research-pipeline

Produce source-grounded research documents with content-addressed evidence and provenance manifests.

2|Updated May 4, 2026
One-click install
npx skills add https://github.com/sfrangulov/skills --skill research-pipeline-sfrangulov
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: research-pipeline
Source: https://github.com/sfrangulov/skills/tree/main/research-pipeline
Command: npx skills add https://github.com/sfrangulov/skills --skill research-pipeline-sfrangulov

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes scripts (resource) and references (resource) components.

What problem does it solve?

This Skill helps you produce research documents whose claims are grounded in sources and remain verifiable across sessions and machines, rather than being “memory of one” or uncheckable summaries.

Core Features & Use Cases

  • Snapshot-backed claims: canonized claims point to content-addressed snapshots (not bare URLs), with a provenance manifest.
  • Deterministic verification gates: deterministic checks run first, with an LLM-judge only on the small remainder when needed.
  • Mandatory adversarial pass: a separate refuting step that classifies load-bearing claims as survived, weakened, or refuted before canonization.
  • Stop-hook enforcement: gate checks can be run automatically at session end so omissions don’t slip through.
  • Explicit non-verbating tiers: avoids using paraphrase-style fetch methods as a verbatim source of record.

Quick Start

Tell your agent: “Run the research pipeline for my topic, store snapshots in the configured cache directory, and end by producing a document with a provenance-manifest block and an adversarial epistemic-status tag for load-bearing claims.”

Frequently Asked Questions about research-pipeline

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I ensure reproducibility in multi-source research documents?

Reproducibility in multi-source research is achieved by generating source-grounded documents with content-addressed snapshots and a provenance manifest. This allows others to re-check findings later by fetching the exact evidence rather than relying on bare URLs or single-session memory.

What is adversarial verification for research claims?

Adversarial verification is a mandatory refutation pass that classifies load-bearing claims as survived, weakened, or refuted before canonization. This ensures only vetted claims are added to your final research artifact.

How do I cite sources verbatim without relying on bare URLs?

You cite sources using content-addressed snapshots instead of bare URLs. A provenance manifest records the verbatim source of record, explicitly avoiding paraphrase-style fetch methods to maintain evidence integrity.

Can I automate verification gates at the end of a research session?

Yes, you can automate verification gates using stop-hook enforcement. This runs deterministic checks automatically at session end, ensuring omissions do not slip through and maintaining reproducibility checks.

Does deterministic verification require an LLM-judge for every claim?

No, deterministic verification gates run first without an LLM-judge. An LLM-judge is only applied to the small remainder of claims when deterministic checks are insufficient.

When do I need content-addressed snapshots for technology evaluations?

You need content-addressed snapshots for technology evaluations when others must be able to re-check findings later. This process canonizes claims with verifiable fetch-contract requirements and provenance manifests.