Academic Research Skills: A five-skill research pipeline

It is not a paper factory; it makes research state explicit

Academic Research Skills (ARS) divides scholarly work into research, writing, integrity checking, simulated peer review, revision, re-review and finalization, then connects those stages through a ten-stage orchestrator. Its central judgment is that AI is strongest at retrieval, formatting, citation checking, consistency checks and repetitive work, while the scholar must still make the defining decisions. Every stage therefore waits for confirmation, and the integrity and review gates cannot be skipped automatically.

That design is more interesting than a prompt that asks for a complete paper because the process control does not disappear inside the model. Stage state is written to a Material Passport, reviewer concerns are linked to revision responses through a traceability matrix, and unresolved issues become acknowledged limitations instead of quietly vanishing in a later pass. Version 3.23.0 still treats its output as working material for the researcher. It does not claim authorship, and simulated review remains author-facing feedback rather than a publication decision.

How the five skills divide the work

Skill	What it handles	Where the boundary sits
deep-research	Eight modes cover Socratic question clarification, quick briefs, literature reviews, systematic reviews and fact checking. Its role set spans research questions, methodology, bibliography, source verification, synthesis, risk of bias and ethics.	Retrieval and synthesis can inform judgment, but they do not replace the scholar's choice of question, method or interpretation.
academic-paper	Eleven modes cover paper planning, outlines, full drafts, abstracts, citation checks, format conversion, revision coaching and reviewer-response work.	The outline must be approved before drafting, and missing material is marked rather than filled with fluent but unsupported text.
academic-paper-reviewer	Journal-fit, domain, methodology, perspective and devil's-advocate reviewers form a multi-perspective review. Re-review, methodology focus, guided review and calibration are also available.	The product is feedback for the author, not an editorial decision. Live reviews remain uncalibrated, so this is not a measured reviewer-accuracy claim.
academic-pipeline	A ten-stage orchestrator manages checkpoints, revision rounds, integrity gates, state handoffs and the final process record.	The orchestrator dispatches work; it does not do the substantive research, writing or review itself. Revision is capped at two full loops, after which unresolved issues must remain visible.
sr-screener	Turns a review protocol into confirmed eligibility rules, screens titles, abstracts and full texts with two isolated reviewers plus an adjudicator, and produces screening logs, PRISMA counts and a handoff corpus.	Protocol comes before screening and piloting comes before scale. Failed calls cannot default to exclusion, and the human team owns the final decisions.

The plugin manifest registers 35 modes and 43 prompt roles. They do not all run as independent runtime agents: four agents are exposed through the plugin layer, while many remaining roles run through in-session prompt flows. The distinction matters because having a written role is not the same as giving every role isolated context, tool permissions and genuinely independent judgment.

Material Passport turns state into a handoff artifact

The Material Passport carries the research question, methodology blueprint, literature corpus, declared experiment provenance, integrity results, revision history and stage handoff materials. It is not a cross-paper memory system and does not retain institutional templates or a long-term author profile; each run reads only the state explicitly supplied to it.

An optional context-reset mode writes a boundary entry at checkpoints so a fresh session can resume from the ledger and its summaries. This addresses context growth and interruption recovery in long workflows. It does not turn model output into a byte-reproducible software build. The repository is explicit that LLM prose and semantic judgments are stochastic: deterministic validators can be replayed, but that is not a reproducibility guarantee for the whole paper.

The two integrity gates are the strongest part of the design

Stage 2.5 checks references, citation context, statistical data, originality and registered claims before the first review. Stage 4.5 repeats the process after revision instead of only rechecking known Stage 2.5 findings. Both gates run a seven-mode AI research failure checklist: implementation bugs that pass self-review, hallucinated citations, hallucinated experimental results, shortcut reliance, implementation bugs reframed as novel insight, methodology fabrication and frame-lock. A suspected mode stops the pipeline so the researcher can confirm, override with recorded reasoning or send the work back for revision.

The deterministic citation gate can establish that a reference resolves across multiple indexes. Existence is not support: a real paper may still fail to support the sentence citing it. Version 3.7.3 added locator anchors and v3.8 added an opt-in claim-to-source alignment audit. That audit is off by default, and the repository records only a synthetic tooling calibration; it publishes no live-judge calibration result. Researchers still need to read the source, not merely accept a pass marker.

Experiment Provenance Intake asks the scholar to declare whether the paper reports experiments they performed. If so, each external experiment records its provenance, planned-versus-executed differences, negative results and known limitations, and the gate checks manuscript claims against that declaration. ARS never runs the experiment and does not judge whether the experiment itself was correct; the check constrains whether the paper overstates the results the scholar reported.

It checks the reported process, not whether an experiment happened

This is the project's most important limitation. ARS can check references, claim-source alignment, reported methodology, figure and table fidelity, reporting coverage and package conformance. Some checks are sampled, some depend on external index coverage, and some remain model-mediated judgments.

It cannot establish that the described procedures were performed, that raw data are authentic or complete, that an analysis reproduces from the underlying materials, or that a real-world intervention occurred as described. A consistently reported fabrication can pass every gate if its citations, claims and package remain internally coherent. The project states that failure shape in its own positioning document and keeps human, institutional and reproducibility review with the researcher.

The install channel determines which controls actually run

The recommended route is the Claude Code plugin. That channel wires slash commands, session-start reminders, the write-scope guard and plugin agents with explicit tool allowlists. Copying skill folders into a project, running inside a repository clone, importing into Claude Science or uploading through Cowork drops different parts of that machinery.

Control	Actual strength	What to know
Stage checkpoints, integrity gates and no-skip rules	Prompt-level controls with audit trails	The session model follows instructions; this is not operating-system or permission-layer enforcement. Overrides require recorded reasoning.
Write-scope guard	Optional hook, conditional	Needs a real Python interpreter, Bash and the plugin channel. Missing dependencies degrade to pass-through rather than locking the user out.
Deterministic citation gate and Python validators	Repository scripts, conditional	They need the plugin install or a repository checkout. Copying the five skill folders alone is not enough.
Dual-blinded systematic-review screening	Requires subagents and screening scripts	The plugin channel is complete. Copy installs need the reviewer agent added separately. Without subagents, only a clearly labelled single-reviewer quick mode remains.

"The repository ships this control" and "this installation runs that control" are different claims. Version 3.23.0 maintains a per-channel availability matrix, which makes the suite more trustworthy than a capability list that ignores installation details.

The licence is not open source, and commercial use needs care

The repository uses CC BY-NC 4.0: source-available with a noncommercial restriction, not an OSI-approved open-source licence. Noncommercial academic research, teaching, method training and shared workflows in noncommercial research groups are allowed. Commercial SaaS, paid consulting, paid enterprise deployment, commercial API wrappers and resale require separate licensing.

The installation path is:

/plugin marketplace add Imbad0202/academic-research-skills
/plugin install academic-research-skills

For a scholar who already has code, data and reviewer comments, the strongest use is not generating a paper from a blank page. It is entering at the relevant stage, running the integrity check first, and then letting review, revision and re-review operate on traceable state. People who need cross-paper memory, actual experiment execution or a fully autonomous research pipeline will find that this tool deliberately stops at the boundary.