Paper Init

Paper Init

Build the local context bundle a later session needs to implement a research paper: the paper itself, its figures and references, what it specifies, what it leaves open, and its data. Context only — no implementation code, no stack setup. Everything the implementer needs must end up on disk; nothing may depend on this conversation's context.

The bundle

One paper/ directory at the project root holds everything paper-derived:

• paper/README.md — bundle index: what this is, the read-first order, current status. Under ~50 lines.
• paper/source/ — pristine originals only: the PDF (supplementary material included) plus the best machine-readable form available (EPUB, HTML, XML, arXiv LaTeX source), exactly as acquired. Derived files (text dumps, extracted markup) never live here.
• paper/source/metadata.json — the provenance record (step 2).
• paper/paper.md — the faithful working transcription (step 3).
• paper/tables.md — overflow tables, only when the transcription takes its escape hatch (step 3).
• paper/figures/ + paper/figures/README.md — every figure, one README line each (file, dimensions, caption) with known defects — duplicates, missing diagrams — at the top.
• paper/references.md — every reference, annotated: role (dataset, baseline, method), resolved link, and fetch priority.
• paper/cited/ — pristine sources of the works the paper, or a fetched work in turn, defers load-bearing content to (step 3), when any were fetched. Nothing derived lives here.
• paper/spec.md — what the paper actually specifies (step 3), ending with the targets the implementation is later judged against.
• paper/open-questions.md — what the paper leaves open (step 3).
• paper/dataset.md — dataset provenance and re-acquisition, when the paper uses datasets (step 6).
• paper/reference-code/ — vendored existing code, when it exists (step 4); gitignored, its provenance tracked in metadata.json.
• data/ — the datasets themselves, gitignored, when any were acquired (step 6).

Every emitted file is tool-agnostic: "a fresh session", "the implementing agent" — never a named AI product, never /optimus:* commands. Only the final chat message may name /optimus:gauntlet.

1. Resolve the paper

The invocation argument is a URL, a local PDF path, a DOI, or an arXiv id; if none was given, ask for one. Resolve DOIs and arXiv ids to the source of record. If the paper is inaccessible (paywall, dead link), say so plainly and either stop or proceed from a file the user supplies.

A .git file (rather than a directory) marks a linked worktree — a real repo; proceed normally. With neither .git/ nor a .git file, read $CLAUDE_PLUGIN_ROOT/skills/init/references/multi-repo-detection.md and apply it: the bundle goes inside the target repo, not above it. When it detects a multi-repo workspace, ask which repo the paper work targets before writing anything — bundle, .gitignore, and README block all land there; when it finds no recognized structure, work in the current directory.

2. Acquire sources

Download into paper/source/, redundantly: the PDF whenever one exists (an HTML-only paper's publisher full text is the primary source), any supplementary material, plus the cleanest structured full text the publisher offers — the transcription cross-checks formats against each other. For arXiv papers, also pull the e-print source bundle (https://arxiv.org/e-print/<id>) when offered: the LaTeX source makes math transcription near-mechanical and ships figures at native resolution. Pull figure rasters into paper/figures/ from whichever source has the best resolution (PDF-embedded usually beats web-served); keep native formats, never re-encode, and write paper/figures/README.md as they land — one line per figure (file, dimensions, caption from the paper text), known defects (duplicates, missing diagrams) at the top. Installing transient fetch or extraction tooling along the way (a PDF library, gdown, pandoc) is fine — that is not the project stack — but install it isolated (pipx, a scratch venv, pip install --target into a temp dir), never into the project's own environment.

metadata.json records at minimum: title, authors, venue, published, doi, url, license, downloaded (date), code_available (with the paper's own availability sentence when it states one), dataset_referenced (name, URL, whether the paper redistributes it — full provenance and re-acquisition live in paper/dataset.md), and local_files — every file in source/ mapped to its role and exact acquisition record (URL or command, and date; for a file the user supplied, the path it came from). Add any further bibliographic fields the source offers. The test: a fresh clone can re-acquire every publicly fetchable file from this record alone.

3. Working forms

• paper/paper.md — a complete transcription, not a summary: mirrored section headings, math in LaTeX, figures as local relative links with their captions, tables inline (escape hatch: a separate tables.md when tables are numerous or large, linked both ways). Open with a provenance header naming the source of record. Its length is the paper's own.
• paper/spec.md — only what the paper states, each fact tagged with its section ref: data, architecture, training procedure (including the compute the paper states — hardware, training time, scale), evaluation, baselines, reported results. Architecture/training/eval tables carry a "defined enough to implement?" column. Anything inferred, chosen, or assumed is labeled as ours or moves to open-questions.md — never present our choices as the paper's. End with a short "targets worth holding the implementation to" section — the quality bar the implementation is later judged against, so it must stand on its own: the externally meaningful results (numbers, and figures where the claim is qualitative), each with the metric, data, protocol, and spread the paper reports, plus the cheaper checks the paper states along the way (dataset counts, parameter counts, a loss reached) that let part of the implementation be judged before a full run exists. A target the bundle cannot yet measure — data not on disk, a scorer or protocol undefined — says what is missing. Keep the file under ~200 lines.
• paper/references.md — every reference the paper cites, annotated: role (dataset, baseline, method), resolved link, fetch priority. Step 4 appends the reference-code summary here when code exists.
• paper/open-questions.md — everything undefined, ordered by how much it blocks work: missing hyperparameters, ambiguous procedures, figure/table defects (ledgered in figures/README.md — point there, don't duplicate), credibility issues, and a suggested framing for the implementation, tech stack included (drawn from the paper and the project's existing stack if one exists). An open point with a defensible default says so and where it comes from (reference code, a cited work, common practice), so an unattended implementation run proceeds instead of stalling on it. Mark a finding [verified] only when checked against the local files during this run — never for inference — open the file with a one-line legend saying what the mark means, and settle now whatever those files can settle: no verifiable-now TODO leaks into the implementation phase. Keep it under ~150 lines.

When the paper defers load-bearing content to a citation — an inherited architecture, a borrowed training procedure, a dataset defined there — the bundle's contract covers that content too: fetch each such cited work's sources now, as in step 2, into paper/cited/<slug>/, and record each in a cited_works array in metadata.json (title, identifier, slug, license, and the acquisition record — URL and date, or the path it came from for a user-supplied file). Pull the specific facts the implementation needs from it into spec.md or open-questions.md, tagged with provenance, and note in references.md why it was fetched — a work the paper itself never cites gets its own entry there. Apply the same test to each fetched work: when it defers content the implementation still needs to a further work, fetch that one too. Sources and targeted extraction only — no bundle per cited work. Expect a handful of works, not a bibliography crawl: the test is dependence, not relevance. An inaccessible cited work stays a visible gap — record what is deferred to it in open-questions.md and report it in the final message so the user can supply a copy.

If producing the bundle took mechanical extraction work a fresh session could not trivially redo (pulling rasters out of a PDF, dumping text from an EPUB), leave one small regenerator script that reproduces those derived artifacts from paper/source/ offline (placement follows project conventions; when the project has none, paper/). Stamp only script-produced files with a do-not-hand-edit header — the model-authored files (paper.md, spec.md, and the rest) stay hand-editable, and a re-run updates them in place. When you leave one, ensure .gitattributes pins the script's outputs — eol=lf for generated text, binary for extracted rasters — so regeneration stays diff-clean on any platform.

4. Reference code

Check the paper's own links and project page, then search for official or third-party implementations. When code exists: vendor it into paper/reference-code/ (gitignored — step 7), record its provenance as a reference_code object in metadata.json (upstream URL, exact commit or tag, vendor date, license) — that tracked file is what lets a fresh clone re-acquire the code; anything left inside paper/reference-code/ itself is gitignored away. Add a what-it-reveals summary to references.md (hyperparameters, architecture details, training procedure). It is reference material, never the implementation. When none exists, record code_available: false in metadata.json.

5. Feasibility

When the paper's experiments plausibly demand substantial compute — model training, large-scale simulation or rendering — assess feasibility before step 6 acquires anything big. Most papers have nothing to gate (a survey, a proof, a small-scale study); skip this step for them entirely.

• Draw the requirements from what spec.md recorded (step 3) and from the vendored reference code's own docs, which often state hardware. When neither states them, estimate from what the paper does record — model scale, dataset size, training steps — and label the figures as estimates.
• Detect the local GPU (model, VRAM) and RAM, and compare against what faithful reproduction needs. Detection tooling is not universal (no nvidia-smi on AMD or Apple Silicon machines): when it cannot answer, ask the user what the machine has — never read a failed detection as "no GPU".
• When the gap makes the reported targets unreachable in practice — not merely slower — ask once with AskUserQuestion: header "Hardware feasibility", the question naming the limiting factor and the concessions that would close the gap (reduced scale, a dataset subset, quantized or distilled variants, different hardware), options "Continue anyway" / "Reduce scope" / "Pause — line up other hardware first".
• Record the outcome where it binds. A reduced scope adjusts spec.md's targets section — that is the bar the implementation is later judged against — gets a line in open-questions.md, may shrink what step 6 downloads, and is recorded in metadata.json (the decision and the adjusted targets) so a refresh re-applies them rather than restoring the paper's reported targets. "Continue anyway" leaves the targets untouched and notes the hardware risk in open-questions.md. "Pause" stops the spend, not the bundle: step 6 writes re-acquisition steps into dataset.md instead of downloading, open-questions.md records the pause, and the remaining steps finish so the bundle commits complete and a re-run resumes it.
• Never write the hardware inventory itself into the bundle — it stays machine-agnostic; only the decision and its consequences go on disk.

No substantial compute, or no mismatch: no gate, and the final message says at most one line about feasibility.

6. Datasets

Identify every dataset the paper uses. Freely downloadable ones go into data/ now; verify what arrived (file counts, sizes, integrity) against what the source promises, and record the verified numbers. Write paper/dataset.md: provenance, exact re-acquisition commands, the verified counts, license and redistribution terms, and anything deliberately not downloaded. Keep it under ~200 lines. If the paper uses no external datasets, say so in one line of paper/README.md's status and skip dataset.md, data/, and the gitignore pair entirely.

Before a large download (GB-scale or hours of time), confirm with AskUserQuestion — header "Dataset download", question stating size and source, options "Download now" / "Skip — write re-acquisition steps only". Small datasets download without asking. When a download is blocked (auth, license acceptance, a manual form), do not ask — write the exact steps into dataset.md and flag it in the final message.

7. Gitignore and routing

Each .gitignore rule below is independent — apply every one whose condition holds, adding only what is missing:

• Datasets in use: data/* plus !data/README.md.
• Vendored reference code: paper/reference-code/.
• License: when metadata.json's license does not permit redistribution (typical for a paywalled publisher PDF), also paper/source/* and paper/figures/* with !paper/source/metadata.json and !paper/figures/README.md exceptions — a fresh clone re-acquires those from the metadata record, which must therefore stay committed. When the license forbids redistribution or derivatives (an ND clause), flag in the final message that paper.md is a full-length derivative: committing it is the user's call when the repo is or will become public.
• Cited works: the same license test applies to each cited work on its own, whatever the main paper's license — one whose license forbids redistribution adds its own paper/cited/<slug>/ directory, never the whole paper/cited/ (other works may be committable). A fetched work is re-acquirable from the record; a user-supplied one exists only at the path it came from — say so in the final message.

A pre-existing data/ line (directory form, common in ML repos) defeats the !data/README.md exception — git cannot re-include a file under an excluded directory. Narrow that line to data/* (same ignore coverage; exceptions become possible), note the change in the final message, and verify with git check-ignore -q data/README.md: it must exit non-zero, finding nothing to ignore (under -v, a match on the ! line is the exception working, not the file being ignored).

When the paper uses datasets, also write data/README.md: what goes here, the counts when known, license terms, and a pointer to paper/dataset.md. Not a git repo? Skip the .gitignore part and note it in the final message.

Write paper/README.md (the bundle index). If a root README.md exists, maintain one short routing block there pointing at the bundle(s), wrapped in marker comments that identify it as managed:

<!-- paper-context:start -->
## Paper context
...
<!-- paper-context:end -->

On re-run, rewrite only what lies between the markers, never duplicate the block; a ## Paper context heading without markers is the user's own — leave it alone and append a new marked block. If no root README exists, skip — the bundle indexes itself.

8. Final message

Close with: what the bundle contains and where; cited works — which were fetched and why, which the user supplied, and which remain gaps in open-questions.md (or that none were needed); dataset status (downloaded and verified, skipped, blocked — with the instructions pointer — or none); hardware feasibility (one line when fine, the mismatch and the recorded decision when not); any transient tooling installed; and the suggested tech stack as recorded in open-questions.md — a suggestion only, nothing is installed.

Then: commit the bundle first (e.g. with /optimus:commit, staying in this conversation) so the implementation loop starts from a clean, tracked baseline — the gitignored parts (data/, paper/reference-code/, license-ignored sources) exist only in this checkout, so a worktree or fresh clone re-acquires them from the records or copies them in. Then start a fresh conversation with /optimus:gauntlet, printed as one paste-ready line: the goal names the bundle root with its index as the read-first entry, the bar its spec.md targets section.

Re-running

Same paper — a resolved identifier (DOI, arXiv id) matches source/metadata.json, or one source names the other's identifier (an arXiv page listing the published DOI): refresh in place — update, don't duplicate, and keep the original acquisition records (append the refresh; a preprint's provenance is not overwritten by its published version's). A scope decision recorded in metadata.json (step 5) is not paper content: re-apply it to the regenerated spec.md targets rather than restoring the reported results. Cited works already in paper/cited/ are kept, not re-fetched; when the revision drops a citation the bundle fetched, ask before removing its directory. When the match is uncertain, ask before touching the existing bundle. A different paper while paper/ already holds one: use papers/<slug>/ (kebab-case slug from the title) as the bundle root everywhere — datasets go in papers/<slug>/data/ with their own data/README.md, and the gitignore entries spell full paths (papers/<slug>/data/* with its !papers/<slug>/data/README.md exception, papers/<slug>/reference-code/, papers/<slug>/cited/<work>/ only when step 7's license rule fires for that work): a pattern containing slashes anchors at the .gitignore location, so the bare step 6–7 paths cannot reach a nested bundle. Leave the existing bundle untouched and add the new one to the routing block. If papers/<slug>/ already holds a different paper, disambiguate the slug (append the year or venue) — never refresh a bundle that is not the same work. Never merge two papers into one bundle; never move an existing paper/ — that restructuring is the user's call.

Boundaries

• Never write under docs/specs/ or docs/product/ — /optimus:tdd auto-detects build specs there and /optimus:brainstorm scaffold owns the steering cascade. The transcription mirrors the paper's own headings verbatim, Scenarios included — spec auto-detection reads only those two directories, so no bundle heading can misfire it.
• Never write .claude/.optimus-version (owned by /optimus:init) and never edit .claude/CLAUDE.md (regenerated by init).