-->

Abstract image of a research assembly line A staged pipeline flowing from research to manuscript, passing through gates at each transition.

Overview

The performance of a coding agent is increasingly determined not by the model itself but by the skills and workflows layered on top. The same model becomes an entirely different specialist depending on which skill pack is loaded. In June 2026, a project that pushed this idea to its limit in academic research gained attention. It connects 10 stages automatically, from literature search to a submission-ready PDF manuscript, with agents playing the roles of supervisor, three peer reviewers, and copyeditor.

ThakiCloud runs a Kubernetes-based AI/ML SaaS platform serving agent workloads for multiple customers simultaneously. The design principle of “multiple agents collaborating to produce verifiable output” is not someone else’s problem for us. This post is a record of actually integrating that skill pack, Academic Research Skills, into our agent environment and examining its architecture. Rather than a promotional overview, we look at what is inside, how much of it can be trusted, and what is genuinely useful from our platform perspective.

What This Skill Pack Is

Academic Research Skills is a noncommercial skill collection for Claude Code that packages research, writing, verification, review, editing, and citation-integrity workflows into step-by-step skills, agents, and gates. The core is not a single prompt but the collaboration of multiple specialized agents and mandatory verification gates between each stage. It is organized into four major skills.

  • deep-research (version 2.9.4): A 13-agent research team that handles any topic. It supports 7 modes: full research, quick brief, paper review, literature review, fact-checking, Socratic guided dialogue, and systematic review with meta-analysis. Roles are divided across research-question formulation, methodology design, systematic literature search, source verification, cross-synthesis, bias-risk assessment, APA 7.0 report writing, editorial review, and devil’s advocate counterargument.
  • academic-paper: A writing pipeline that produces actual paper output from research findings. It accepts a style profile, applies it, and runs a writing-quality check immediately before drafting, flagging AI-typical overused phrases, monotonous sentence length, and padding in introductions.
  • academic-paper-reviewer: Serves as the reviewer who assesses a completed manuscript.
  • academic-pipeline (version 3.11.1): The orchestrator that ties all three together. It does no work itself; it only handles stage detection, mode recommendation, skill dispatching, and status tracking. A thin conductor.

The 10 stages the orchestrator connects are as follows.

[1] RESEARCH
      |
[2] WRITE
      |
[2.5] INTEGRITY <- 100% citation and data verification gate
      |
[3] REVIEW (first peer review)
      |
[4] REVISE
      |
[4.5] RE-REVIEW (second focused review)
      |
[5] RE-REVISE
      |
[5.5] FINAL INTEGRITY (re-verification)
      |
[6] FINALIZE (PDF with provenance record)

Two design decisions here are particularly interesting from a ThakiCloud perspective. First, both core skills explicitly declare data_access_level: verified_only in their metadata. Second, there is a mandatory verification gate between each stage. Claims about the literature are not generated from the model’s memory; they are checked against an actual paper database via the Semantic Scholar API. Instead of producing plausible-sounding citations, citations that do not exist are blocked at the gate.

Installation and Integration

Installation was done by cloning the external repository and linking it into our .claude/skills/ directory via symbolic links. Rather than maintaining a separate copy, we follow the source so upstream updates are reflected automatically.

# Clone the external repo (home directory)
git clone https://github.com/Imbad0202/academic-research-skills.git ~/academic-research-skills

# Link into our agent skills directory
cd .claude/skills
ln -s ~/academic-research-skills/deep-research            deep-research
ln -s ~/academic-research-skills/academic-paper           academic-paper
ln -s ~/academic-research-skills/academic-paper-reviewer  academic-paper-reviewer
ln -s ~/academic-research-skills/academic-pipeline        academic-pipeline

Checking which skills are connected after linking shows them resolved as symlinks:

academic-paper          -> ~/academic-research-skills/academic-paper
academic-paper-reviewer -> ~/academic-research-skills/academic-paper-reviewer
academic-pipeline       -> ~/academic-research-skills/academic-pipeline
deep-research           -> ~/academic-research-skills/deep-research

Internally at ThakiCloud, we wired these four skills into our research and report pipeline (the jarvis family) so that paper-oriented inputs invoke deep-research in paper-review or literature-review mode. The “full review document (DOCX)” output attached to the paper-category posts on our blog also sits on the same philosophy of verification-first review.

What Is Actually Inside

Opening the orchestrator’s agent configuration directly reveals not a simple bundle of prompts but agents with clearly separated roles.

academic-pipeline/agents/
+-- pipeline_orchestrator_agent.md      # stage switching and dispatch
+-- state_tracker_agent.md              # progress state and checkpoint tracking
+-- integrity_verification_agent.md     # citation and data integrity verification
+-- collaboration_depth_agent.md        # records depth of human-AI collaboration
+-- claim_ref_alignment_audit_agent.md  # claim-citation alignment audit (L3)

Several safeguards stand out.

  • Integrity verification stages (2.5, 5.5): Citations and data are verified 100% twice, once after writing before review submission, and once after revision is complete. Failing this gate blocks progression to the next stage.
  • L3 claim-citation alignment audit: Enabling the ARS_CLAIM_AUDIT=1 flag adds a claim-fidelity audit gate at the transition from stage 4 to stage 5. It collects and classifies unsupported assertions, claim drift, and constraint violations, and if the risk level is high enough the formatter refuses to produce output. Off by default, designed to be enabled gradually as calibration data accumulates.
  • Material Passport: Enabling ARS_PASSPORT_RESET=1 promotes stage checkpoints to context-reset boundaries. In a new session, you can resume from a recorded stage using resume_from_passport=<hash>. This is a practical mechanism for cutting long tasks at stage boundaries to avoid bloated context windows.
  • Provenance record PDF: After the pipeline completes, it automatically generates a document recording how humans and AI collaborated. This makes it possible to trace “how far AI was involved” from an academic integrity perspective.

The README is provided in English, Japanese, Simplified Chinese, and Traditional Chinese, showing broad consideration for multilingual users. The license is noncommercial, so if you are evaluating commercial adoption, review the license terms first.

Application to ThakiCloud K8s AI/ML SaaS Platform and Implications

The aspect that resonated most from integrating this skill pack is that it overlaps precisely with two principles we have emphasized in our platform design.

First, verified-only and hallucination blocking. ThakiCloud serves customers with significant on-premises and self-hosting requirements. In such environments, the most feared failure mode is “plausible but wrong output.” The way Academic Research Skills cross-checks citations against an actual database rather than the model’s memory reflects the same mindset as our internal rule of “only cite URLs you actually retrieved, mark unverified figures as estimates.” The structure of code enforcing gates while models only generate content also resembles our principle of delegating format and verification in batch outputs to deterministic code.

Second, the thin conductor and thick skills, with loops closed by verification. academic-pipeline does no work itself; it handles only stage detection, dispatch, and status tracking. Real capability is packed into each specialized skill. Inserting an integrity verification stage twice before merging multi-agent fan-out results is the same conclusion as our operational lesson that hallucinations accumulate when parallel agent results are merged without a verification gate.

From a platform product perspective, this kind of research pipeline is well-suited to run as a multi-tenant workload on our Kubernetes infrastructure. In an environment where Kueue manages the GPU queue and vLLM serves models, spinning up isolated research agent teams per customer to produce verified reports limited to internal documents and internal data is a natural scenario. For public institutions and research organizations that cannot upload sensitive research materials to an external SaaS, the ability to run an agent pipeline with built-in verification gates inside their own infrastructure becomes straightforward differentiation.

Limitations and Counterarguments

Good design, but not something to accept uncritically. There are several clear limitations.

  • Verification gates have a cost. Running integrity verification twice across 10 stages, plus an optional claim audit, adds substantial tokens and time to each output. For cases where a quick draft is needed, this is overkill. That is why a quick-brief mode exists separately, and forcing the full pipeline on every task is not wise.
  • Semantic Scholar verification is not comprehensive. Recent preprints not yet indexed in the database, non-English literature, and grey literature may fall outside verification scope. “Verified” does not guarantee 100% accuracy.
  • The noncommercial license is fine for internal experimentation but requires legal review before embedding in a product. If commercial redistribution is the goal, it cannot be used as-is.
  • Agent count does not scale linearly with quality. Numbers like 13 or 12 agents are impressive, but the key question is whether each agent genuinely adds verification from a distinct perspective. Running the same prompt multiple times is not verification. This skill pack does separate roles clearly, which is good, but adopters should evaluate by “diversity of verification” rather than “number of agents.”

In summary, Academic Research Skills is a solid reference for taking the idea of “producing verifiable academic output with multiple agents” to its logical conclusion. What matters to ThakiCloud is not the tool itself but the design principles: enforcement of verification through code gates, a thin conductor coordinating thick skills, and closing multi-agent results with verification. These principles apply not only to academic papers but directly to the multi-tenant agent platform we operate every day.

Sources

참고

추론 엔진을 고르지 마세요, 하드웨어 전략을 고르세요

 

vLLM이냐 SGLang이냐 TensorRT-LLM이냐를 먼저 고민하는 것은 순서가 틀렸습니다. 한 엔지니어가 제시한 하드웨어 우선 선택법은, 가진 하드웨어와 워크로드 모양이 엔진을 결정하게 두라고 말합니다. 이 프레임을 ThakiCloud의 쿠버네티스 GPU 서빙 운영 관점에서 ...

Hermes Bible: Hermes Agent 문서와 실전 워크플로를 한 번에 검색합니다

 

Nous Research의 Hermes Agent 공식 문서 169페이지와 커뮤니티가 만든 실전 워크플로 28개를 한곳에 색인해 ⌘K 한 번으로 검색하는 비공식 커뮤니티 사이트입니다. 무엇을 담고 있고, 공식 문서와 어떻게 다른지, 그리고 1000개가 넘는 스킬과 룰을 운용하는 T...

Fable 5가 내 코드의 100%를 쓴다는 말, 플랫폼 관점에서 따져봅니다

 

Claude Code를 만든 사람이 ‘Fable 5가 내 코드의 100%를 쓰고 있고 Opus보다 최소 3배 강력하다’고 말했다는 인용이 돌았습니다. 검증되지 않은 수치는 인용으로만 두고, 공식적으로 확인되는 사실(가격, 컨텍스트, 벤치마크)과 에이전트형 코딩이 실제로 무엇을 바꾸...