bigspin

A Claude Code plugin that analyzes your local session history and renders a personal "practice mirror" — your archetype, your signature moves, and how you compare against a measured baseline of 4,846 real sessions from 172 Claude Code users. Everything runs on your machine, with no API key required. No data ever leaves your laptop.

Two ways to run this

Option A — as a Claude Code plugin

This works best from CLI. We do not recommend using the Claude desktop app.

Open Claude Code and start a new session. Paste each slash command, one at a time. Once it finishes running, you'll receive an HTML report that opens automatically.

/plugin marketplace add bigspinai/plugins
/plugin install bigspin@bigspinai
/persona

Option B — from a local clone

Clone the repo and ask any agentic coding tool (Claude Code, Codex, Cursor, Copilot, …) to follow the skill. If unfamiliar with cloning a repo, you can simply ask your agentic coding tool to do it for you, then continue with the following instructions. From the repo root, paste this prompt:

I want to discover my Claude Code archetype.

I've cloned https://github.com/bigspinai/plugins and am running from the repo root.

Please follow `skills/persona/SKILL.md` end to end. The skill will analyze my local session history in ~/.claude/projects against the project's measured baseline corpus and produce the report. SKILL.md handles opening report.html in my browser when it's ready.

Note: This is a better experience when Auto mode is turned on so you don't have to accept each item. But we leave that choice up to you.

Either path produces the same artifacts at the same location — see below.

What you get

The /persona slash command (or the clone-and-run prompt) reads ~/.claude/projects/ (where Claude Code already keeps your session history), runs a nine-step analysis pipeline, and writes a fresh batch of artifacts to ~/.claude/bigspin/<timestamp>/:

• report.html — the full slide-style report. Opens automatically in your default browser. Mobile-vertical, screenshot-friendly.
• report.md — the same report as markdown. Portable, no images, can be pasted back into Claude Code.
• hero.md — a tight summary (~10 lines) printed inline in the chat right after the run finishes. The thing you actually read first.
• hero_card.txt + hero_card.plain.txt — CLI hero card with and without ANSI color.

Report style

One HTML report ships with the plugin: a Bigspin-branded slide reveal with build-up cards and per-archetype illustrations, rendered from report_content.json. Best for sharing on social or one-screen-at-a-time discovery.

Privacy

Everything runs locally on your machine. No upload, no telemetry, no third-party request.

• Session data: read from ~/.claude/projects/ (where Claude Code already stores it). Never copied off-disk.
• Analysis: a Python pipeline + Claude Code subagents you spawn yourself. Subagents inherit your Claude Code session — no separate API key, no separate vendor.
• Output: written to ~/.claude/bigspin/<timestamp>/ on your machine.

You can audit the whole pipeline — it's ~4 K lines of Python plus markdown skill instructions, all bundled in the plugin.

Requirements

• A Claude Code-compatible agent for the structured tagging step. Claude Code itself is the reference target; the clone-and-run flow works with any agentic tool that can spawn subagents.
• Python 3.10+ OR uv (curl -LsSf https://astral.sh/uv/install.sh | sh). On first run, the pipeline auto-bootstraps a venv at ~/.claude/bigspin/.venv and installs jinja2 + jsonschema into it. Idempotent on subsequent runs.
• Some Claude Code history. 30+ sessions makes positioning stable; 10–30 still works with reduced confidence; under 10 produces a graceful "small history" version of the report.
• Time. ~5–10 minutes wall-clock for a 30-session run, almost all of it the structured tagging step where subagents read transcripts in parallel.

How it works

Three layers of analysis combine into one report:

1. Deterministic signals (iteration count, tool diversity, course corrections, tests attempted, …) — computed from message structure in <1 second.
2. Structured interpretive tagging — 5 persona-tagger subagents in parallel tag ~36 signals against a fixed taxonomy. Produces aggregated rates positioned against the corpus baseline.
3. Open behavioral observation — one persona-tagger subagent in open mode reads ~20 sessions schema-free and writes the rich behavioral findings (distinctive patterns, sensitivity, suggested experiments).

The two tracks synthesize into the final report: the structured side gives the archetype label and the comparison bars; the open side gives the recognition lines, the suggested moves, and the framing voice.

Full methodology lives in skills/persona/analysis/interpret.md. The corpus baseline (measured 2026-05-01) is documented in skills/persona/baselines/README.md.

Contributing

This repo is a synced mirror — the source of truth lives in a private monorepo. Do not file PRs here. Open issues for bug reports, but pull requests against bigspinai/plugins will be closed in favor of upstream changes.

License

MIT. See LICENSE.