findings-workflow

Plugin findings-workflow from marketplace findings-workflow in mriffle/findings-ai-collab-workflow. Commands read from the repo's manifest — 14 days ago.

Skills (7)

• figure-generation

  How to render a publication-ready figure in a Findings Workflow project: read the color registry, apply the publication defaults, dual-export (SVG + 300 DPI PNG), render a separate legend image, and handle the >8-category problem. Use whenever generating a figure for a finding or report.

• report-writing

  How to assemble a report in a Findings Workflow project as a projection of the findings graph: structure, the shared-spec discipline for parallel writers, the per-section write/review loop, and the coherence/assembly pass. Use when running Stage 6 (the stage6-report command) or drafting/assembling r

• research-protein

  How to research a protein or gene for a Findings Workflow project using the authoritative databases (UniProt, PDB/AlphaFold, STRING, GO), with identifier and fact verification. Use when a finding involves a specific protein/gene and needs grounded background. Output feeds a research-finding.

• research-publications

  How to research the scientific literature properly for a Findings Workflow project: find the relevant primary sources, extract specific claims, and cite them with verifiable identifiers (DOI/PMID). Use when researching a question, a paper, or a body of work. Output feeds a research-finding.

• research-source-code

  How to ground a methodological claim in what a tool ACTUALLY computes by reading its source code and documentation, rather than trusting a methods-section description. Use for claims about how software like DIA-NN, limma, MSstats, or a search engine processes data. Output feeds a research-finding.

• statistical-analysis

  How to run a statistically sound analysis in a Findings Workflow project: seed from the vetted lib/ templates, and apply the leakage-safe, multiplicity-honest patterns the conventions require. Use when performing differential abundance, group comparisons, classification, regression, or dimensionalit

• verification-task-builder

  Mechanically construct a blind verification task for the verifier agent from a finding's structured fields, with the answer (evidence and verdict) stripped, and pre-specify the concordance criterion before the verifier runs. Use this whenever preparing to validate a finding (analytic or data replica

Slash commands (11)

• /findings-workflow:README
• /findings-workflow:init

  Scaffold a Findings Workflow project in the current directory — create the standard layout, seed the color registry, initialize the findings graph, and write the project CLAUDE.md.

• /findings-workflow:setup-env

  Ensure this project has a usable, project-local Python environment (>= 3.11). Detects existing Python; if none is suitable, transparently offers to download uv and install Python into this project (zero global footprint), then creates the project .venv, lockfile, and baseline tooling.

• /findings-workflow:stage0-science

  Stage 0 — State the science. Interview the scientist about the study and write state/PROJECT.md, the framing that shapes everything downstream.

• /findings-workflow:stage1-metadata

  Stage 1 — Understand the metadata. Infer and validate the meaning of every column, test the relationships that should hold, detect confounds, and write state/METADATA.md.

• /findings-workflow:stage2-data

  Stage 2 — Understand the data. Determine orientation, transformation/normalization state, and missing-value semantics of the data matrix; characterize missingness, contaminants, decoys, and duplicates; write state/DATA_DESCRIPTION.md.

• /findings-workflow:stage3-loaders

  Stage 3 — Loaders, pairing, QC [INTEGRITY GATE]. Build and verify tested data+metadata loaders that pair every sample exactly, verify the loaded data against the source, produce QC, and pass the integrity gate. Unlocks all analysis.

• /findings-workflow:stage4-explore

  Stage 4 — Explore ⇄ record findings (the heart). Run boilerplate analysis to spark discussion, investigate with the scientist, and capture every substantive insight as a finding the moment it emerges.

• /findings-workflow:stage5-validate

  Stage 5 — Independent validation. Promote a candidate finding toward validated through the blinded independent-validation gate: build the answer-stripped task, dispatch the verifier, check pre-specified concordance, and record the outcome.

• /findings-workflow:stage6-report

  Stage 6 — Reporting. Compile findings into a report that is a projection of the findings graph (not a fresh write-up). Two modes: an exhaustive QC report or a selective research report.

• /findings-workflow:status

  Show where the project stands in the Findings Workflow — pipeline position, integrity-gate state, and a breakdown of findings by status and phase.

Subagents (14)

• README
• code-reviewer

  Independently review a Python analysis script or loader for correctness, reproducibility hygiene, and data-handling bugs before it is promoted. Use before any script moves from scripts/scratch/ to scripts/promoted/ (and thus before a finding may link to it). Checks the artifact — runs the tests, typ

• coder

  Write Python analysis scripts and data loaders for a Findings Workflow project. Use to turn a concrete analysis question into a parameterized, tested, regenerable script. Seeds new scripts from the vetted lib/ templates and reuses the project's existing scripts and shared modules rather than writing

• figure-generator

  Produce a publication-ready figure for a Findings Workflow project from a spec and data. Use to render any figure a finding or report will use. Emits the matplotlib script plus a dual export (SVG vector master + 300 DPI PNG) and a separate legend image, using the project color registry. Figures are

• figure-reviewer

  Independently review a rendered figure for accuracy and publication standards by inspecting the actual PNG, not the code that produced it. Use before any figure is accepted into a finding or report. Catches what code review can't: clipped labels, misleading axes, wrong color mapping, overplotting.

• findings-manager

  Owns the findings graph and manifest for a Findings Workflow project. Use to create a new finding, update an existing one, record a validation outcome, assert a relationship between findings, change a finding's status, or run a consistency/staleness check over the graph. The single writer of finding

• librarian

  Controller of the research subsystem. Use to answer a question from the existing research corpus, to judge whether a research question is already covered, and to decide what new research is needed and with what scope. Knows what research-findings exist and prevents redundant or unverified research f

• report-reviewer

  Independently review report sections (and the assembled report) for claim traceability, faithful status/caveat propagation, and citation integrity, then perform the coherence/editor pass. Use after writers draft sections and before a report is finalized. This is where an LLM most easily overstates —

• research-reviewer

  Independently fact-check a research-finding before it enters the corpus. Use on every research-finding a researcher produces. Verifies factual accuracy and, for EVERY reference, that it exists AND supports the claim attributed to it. Hallucinated or misattributed citations are the failure mode this

• researcher

  Thoroughly research one bounded topic — a protein, gene, disease, pathway, publication, or a piece of software's behavior — and produce a research-finding document with verified, specific references. Dispatched by the librarian with a scoped topic. Context-isolated so research detail never fills ana

• statistician

  Perform a statistical analysis for a Findings Workflow project and return the result plus the analysis script. Use for differential abundance, group comparisons, correlation/association, classification, regression, and dimensionality reduction. Seeds analysis scripts from the vetted lib/ templates a

• stats-reviewer

  Independently review a statistical analysis against the project's statistical conventions. Use before any finding built on an analysis is promoted. Checks for the failure modes that look fine but aren't — bare p-values, missing corrections, data leakage, CV mismatched to the generalization target, m

• verifier

  Blind, independent validator of a finding. Dispatch it with a mechanically derived verification task (the question, comparison, and feature scope) in which the finding's answer — its evidence and verdict — has been stripped. It re-derives the result from scratch in a clean context and reports what i

• writer

  Draft one section of a report by projecting selected findings into prose. Use per logical section during reporting (Stage 6). Invents nothing — every claim traces to a finding (or vetted research-finding) id, and a finding's status and caveats are preserved in the writing. Works from a shared spec s

Hooks (10)

• PreToolUse

  ${CLAUDE_PROJECT_DIR}/.venv/bin/python

• PreToolUse

  ${CLAUDE_PROJECT_DIR}/.venv/Scripts/python.exe

• PreToolUse

  ${CLAUDE_PROJECT_DIR}/.venv/bin/python

• PreToolUse

  ${CLAUDE_PROJECT_DIR}/.venv/Scripts/python.exe

• PreToolUse

  ${CLAUDE_PROJECT_DIR}/.venv/bin/python

• PreToolUse

  ${CLAUDE_PROJECT_DIR}/.venv/Scripts/python.exe

• PreToolUse

  ${CLAUDE_PROJECT_DIR}/.venv/bin/python

• PreToolUse

  ${CLAUDE_PROJECT_DIR}/.venv/Scripts/python.exe

• PostToolUse

  ${CLAUDE_PROJECT_DIR}/.venv/bin/python

• PostToolUse

  ${CLAUDE_PROJECT_DIR}/.venv/Scripts/python.exe