craft-research

plugin 0.3.0

Systematic research with evidentiary provenance. Every claim traces to a source quote, through discourse, extraction, verification, and synthesis.

5 agents6 skills
cix add craft-research

When to Use

  • Scoping research questions and identifying sources
  • Extracting atomic claims from academic papers
  • Verifying citations and claims against sources
  • Synthesizing findings across multiple papers
  • Literature review with traceable provenance
  • Research gap analysis (theoretical, methodological, empirical, practical)
  • Quality audit of research outputs

Pipeline

discourse (elicit) → extract (extract) → verify (scrutiny) → synthesize (synthesis) → audit (audit)

Elicit draws out the inquiry. Agents execute the methodology. Human reviews and decides.

Agents

AgentRolePhase
elicitDraw out research questions, boundaries, sourcesDiscourse
extractExtract atomic claims from sources (Claimify pipeline)Extraction
scrutinyCoVE verification — independently verify claimsVerification
synthesisCross-source integration, convergence/divergence/gapsSynthesis
auditQuality gate — trace provenance chain, ship/returnAudit

Skills

SkillTypeContent
researchHubPipeline, workspace, orchestration, routing
elicitingSpokeResearch discourse, question sharpening, source identification
extractingSpokeClaimify pipeline, source tiers, extraction protocol
verifyingSpokeCoVE protocol, verification verdicts, independence
synthesizingSpokeEvidence weighting, convergence/divergence, gap analysis
auditingSpokeProvenance chain, chain integrity, ship/return

Workspace

.research/
├── scope.md              # Co-created (elicit + human): questions, boundaries
├── PLAN.md               # Orchestrator tracking
├── sources/inventory.md  # Co-created (elicit + human): source list with metadata
├── extraction/           # Per-source claim files
├── verification/         # Per-source verified claims
├── synthesis/            # Per-question findings
└── audit/report.md       # Evaluation report

The Provenance Chain

Source quote (verbatim) → Extracted claim → Verified claim → Finding → Audit trace

Confidence can only decrease through the pipeline — never increase without new evidence.

Key Principles

  • Discourse first — draw out the inquiry before any pipeline work
  • Evidentiary provenance — every claim traceable to a source quote
  • Independence — verification re-reads sources independently (CoVE)
  • Structural confidence — evidence patterns determine confidence, not assertions
  • Named gaps — what's missing is as important as what's found
  • Human scoping — no agent substitutes for defining what questions matter

License

MIT