insight-harness

A Claude Code skill that generates a comprehensive profile of your Claude Code harness — skills, hooks, workflow patterns, tool usage, token consumption, and plugin inventory across the last 30 days.

It's a superset of /insights — everything /insights provides, plus token breakdowns, tool usage stats, skill inventory, hooks, agent patterns, and cost.

What it captures

• Tokens — input/output/cache usage over the last 30 days
• Tools — which built-in and MCP tools you invoke most
• Skills — inventory of skills you have installed and which ones you actually use
• Hooks — your configured hooks from settings.json
• Agents — subagent usage patterns
• Cost — dollar breakdown by model and by day
• Plugins — installed plugin marketplaces and plugins

How it runs (and what it doesn't do) 🧘

One Python script and an HTML template. No npm installs, no pip installs, no native binaries, no background daemons.

🐍 Runs on your machine:

• One python3 invocation using the standard library only (json, re, pathlib, subprocess, base64, etc.). Nothing to pip install.
• One optional gh api user call to read your own GitHub login for the filename — skipped if gh isn't installed. No data transmitted.
• Exits when the report is written. Nothing stays resident.

📂 Reads (locally, read-only):

• ~/.claude/settings.json, ~/.claude/plugins/installed_plugins.json, and every SKILL.md frontmatter under ~/.claude/skills/
• ~/.claude/projects/*/*.jsonl — field-whitelisted: only tool names, skill names, hook events, tool-transition sequences, and workflow-phase classifications. Never tool arguments, message text, tool results, or project file paths.
• Per shareable skill: README.md and assets/hero.{png,jpg} (see scrubbing below).

🚫 Doesn't do:

• No network calls from the script itself. urllib is imported but never invoked; the only outbound traffic is the optional gh call above.
• No uploads. Output is written locally to ~/.claude/insight-harness/report.html. You choose whether to share it at insightharness.com.
• No telemetry, analytics, cookies, or service workers. The HTML is a single static file with inlined images — open it offline, email it, host it yourself.

🛡️ PII scrubbing before anything ships: git name/email, OS username paths (/Users/<you>, /home/<you>), GitHub URLs with your username, and @<you> mentions are redacted. Hero images are text-scrubbed only — pixels are out of scope, so review screenshots before uploading. SVG heroes are rejected outright.

Install (plugin marketplace)

Run these two commands inside Claude Code:

/plugin marketplace add kabirdos/insight-harness
/plugin install insight-harness@kabirdos-insight-harness

Then in any Claude Code session:

/insight-harness:insight-harness

…or just ask: "run insight harness", "show my harness", "what skills do I use".

Install (curl fallback, for power users)

If you'd rather not use the plugin system, you can drop the skill directly into your skills directory:

mkdir -p ~/.claude/skills/insight-harness
curl -sSL https://raw.githubusercontent.com/kabirdos/insight-harness/main/skills/insight-harness/SKILL.md \
  -o ~/.claude/skills/insight-harness/SKILL.md
curl -sSL https://raw.githubusercontent.com/kabirdos/insight-harness/main/skills/insight-harness/scripts/extract.py \
  --create-dirs -o ~/.claude/skills/insight-harness/scripts/extract.py

Or clone the whole repo and symlink:

git clone https://github.com/kabirdos/insight-harness.git
ln -s "$(pwd)/insight-harness/skills/insight-harness" ~/.claude/skills/insight-harness

Output

The skill writes a single-file HTML report to ~/.claude/insight-harness/report.html and opens it in your browser. Read it locally, or upload it to insightharness.com to share as a public profile.

License

MIT — see LICENSE.