Auto Paper

SkillDev tools

Internal Harness instruction source for auto-paper. Route through visible Harness aliases instead of invoking directly.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Auto Paper skill

What this skill tells your AI

The instructions your AI receives, as published by linzhe001/harness-research in .agents/skills/auto-paper/SKILL.md and read by ahel’s review.

Use this as the auto-paper orchestrator, not a polishing tool. It routes phases and checks artifacts without inventing paper state.

Boundaries

  • Author evidence outranks exemplars: local results, figures, tables, notes, approved claim boundaries, and operator facts.
  • Do not invent experiments, results, mechanisms, statistics, novelty, clinical meaning, data availability, citations, or source support.
  • Blogs, reviews, surveys, and tutorials are citation-supported by default unless the operator explicitly asks for an uncited opinion memo.
  • External papers are examples for structure, style, terminology, and positioning, not author claims.
  • Patch LaTeX only from written artifacts.
  • Read experiment results primarily through docs/30_evidence/Experiment_Evidence_Index.{json,md} and artifact paths named there. Direct iteration_log.json reads are weak intent signals; paper claims require reports, configs, logs, metrics, or run artifacts.
  • Read docs/context/experiments.md when present for Research Wiki context, stable observations, run requests, assurance gaps, and open questions. Legacy docs/45_discoveries/Research_Wiki.md is a fallback before dynamic-context-v2 migration; neither file is approval.
  • Do not write iteration_log.json, .auto_iterate/**, or hidden controller state. Durable paper artifacts live under auto_paper_output/<paper_id>/.

Required Config

Before the loop, establish paper_id, workflow, venue/target, draft path, TeX roots, bibliography paths, figures, materials, references, artifact dir, optional experiment evidence index, output language, citation target count, figure/table requirement status, compile command, human gate policy, and forbidden directions in auto_paper_output/<paper_id>/config.yaml or an intake artifact. If missing, route to $auto-paper-intake or return USER_GATE.

After Grill establishes an Automation Policy, paper phases auto-proceed within that policy. Use USER_GATE only for missing operator intent, explicit approval tools, or irreversible external submit.

Use the deterministic checker when the operator requests an automatic paper chain:

python tooling/auto_paper/scripts/auto_paper/artifact_check.py \
  auto_paper_output/<paper_id> --phase <phase> --json

The checker reports artifact readiness only; it does not invent manuscript content or prove scientific claims.

Phase Route

Run in order unless an audit routes back: $auto-paper-research, $auto-paper-argument, $auto-paper-citation, $auto-paper-layout, $auto-paper-patch, $auto-paper-harden.

Optional branches: $auto-paper-response, $auto-paper-data, $auto-paper-figure. Enter $auto-paper-figure when the operator asks for figure work, TeX or source files mention needed figures/tables, or harden finds figure/caption risk.

Artifacts

Durable decisions live under auto_paper_output/<paper_id>/.

  • Intake: config.yaml, source_index.md, tex_inventory.json, intake_report.md, experiment_source_map.md, optional figure_requirement_scan.md
  • Phase artifacts: research dossier, style/gap maps, confirmed motivation, claim register, citation support bank, section blueprints, rationale matrix, citation and patch plans, figure contracts, patch ledgers, audits, and final gate ledger
  • run_request_register.{json,md} when missing experiments are needed

Use references/artifact-contract.md for identifiers and stale-artifact policy.

Gate And Routing

Do not patch around missing upstream artifacts. Every audit finding names root_cause, owning_phase, required_artifact, fix_action, and downstream_risk.

Decisions: NEXT_UNIT, REWORK_RESEARCH, REWORK_ARGUMENT, REWORK_CITATION, REWORK_LAYOUT, REWORK_PATCH, RUN_REQUEST, USER_GATE, COMPLETE, ABORT.

RUN_REQUEST means writing found missing experiment evidence for $run. Record the blocking claim, needed evidence, minimum artifacts, suggested prompt, and acceptance check in run_request_register.{json,md}.

When a draft introduces, narrows, or removes a result claim, record Claim Delta Evidence in the phase artifact or Gate ledger. Unsupported claims become RUN_REQUEST, not a default approval pause.

References

Load only active-phase references from references/: loop, auto-iterate boundary, artifact contract, writing rationale, motivation thread, citation support, and LaTeX source control.

Use deterministic helpers when available:

  • .agents/skills/auto-paper/scripts/figure_requirement_scan.py
  • ../../../.agents/references/research-supervision-patterns.md
  • ../../../.agents/references/research-supervision/README.md

Output

Final responses report artifact manifest, modified files, guard/compile results, remaining reviewer risks, and next human gates.

Signals

GitHub stars
86
Forks
9
Last commit
Aug 2026
Advanced
Catalog kind
skill
Gateway key
auto-paper
Source
github.com/linzhe001/harness-research