Lineage

April 23, 2026 · View on GitHub

This bundle stands on four other people's work. Naming them explicitly, because the design is deliberate borrowing.

K-Dense scientific-agent-skills

What we take: The skill library itself. Ten of the 134 skills — hypothesis-generation, scientific-critical-thinking, statistical-analysis, peer-review, scholar-evaluation, scientific-writing, scientific-schematics, citation-management, venue-templates, scientific-brainstorming — cover the full paper lifecycle from question formulation to venue formatting. The bundle does not fork or vendor these; it references them upstream.

What we add: Orchestration. K-Dense's 134 skills are narrow and unopinionated by design — that's the right call for a skill library. A working researcher still has to know which skills to compose, in which order, with which defaults. The bundle does that composition.

Credit: K-Dense Inc. License of upstream skills applies.

Denario

What we take: The multi-agent topology. Denario's modules (idea → literature → plan → execute → paper → review) map almost 1:1 to our modes, because the underlying lifecycle is the same. The insight that a research system benefits from specialized agents per phase — rather than one generalist agent — is Denario's and is correct.

What we do differently: Denario targets autonomous research where AI is the primary author. This bundle explicitly keeps the human as primary author. Denario produces a paper from a problem statement; this bundle produces a paper with the human making the substantive calls at each mode transition.

Credit: Pablo Villanueva-Domingo, Francisco Villaescusa-Navarro, Boris Bolliet. GPL-3 / Apache 2 as applicable to any code we borrow (none yet — we borrow design only).

Superpowers + amplifier-bundle-superpowers

What we take: The process-discipline pattern. Superpowers enforces /brainstorm → /write-plan → /execute-plan → /verify → /finish for software development. This bundle applies the same pattern — design before execution, lock the plan, verify against the plan, report honestly — to scientific research. The structure transfers cleanly because the underlying virtue is the same: disciplined workflow beats improvisation.

What we do differently: Different domain, different agents, different verification criteria. Software verification means tests pass; research verification means the analysis matches the pre-registration.

Credit: Jesse Vincent (Superpowers methodology). Microsoft Amplifier team (bundle packaging). The inspiration for this bundle's existence, especially the mode-driven UX.

amplifier-bundle-stories

What we take: The UX bar. Install is one command. First run requires zero configuration. Recipes handle the 80% case out of the box. A non-expert can produce an artifact in their first session without reading docs.

What we match: File layout, bundle.md conventions, README structure, recipe YAML shape.

Credit: Michael Jabbour (author). This bundle targets the same user ease as stories; if we land meaningfully above or below that bar, it's a bug.

amplifier-bundle-scientificpaper (Michael Jabbour, 2025)

What we take: Craft material. The earlier bundle is a substantially more implemented piece of work on the narrow problem of academic-paper production. We absorb its content:

  • Six conference format specifications (NeurIPS, ICML, ACL, IEEE, ACM, arXiv) — now in references/conference-formats/
  • Matplotlib-scientific and TikZ/PGFPlots reference material — now in references/imaging/
  • PaperBanana figure-generation methodology (arXiv 2601.23265) — now in references/figure-generation/
  • Python scripts for deterministic operations (LaTeX compilation, format validation) — now in scripts/
  • Writing-craft content from the paper-architect agent: 5-component abstract framework, section flow principles, active-voice and precision-of-claim patterns — absorbed into our technical-writer agent
  • The approval_required / approval_prompt pattern from their paperbanana-figure.yaml recipe — adopted for our recipe schema at mode boundaries

What we do differently:

  • Broader scope. Their bundle targets academic-paper production; ours covers patent briefs, policy briefs, white papers, replication studies, grants, and more alongside papers.
  • Three-persona framing (non-scientist / researcher / reviewer) rather than an implicit researcher audience.
  • Mode-based UX (/question, /plan, /execute, /critique, /draft, /publish) rather than keyword-triggered agent routing.
  • Pre-registration discipline (hash-locked preregistration, honest-pivot behavior, exploratory-labeling) — core differentiator not present in the earlier bundle.
  • Methodology critique via GRADE / Cochrane ROB / CONSORT / STROBE / PRISMA — the earlier bundle's critique is structural-only.

Credit: Michael Jabbour (author). The earlier bundle is the direct ancestor of this one. The content in references/ was written for that bundle and is reused here under the same license. See docs/GAP_ANALYSIS.md for the full merge analysis.

PaperBanana acknowledgment: The figure-generation methodology cited in references/figure-generation/paperbanana-methodology.md is original research by Dawei Zhu, Rui Meng, Yale Song, Xiyu Wei, Sujian Li, Tomas Pfister, and Jinsung Yoon (Peking University / Google Cloud AI Research), published at arXiv 2601.23265. Their work is credited as the source of the 8 quality veto rules and the 5-agent multi-stage refinement architecture.

Amplifier itself

What we rely on: The runtime, module system, session management, provider routing, agent-delegation mechanism, and bundle loader. This bundle is pure configuration on top of Amplifier; it ships no runtime modules of its own.

Credit: Microsoft Amplifier team. The bundle would not exist without the host.

The broader ecosystem

Papers, tools, and ideas that shaped the thinking:

  • The AI Scientist / AI Scientist-v2 (Sakana, UBC, Oxford) — published in Nature, 2026. Demonstrated that end-to-end autonomous paper generation is possible. Sharpened our thinking about what not to automate: the substantive judgment calls.
  • CONSORT, STROBE, PRISMA reporting guidelines — the baseline for methodology transparency. honest-critic references these directly.
  • GRADE, Cochrane Risk of Bias — evidence-quality frameworks, embedded in methodologist and honest-critic.
  • OSF pre-registration standards — the model for our pre-registration artifact.
  • Reflection Tokens, ROI on Inference, and the RCE benchmark — in-progress research programs by the author whose methodology discipline this bundle codifies and generalizes; the RCE benchmark is this bundle's planned primary evaluation target.

What's original here

The bundle's contribution is not any single layer. It's the composition:

  • A specific ten-skill subset (not all 134)
  • Mapped to eight agents (not one, not one-per-skill)
  • Behind six modes (not the raw agent list)
  • Composed into seven recipes (not an open-ended recipe space)
  • With two honesty-enforcing behaviors (novel — not present in Denario or Superpowers)
  • Packaged for one-command install (matching stories)
  • Sized for non-expert users as primary audience (distinct from Denario and Superpowers)

Everything else is borrowed, credited, and better for it.