Ivan Charapanau

Skilled - Dashboard to find agent skills you no longer need

Your AI coding tools keep traces. Skilled reads them. Live TUI dashboard that aggregates skill usage across Claude Code, OpenCode, Codex, Grok, and Droid. Reads local history files only. Zero network, zero telemetry. Frequency counts, weekly trends, hourly distribution, per-project breakdowns, and audit heuristics (rising = 50%+ increase over 4 weeks, stale = unused 30+ days, etc.). No data leaves your machine. No accounts, no config files, no API keys.

Add a comment

Replies

Best
Ivan Charapanau

I use dozens of custom skills across Claude Code, OpenCode, Grok, Codex, and Droid. After a while I realized I had almost no grounded view of which skills I was actually using, how heavily, or which ones were fading.

Skilled is the terminal dashboard I built to surface the real usage data.

It reads the local history files from those tools and renders a TUI with:

- Real skill frequency (bar charts using actual Unicode blocks)
- 16-week activity heatmap
- Time-of-day histogram
- "Audit" view showing heavy hitters, rising/declining skills, stale skills (30+ days), one-offs, and cross-project usage

The agents keep the logs. Skilled reads them.

Everything is local-only. No network, no telemetry, no accounts, no config. If the history files don't exist, it just shows nothing.

I keep it running in a bottom split now. It's become a quiet, slightly uncomfortable mirror for "am I actually expanding the skills I reach for, or just getting faster at the same five things?"

GitHub: https://github.com/av/skilled

Install with the one-liner in the README or npm install -g @avcodes/skilled. There's also a short demo video on the repo.

Curious what patterns it surfaces for other heavy users. Especially anything that surprised you.

Keith Taylor

hi ivan,

dead-skill audit tends only to get noticed once the library hits 50+ and search starts feeling slow. clean fit.

which axis do you default to for the obsolescence call: usage decay over a window, explicit user kill, or the model judging from metadata + output traces? decay misses seasonal skills, explicit-kill never gets done, model-judged is the one nobody trusts until they do. congrats on your launch, and good luck!

Ivan Charapanau

@hiyamojo thanks for stopping by!

it's a mixture of both, the tool includes multiple dimensions to analyse decay such as usage over time, cross-project usage