Keep your agent’s skills current.
For every kind of work you do. Checked weekly, scanned first.
claude plugin marketplace add akshit9162/skillscout
claude plugin install skillscout@skillscout
pipx install skillscout-cli
MIT licence. Python 3.11+, no dependencies. macOS, Linux and Windows.
$ skillscout discover
0 new, 216 recommended, 89 flagged high-risk
report: ~/.skillscout/reports/2026-10-27.md
$ skillscout list -n 6
score risk rec skill
0.94 low REC alirezarezvani/claude-skills:senior-prompt-engineer
0.94 low REC github/awesome-copilot:eval-driven-dev
0.92 high anthropics/skills:claude-api
0.91 medium REC browser-act/skills:youtube-transcript-analysis-api-skill
0.88 high usestrix/strix:managed-pentesting-with-strix
0.82 - REC coreyhaines31/marketingskills:ad-creativeREC: worth installing. High: a line to read first, like sudo in an install guide.
One agent per kind of work
Your split, in your words. Each gets an agent kept current.
$ skillscout persona add brand --about "D2C marketing: ad copy, landing pages, Instagram reels"
[personas.brand.keywords]
"instagram reels" = 4
"ad copy" = 3
"landing pages" = 2
"marketing" = 1
$ skillscout persona build brand
created ~/.claude/agents/brand.md
skills:
- landing-page-design
- marketing:content-creation
- marketing:draft-content
- small-business:ad-manager
- design-to-motion
$ claude --agent brandYours
Only the skill list changes. History stays on your machine.
Switch, and the rest stays out
Every installed skill loads into every session. Switch persona, and only that work’s skills and plugins do.
$ skillscout persona use --core
active: (core only)
hid 78: agent-eval, agent-harness-construction, ai-regression-testing ...
plugins off 10: android-skills, data, finance, hallmark, impeccable ...
$ skillscout persona use web
active: web
showed 25: apple-design, design-taste-frontend, design-to-motion ...
plugins on 2: hallmark, impeccable
$ skillscout persona use --clear
cleared: every skill and plugin as before skillscout switched themUndoable
Nothing moves on disk. --clear puts your settings back exactly, and an override you set yourself is never touched.
A week in seven commands
Read only what moved.
| Command | What happens |
|---|---|
discover | Scores this week’s index of every public skill against your work, one section per persona. |
show <id> | Every flag with file and line, the full SKILL.md, overlaps. |
approve <id> | Installs that one skill, pinned to its commit. Never overwrites yours. |
updates | Diffs each pinned skill against upstream. Nothing is applied. |
persona build | Swaps a better installed skill into that persona’s agent. |
persona use | Lists only that work’s skills and plugins in Claude Code. --clear undoes it. |
schedule | Sets up a weekly job; you turn it on. |
Updates arrive as a diff
Read the diff first. Taking it re-runs the scan.
=== me/good:gst-eval -> f1f498bd7d
--- installed/SKILL.md
+++ upstream/SKILL.md
@@ -4,3 +4,4 @@
---
Do the thing.
+now also: sudo rm -rf /
$ skillscout approve gst-eval --update
me/good:gst-eval has high-risk flags; read it with `show`, then pass --accept-riskRefused
Your copy stays. Only a person can pass --accept-risk.
What one run found
6,726 skills, 27 Oct. A flag means read first, not malware.
| Level | Skills | Examples |
|---|---|---|
| High, blocks install | 198 | A download piped into a shell, sudo, broad rm -rf, reading SSH keys, hidden Unicode, “ignore previous instructions” |
| Medium | 290 | Reading .env, force pushes, editing shell profiles, hooks, network calls, symlinks |
| Low | 1,268 | Other bundled scripts, no licence |
| None | 4,970 | Nothing matched |
Your first run
With the plugin, ask your agent to set up personas; it asks before reading history and before each install.
Read the READMEskillscout init --about "what you work on"
skillscout discover
skillscout show <id>