maxwell2732/claudecode-research-harness-workflow/research-harness-setup

Research Harness Setup

Initialize a new empirical research project with source-of-truth files, a controlled folder structure, and data protection rules.

This skill runs once at the start of a project. It does not run analysis, audit data, or claim feasibility.

Quick Reference

Procedure

Step 1 — Collect study context

If arguments are not provided, ask the user for:

1. Study name or working title
2. Path to raw data file(s) (may be unknown if data not yet available)
3. Brief description of the research question (one or two sentences; may be unknown)

Do not invent answers. If the user says unknown, write unknown in the spec.

Step 2 — Create folder structure

Create the following directories if they do not exist. Do not delete or overwrite existing content.

data/raw/
data/intermediate/
data/processed/
scripts/
analysis/
output/
logs/
reports/

Step 3 — Create study_spec.md

Copy templates/study_spec.md to study_spec.md at the project root.

Fill in:

• Study name from Step 1
• Raw data path from Step 1
• Research question from Step 1

Leave all other fields as unknown. The researcher fills in the full spec before Stage 2.

If study_spec.md already exists: read it, confirm with the user before overwriting, and preserve existing content unless the user explicitly requests replacement.

Step 4 — Create analysis_plan.md

Copy templates/analysis_plan.md to analysis_plan.md at the project root.

Fill in:

• Study spec path: study_spec.md
• Date: today's date
• Analyst: Claude Code

Leave all task fields as cc:todo with unknown paths.

If analysis_plan.md already exists: do not overwrite. Notify the user.

Step 5 — Write data protection notice

Append the following block to the project .gitignore if it does not already contain it:

# Research Agent Harness — raw data is read-only
# Never commit cleaned or processed datasets unless explicitly approved
data/processed/
data/intermediate/
logs/

Write a data/raw/READONLY.md file with the following content:

# data/raw — READ-ONLY
 
This folder contains raw, unmodified data files.
 
No script, agent, or tool may write, edit, rename, or delete files in this folder.
All data transformations must produce new files in data/processed/ or data/intermediate/.
 
See docs/INTEGRITY-RULES.md Rule 1.

Step 6 — Confirm and report

Print a setup summary:

Research Agent Harness — Setup Complete
 
Study spec:     study_spec.md         [created / already existed]
Analysis plan:  analysis_plan.md      [created / already existed]
Folders:        data/raw, data/intermediate, data/processed,
                scripts, analysis, output, logs, reports
Data protection: data/raw/READONLY.md written
 
Next step: Fill in study_spec.md, then run /research-harness-audit

Forbidden Actions

• Do not modify any file under data/raw/
• Do not invent a research question if the user did not provide one
• Do not claim the data is feasible for the proposed design before an audit is run
• Do not run any analysis or data inspection in this step
• Do not overwrite study_spec.md without user confirmation if it already exists

Completion Criteria

• [ ] study_spec.md exists with study name, data path, and research question filled in (or unknown)
• [ ] analysis_plan.md exists with header fields populated
• [ ] All 8 folders exist
• [ ] data/raw/READONLY.md exists
• [ ] No files under data/raw/ were modified

Handoff to Stage 2

Tell the user:

Fill in the remaining fields in study_spec.md — particularly identification strategy, variables, and sample restrictions — before running /research-harness-audit.

When study_spec.md is complete and you have approved it, run /research-harness-audit.