Academic Research Skills: human-in-the-loop paper pipeline for Claude Code

What they did

Installed the academic-research-skills plugin into Claude Code via the plugin marketplace, then used the /ars-plan skill to walk through a paper's structure through Socratic dialogue. From there they ran the staged pipeline: defining the research question, gathering and verifying references, drafting sections, formatting citations, and passing through the Stage 2.5 and Stage 4.5 integrity gates, which run a 7-mode blocking checklist against known AI-research failure modes (hallucinated results, fabricated methodology, citation hallucinations, and similar). They kept the human-in-the-loop contract intact, confirming at each stage rather than letting the pipeline run end to end. They also enabled the optional claim-audit pass (ARS_CLAIM_AUDIT=1) so that each citation's locator anchor was fetched and checked against the claim it was attached to, with HIGH-WARN classes like claim-not-supported and fabricated-reference able to gate-refuse output through the formatter's terminal hard gate. Style Calibration was used to learn their voice from past writing, and the Writing Quality Check caught machine-sounding prose patterns before submission.

How it went

Produced a structured draft with verified citation anchors and a documented audit trail of which claims passed the integrity gates. The pipeline's stage-by-stage confirmation kept the author in control of the question, method, and interpretation. Polish level depends heavily on the quality of the source material and the user's own review; the claim-audit calibration runner ships with a synthetic gold set and a stub judge, so the tooling is tested but no live-judge calibration result is recorded yet.

Worth knowing

The claim-audit pass is opt-in and its live-judge accuracy is uncalibrated, so treat HIGH-WARN flags as advisory signals rather than proof. Corpus-scale evaluation of ARS itself remains future work. The plugin requires Claude Code v3.7.0+ and the marketplace install flow; the traditional symlink install needs manual prerequisites. The tool explicitly does not help conceal AI use, and the author remains accountable for every claim.

Wiring

Install via the Claude Code plugin marketplace: `/plugin marketplace add Imbad0202/academic-research-skills` then `/plugin install academic-research-skills` (requires Claude Code CLI, VS Code, or JetBrains v3.7.0+). The traditional path uses symlinks into the skills directory. Skills are invoked as slash commands, starting with /ars-plan for Socratic structure planning. The pipeline runs in ordered stages with mandatory human confirmation between them; Stage 2.5 and Stage 4.5 integrity gates execute the 7-mode blocking checklist from academic-pipeline/references/ai_research_failure_modes.md. Citation trust-chain frontmatter carries source provenance, and three-layer locator anchors attach to each citation. Set ARS_CLAIM_AUDIT=1 to enable the opt-in audit pass that fetches each cited source against its anchor and judges claim support; HIGH-WARN classes route through the formatter's terminal hard gate. A calibration runner with a 25-tuple synthetic gold set and FNR<0.15 / FPR<0.10 thresholds ships alongside, currently driven by a stub judge.

Try it yourself

Install the academic-research-skills plugin in this repo's Claude Code environment, then run /ars-plan on my draft paper. Walk me through the structure Socratically, one decision at a time, and stop for my confirmation before moving to the next stage. After the outline is settled, gather and verify references, attach locator anchors to every citation, and run the Stage 2.5 integrity gate. Enable ARS_CLAIM_AUDIT=1 and report any HIGH-WARN findings (claim-not-supported, fabricated-reference, anchorless) with the specific anchor and source so I can adjudicate each one myself. Finally, run the Writing Quality Check against my past writing samples for style calibration and list the passages that read as machine-generated.