Harness Auto Paper

SkillDev tools

Build, revise, audit, or rebut a research manuscript from Harness Research evidence. Use for paper intake, related-work research, claim and citation control, section architecture, LaTeX patching, figure planning, reviewer response, or final manuscript hardening when the workspace contains Harness research artifacts.

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 Harness Auto Paper skill

What this skill tells your AI

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

Treat this as an optional manuscript loop attached to the core $write Recipe. Do not turn it into a required research stage.

Establish the source boundary

  1. Locate the Harness workspace root and read harness.yaml, the current Research Brief, relevant experiment context, and release evidence.
  2. Prefer indexed reports, configs, logs, metrics, figures, and Gate ledgers over summaries. Treat iteration_log.json as experiment routing state, not sufficient support for a manuscript claim.
  3. Keep verified facts, interpretations, proposed prose, and open evidence gaps distinguishable.
  4. Never invent experiments, values, mechanisms, novelty, citations, approvals, or compile success.

If the framework checkout is not available at tooling/auto_paper/, continue manually with the same artifact contract and report that deterministic checks were NOT_RUN.

Run the bounded loop

Use one artifact root: auto_paper_output/<paper_id>/. Read phase-contracts.md only for the active phase.

  1. intake — establish target, draft, sources, TeX roots, figures, compile command, claim boundary, and operator constraints.
  2. research — build the source and exemplar dossier; learn organization and terminology without borrowing claims.
  3. argument — create a claim register, explicit scope limits, and claims to avoid.
  4. citation — map each citation-dependent claim to checked support and mark unsupported claims.
  5. layout — define section roles and a rationale matrix before editing prose or TeX.
  6. patch — patch only from accepted artifacts and keep a patch ledger.
  7. harden — run claim, citation, revision, integrity, and compile checks; write the final Gate ledger.

Optional branches are figure/table work, data statements, and reviewer response. Return to the owning phase when an audit finds an upstream defect.

Use artifact-driven state

The optional plugin has no controller or hidden resumable state. Keep the phase artifacts under auto_paper_output/<paper_id>/ and advance only when the active phase contract is satisfied. Use the deterministic checker before moving forward:

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

The checker proves only artifact-contract readiness. It does not prove a scientific claim or a successful compile. Record any skipped checker or compile step as NOT_RUN in the phase or final Gate ledger.

Route evidence gaps

When a claim needs a missing experiment, write a bounded entry in run_request_register.json or .md with the blocking claim, minimum evidence, expected artifacts, and acceptance check. Route that request to $run; do not fabricate a result or silently weaken a central claim.

Pause for the operator only when intent is missing, an explicit approval record is required, or an irreversible external submission is requested. Ordinary paper iterations proceed inside the accepted Automation Policy.

Report the result

Return the artifact root, files changed, evidence used, deterministic checks, compile result, unresolved reviewer risks, and the next human decision. A release-ready manuscript requires both Conclusion Evidence and an explicit Gate ledger.

Signals

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