Skip to main content

Rigorous methods, from question to paper.

Give your coding agent the habits of a careful methods team. Plan the analysis, review the methods, and trace the results in R or Python. You keep the scientific decisions.

A research plugin for Claude Code and Codex, with workflows for clinical statistics, machine learning and meta-analysis.

Set up MethodsmithSee an example

Open source. MIT licensed. Your team stays accountable.

Start in Claude Code
# install the plugin
claude plugin install methodsmith --marketplace sajor2000/methodsmith

# in your project, inside Claude Code
> /methodsmith:method-setup
> /methodsmith:method-workflow

# then plan before any analysis runs
> /methodsmith:method-plan
  a cohort study of 30-day mortality

Requires Claude Code 2.1.292 or later. Other agents and managed laptops.

67 skills, 55 specialist agents, 49 reporting standards.

Browse the full toolkit

Watch the 90-second tour​

A researcher's questions, one scene each: which reporting guideline applies, where to start the literature, how to sharpen the question, the analysis plan, data checks in R or Python, figures and tables, and the first Results draft. You make every call.

Silent, scripted illustration on synthetic data. Screens are mock-ups of the formats the skills define, not a recording of a live run.

You decide. Methodsmith does the legwork.​

Your team frames the question, approves the plan and judges the findings. Methodsmith drafts, runs and checks the work around those decisions. The plan, work, review and compound loop comes from Every's Compound Engineering.

Swimlane. You, the scientist, frame the question, approve the plan, judge the findings and sign off the result. Methodsmith drafts the SAP with /method-plan, runs the approved analysis with /method-work, reviews every step with /method-review, and records the lessons with /method-compound, which feed the next study's plan.Swimlane. You, the scientist, frame the question, approve the plan, judge the findings and sign off the result. Methodsmith drafts the SAP with /method-plan, runs the approved analysis with /method-work, reviews every step with /method-review, and records the lessons with /method-compound, which feed the next study's plan.

  1. Plan

    Your teamFrame the question and approve the analysis plan.

    MethodsmithDraft the SAP, assumptions and planned outputs.

  2. Work

    Your teamResolve decisions and authorize the analysis.

    MethodsmithExecute approved tasks and track SAP outputs.

  3. Review

    Your teamJudge findings and decide what needs to change.

    MethodsmithCheck methods, code, reporting and reproducibility.

  4. Compound

    Your teamChoose which validated lessons to retain.

    MethodsmithDocument decisions for the next study.

Validated lessons feed the next plan. Every study compounds.

See it on synthetic data​

Find the next step, link a causal claim to its estimand, then review the analysis before anyone reads the results.

Illustrative terminal output: workflow recommends method-sap; the SAP links a claim to an estimand and flags two unresolved decisions; review returns Not ready.
Scripted replay on synthetic data. Output abridged from the formats the skills define.

In this example, the plan has two unresolved decisions. Review returns Not ready, with findings for the researcher to resolve, not an automatic sign-off.

Built for three kinds of study​

67 skills in total. These are the ones each kind of study reaches for first.

Clinical statistics

Observational studies, trials and registries, from question to Table 1.

  • /method-questionPICO, FINER and a hypothesis
  • /method-powersample size and power, with a script
  • /method-sapclaims linked to ICH E9(R1) estimands
  • /method-cohortreproducible cohort and waterfall counts
  • /method-table1traceable baseline-characteristics table
  • /method-numbersevery reported number traced to the analysis

Python and ML papers

Prediction models and notebooks that survive peer review.

  • /method-model-choicejustify the model after the estimand
  • /method-ml-trackdata hashes, splits and metrics
  • /method-verifyleakage, effect direction and PHI gate
  • /method-model-cardcalibration, fairness and limits
  • /method-marimoreactive, reproducible Python notebooks
  • /method-figurecalibration, ROC and forest plots

Meta-analysis

Systematic reviews and pooled effects with every source cited.

  • /method-lit-searchPubMed, full text or authoritative sources
  • /method-evidence-mapa source-backed evidence map
  • /method-preregPROSPERO, OSF or ClinicalTrials.gov
  • /method-meta-analysisrandom effects, REML with Hartung-Knapp
  • /method-checklistPRISMA 2020 and its extensions
  • /method-writeevidence-bound manuscript drafts

Why you can check its work​

Methodsmith drafts and checks. The researcher decides, and stays accountable for every result.

Research use only

Methodsmith is not a medical device for clinical care and is not intended to diagnose, treat, or guide decisions about individual patients. Research outputs require qualified human review.

Read the research-use scope
49 reporting standards

Reporting guidelines, appraisal tools and templates, including CONSORT, STROBE, PRISMA, TRIPOD+AI and TARGET, encoded with their official item IDs. A test enforces routing to the newest version of each guideline.

Check the guideline sources
An explicit sensitive-data boundary

Claude Code’s pre-tool hook restricts data reads in schema-only mode and checks identifiers in supported WebSearch, WebFetch, MCP, and shell egress inputs in every mode. Other agents rely on skill instructions. These safeguards are not a sandbox or a substitute for institutional approval.

Read the safeguards and their limits
55 specialist agents

Methods, causal inference, leakage, calibration, fairness and reproducibility reviewers are part of a broader team of research and workflow agents. Review findings are deduplicated and confidence-gated.

Explore the skills and agents
Numbers traced to the analysis

Manuscripts, tables and abstracts cite a registry written by the analysis code. A fail-closed gate traces each number to its SAP output, script commit and source hash. It does not read numbers inside images.

See how number provenance works

The skill map​

Thirty of the 67 skills, in the order a study meets them, with the guardrails that apply at every step.

View the study lifecycle map
Skill map across Frame, Design, Data, Plan, Execute and Review. The skill lists above provide text descriptions; the diagram gives an overview, not proof of enforcement.

Install once, set up each project​

Start with Claude Code or Codex. The setup guide also covers Pi, OpenCode, Gemini CLI, Kiro, Qwen Code and approved offline installs. Then open your study folder and run setup for that project.

Choose your install routeView the source on GitHub
View the installation workflow
Install the plugin, run method-setup in the project, then use method-workflow to choose the next skill.

Credits​

Methodsmith is a fork of Compound Engineering by Kieran Klaassen and Every, which introduced the plan, work, review, compound loop for software. Methodsmith adapts it for research methods and is maintained by Juan Carlos Rojas. MIT licensed.