ThinkReader Codebase Skill

/codebase turns a paper into a clean reproduction workspace: enough grounded detail for an engineer to implement the method and verify it against the paper without pretending that missing details are known.

Required reading order

1. Read prompts/base.md and add prompts/pro.md for deep mode.
2. Read rules.md, tools.md, writer.md and data/quality.json.
3. Read the App-provided codebase context, PaperMap, relevant full sections, equations, figures/tables and declared resources.
4. Existing annotations are optional context, never a prerequisite. If the user separately requests a reproduction-focused annotation pass, use /annotation or /deep-annotation; do not make it a hidden dependency.

File ownership

• prompts/base.md owns implementation judgment and SPEC quality.
• prompts/pro.md owns the deep-mode delta.
• rules.md owns file layout, evidence and acceptance-target consistency.
• tools.md owns assumed context, resource, evidence, validation and output capabilities.
• data/quality.json owns deterministic quality floors.
• writer.md owns proposal, validation and commit flow.

Non-negotiable method

• Separate paper-stated facts, evidence-backed inferences, implementation decisions and unknowns.
• Preserve equations, symbol names, shapes, hyperparameters, datasets, metrics and conditions exactly where stated.
• Make SPEC.md and results.target.json agree on every target value and acceptance condition.
• Let the LLM infer the architecture and write the specification; use tools for exact evidence copying, resource inspection, JSON assembly, cross-file consistency checks and output writes.
• Never invent an omitted implementation detail; record it as a reproducibility risk or explicit decision.

Boundaries

Outputs live only under output/codebase/<paper_id>/. The semantic proposal never mutates PDF or Intelligence. After commit the App first appends the codebase tag to an existing annotation covering the cited evidence, including source URLs; preserve its content, author, marks, hints, replies and layout. Keep uncovered evidence and source URLs in the output citations; do not create PDF cards for evidence or URL bookkeeping. Repeated runs must not multiply cards or tags; matching and tag writes are tool work, not extra LLM output. Legacy cleaner/HTML Reader outputs, public CLI commands and old schemas are not inputs.