insight-blueprint

A Claude Code skills plugin for hypothesis-driven EDA. Analysis designs, journals, reviews, and the data catalog live as YAML under .insight/; skills read and write them directly. Integrity is enforced by an embedded validation library (validate.py) plus a pre-write hook — no server, no daemon, no SQLite. See docs/ARCHITECTURE.md and docs/PRD.md.

Quickstart

Prerequisites: Claude Code, uv, and Python 3.11+. The skills shell out to uv run …, which resolves the Python dependencies automatically — there is nothing to pip install.

1. Install the plugin. In Claude Code, register this repo as a marketplace and install:

   /plugin marketplace add etoyama/insight-blueprint-skills
   /plugin install insight-blueprint@insight-blueprint-marketplace
   

2. Open your analysis project in Claude Code and take the first step:

   /analysis-framing        # explore your data & frame a direction
   /analysis-design         # turn it into a hypothesis design document
   /analysis-journal        # record reasoning as you investigate
   /analysis-reflection     # conclude (or branch) the hypothesis
   /analysis-report         # write up a distributable APA-style report
   /knowledge-extract       # save reusable, source-scoped knowledge
   

There is nothing to launch — skills read and write YAML under .insight/ in your project (created on first use). Writes to design documents are validated automatically by a pre-write hook.

Installation

Claude Code plugin (recommended)

In Claude Code, add this repository as a plugin marketplace and install from it:

/plugin marketplace add etoyama/insight-blueprint-skills
/plugin install insight-blueprint@insight-blueprint-marketplace

Here insight-blueprint is the plugin name and insight-blueprint-marketplace is the marketplace defined in .claude-plugin/marketplace.json.

Local clone (for development)

git clone https://github.com/etoyama/insight-blueprint-skills.git
claude --plugin-dir ./insight-blueprint-skills

--plugin-dir loads the plugin for the current session only. Add a shell alias to make it stick:

alias claude-ib='claude --plugin-dir /path/to/insight-blueprint-skills'

Optional: Python package (only for hand-written lineage)

You do not need to install anything for the skills to work. The plugin is self-contained: its bundled uv environment provides insight_blueprint (models, validation, lineage) and /analysis-notebook runs marimo/pandas through that same environment (--extra notebook). See ADR-0006.

Install the insight-blueprint-lineage package only if you want to import tracked_pipe directly in your own notebooks/scripts (i.e. lineage tracking outside /analysis-notebook). The import name stays insight_blueprint:

uv add insight-blueprint-lineage

See CHANGELOG.md for release notes.

How it works

• Skills (skills/) drive the workflow and read/write .insight/ YAML via small server-free helpers (skills/_shared/design_io.py, catalog_io.py).
• Validation is centralized in src/insight_blueprint/validate.py (Pydantic schema + state-transition guard). A pre-write hook (hooks/validate-design.py, shipped with the plugin via hooks/hooks.json) calls the same library to block invalid writes to *_hypothesis.yaml.
• Lineage (src/insight_blueprint/lineage/) records DataFrame transformations and exports Mermaid diagrams.

Skills

• /rq-problematization — Generate impactful research questions by problematizing prior-research assumptions (upstream of framing)
• /analysis-framing — Explore available data and existing analyses to frame a direction
• /analysis-design — Guided creation of hypothesis design documents
• /analysis-review — Produce a structured review of a design and record it as a review batch
• /analysis-notebook — Generate a marimo notebook from the design's methodology, run it, and record results to the journal
• /analysis-journal — Record reasoning steps during analysis (observations, evidence, decisions, questions)
• /analysis-reflection — Structured reflection to draw conclusions or branch hypotheses
• /analysis-report — Assemble a distributable APA-style Markdown report from a concluded analysis (read-only consumer)
• /analysis-revision — Guided revision workflow for addressing review comments
• /catalog-register — Step-by-step data source registration
• /knowledge-extract — Extract reusable, source-scoped domain knowledge from a concluded analysis
• /data-lineage — Track data transformations and export lineage diagrams (Mermaid)
• /premortem — Report-only pre-flight cost/risk evaluation of designs before expensive data access
• /analysis-auto — Guided autopilot: drives the pipeline, pausing only at genuine decisions

Skills support both English and Japanese trigger phrases.

Analysis Workflow

/rq-problematization (problematize assumptions → research questions)  ← optional upstream
    ↓ (RQ Brief)
/analysis-framing (explore data, frame direction)
    ↓
/analysis-design (create hypothesis)
    ↓ ↘ /analysis-review (review the design) → /analysis-revision (address comments)
    ↓ ↘ /premortem (optional: cost/risk report before expensive data access)
/analysis-notebook (generate & run a marimo notebook from the methodology → record results)
    ↓ ↘ /data-lineage (optional: track transformations, export Mermaid)
/analysis-journal (record reasoning during analysis)
    ↓
/analysis-reflection (reflect → conclude or branch)
    ↓ ↘ /analysis-report (concluded → distributable APA-style report)
    ↓ ↗ back to /analysis-framing (new direction needed)
/knowledge-extract (save reusable, source-scoped knowledge)

Skills are invoked explicitly (/command) by default and the flow is interactive. The actual analysis is run by /analysis-notebook, which generates a marimo notebook from the design's methodology, executes it, and records the results to the journal (uv add "insight-blueprint-lineage[notebook]" for the runtime deps).

Guided autopilot. /analysis-auto drives the whole pipeline for you — auto-advancing the low-friction steps and pausing only at genuine decisions: confirming the hypothesis, registering a data source, a HARD_BLOCK/HIGH premortem, a notebook that would need out-of-allowlist packages or external communication beyond the declared source, the conclusion, and — after a terminal conclusion — offering the distributable report. It is opt-in and still interactive — not an unattended pipeline. The individual skills stay explicit everywhere else. See ADR-0005.

/catalog-register sits upstream (register a data source before you frame against it); /knowledge-extract sits downstream (harvest what a concluded analysis taught you about that source).

Each design has an analysis_intent field (exploratory, confirmatory, or mixed). The Insight Journal (.insight/designs/{id}_journal.yaml) tracks your reasoning with event types mapped to the Narrative Scaffolding framework (Huang+ IUI 2026).

Designs, status & review

There is no web UI — everything is skill-driven over .insight/ YAML:

• Status transitions are performed by the skills, not by hand. /analysis-reflection proposes the terminal transition (e.g. analyzing → supported) and runs it via design_io transition; the pre-write hook enforces the allowed transitions.
• Review is a producer/consumer pair (no dashboard): /analysis-review critiques a design and records a review batch under .insight/designs/{id}_reviews.yaml (via design_io review-batch), setting the design to revision_requested or analyzing. /analysis-revision then consumes a revision_requested batch, walking you through each comment — tracking per-comment progress in .insight/designs/{id}_revision.yaml — and re-submitting for review.

Capturing knowledge (/knowledge-extract)

When an analysis concludes, /knowledge-extract reads its journal/reflection/review and proposes source-scoped knowledge entries (a caution, a definition, a methodology note), which — after you confirm — are persisted to the catalog:

echo '{"entries": [
  {"key": "pop-null-pre-2019", "title": "population null before 2019",
   "content": "population is null before 2019; start time series at 2019.",
   "category": "caution", "importance": "high", "affects_columns": ["population"]}
]}' | catalog_io add-knowledge --id <source_id>

category is an open string (conventional values: methodology / caution / definition / context). Analytical conclusions stay in the reflection, not the catalog.

Pre-flight Risk Evaluation (/premortem)

/premortem evaluates designs for cost/risk before expensive data access and prints a risk report (HARD_BLOCK / HIGH / MEDIUM / LOW / SKIP). It is report-only — it issues no tokens and writes nothing; it advises. It exits non-zero when any design is HARD_BLOCK/HIGH, so you can gate a script on it if you want.

Development

Requires Python 3.11+ and uv.

git clone https://github.com/etoyama/insight-blueprint-skills.git
cd insight-blueprint-skills
uv sync --all-extras

# Run lint + typecheck + test
uv run poe all

See CONTRIBUTING.md for setup, code style, and PRs. Contributor/architecture docs: CLAUDE.md, docs/ARCHITECTURE.md, docs/PRD.md.

Tech Stack

Support

If you find this project useful, consider buying me a coffee.

License

MIT