already in the draft become unsourced claims, not silent debt.
/stage-evid-curatorRun the import, register hand-dropped files, normalize messy exports, propose claim⇄evidence mappings.
✦
PlanShape the story and skeleton the paper.
/stage-stry-coach †Dialogue-first story shaping: pitch, contributions with claim IDs, venue rationale — seeds the claim ledger.
/stage-outl-planner †Story → skeleton: section page budgets, figure and table plans, claim→section assignment.
/stage-refs-curatorBibliography hygiene, reading notes with citable facts, topic search for work nobody named, related-work positioning.
✦
WriteDraft from evidence, never from memory.
/stage-sect-drafterOne section per run, from its brief, mapped evidence, and claims — missing numbers become \todo{}.
/stage-tabs-builderTables from mates/ evidence only — one % src: fingerprint comment per data row.
/stage-figs-designerEach figure end to end: purpose, editable source under figs/srcs/, rendered PDF.
✦
Polish & auditTrace every number. Check every citation.
/stage-copy-editorClarity, flow, and notation consistency — meaning and numbers untouchable.
/stage-clms-auditorExtracts every number and traces each to a fingerprint: matched / mismatched / unsourced.
/stage-cite-auditorEvery \cite resolves; every assertion about a cited work checkable against a reading note.
✦
Submission cycleReview, respond, freeze.
/stage-peer-reviewerA five-perspective simulated program committee under a citation-integrity contract.
/stage-resp-writer †Reviews → point ledger → response within the venue's limit, every promised change a checkbox.
/stage-subm-packer †Preflight, venue-template conversion from the official kit, package, submission record, freeze tag.
✦
Read acrossReport the whole flow, name the next action.
/stage-flow-statusRead-only map of the whole flow — outline, ledger, manifest, cycle state — and the single next action with its exact command. Never writes.
† marks the six slash-only skills: they run only when you name them, because the
decision points — adoption, story, outline, response, submission, and poster selection — stay yours.
Quick start
Five steps to a compiling paper.
Start fresh from the template — or run update.sh --adopt
inside an existing paper repo and STAGE installs its skeleton around your draft without moving a single file.
# 1 · Start a paper repo from the STAGE template — one paper, one repo
❯ git clone https://github.com/wanghao9610/STAGE my-paper && cd my-paper
❯ rm -rf .git && git init # drop STAGE's history
# 2 · Configure the local runtime
❯ cp .env.example .env && $EDITOR .env # STAR_HOME (optional) · LATEX_ENGINE · ANON
# 3 · Import evidence from the paired STAR repo — or drop files into mates/manual/
❯ bash execs/scpts/import.sh
# 4 · Build & lint — a fresh checkout compiles out of the box
❯ bash execs/run.sh
❯ bash execs/scpts/lint.sh
# 5 · Pick up the workflow — it names the single next action for you
❯ /stage-flow-status
❯
Ready when you are
Give your paper a
chain of custody worth submitting.
STAGE is free, MIT-licensed, and agent-agnostic. Use it as a GitHub template, or adopt it into the paper you already have.