Skillkeel

Guard hooks, tested skills and evals for Claude Code. Each skill ships with a fixture repo, graders and the unedited transcript of a real run, so you can read what it does before you install it.

Terminal recording: guard-bash refusing five destructive commands inside Claude Code

Every call in the recording is the real hook fed the JSON Claude Code sends.

Starter, free and MIT

guard-bash refuses destructive shell commands (rm -rf on root paths, force-push without lease, git reset --hard, recursive rm hidden behind find -exec, git settings that run commands). secret-scan refuses file writes that contain credentials. session-start prints a short project brief. The tests, the eval cases and the per-version results are all in the repository.

/plugin marketplace add skillkeel/skillkeel-starter
/plugin install skillkeel-starter@skillkeel

Source, tests and transcripts on GitHub · ZIP on Gumroad

Listed on ClaudePluginHub and ClaudeMarketplace, two directories of Claude Code plugins and skills.

Kit

The playbook, the six skill packs, the rest of the hooks, the subagents, the checklists and the CLAUDE.md templates, itemised below. Same rule: every skill was run against a fixture before it shipped. Two tiers, the second with access to the private repo. The price is on the Gumroad page; it is before tax, and Gumroad adds the tax of your country and shows the total at checkout.

Read before you buy: chapter 4 of the playbook, one skill's eval case with its unedited transcript, and the pass counts per Claude Code version, copied from the ZIP. · Skillkeel Kit on Gumroad

Current: Starter 0.1.14 (2026-10-21) and Kit 0.1.11 (2026-09-30). Transcript evals: Starter 8 of 8 on Claude Code 2.1.306 (2026-10-21), Kit 24 of 24 on 2.1.284 (2026-09-29). Every version tried is in docs/compat.md of each download.

What is in each one

Starter 0.1.13, free and MIT

  • 3 guard hooks: guard-bash, secret-scan, session-start
  • 8 repo skills: changelog, claude-md-init, commit-message, dependency-audit, pr-description, readme-refresh, secret-audit, test-gap-finder
  • 89 hook tests and 8 eval cases, each with its fixture, its graders and a transcript

Kit 0.1.11

  • 5 guard hooks: prod-env-guard, migration-guard, test-before-commit, large-diff-warn, cost-guard
  • 24 skills in 6 packs, with 24 eval cases
  • A playbook of 10 chapters
  • 5 subagents, 3 guardrail checklists, 8 CLAUDE.md templates
  • 41 hook tests

Counts on this page are read from the shipped repositories by tools/site_contents.py and checked against them again by make surfaces.

How it works

  1. Install the plugin with the two slash commands above, or unzip the Gumroad download into the project. Everything runs on your machine; no hook in either product makes a network call.
  2. Claude Code hands every Bash, Write and Edit call to the hooks as JSON before it runs them. A hook that refuses exits 2 and prints why, and Claude reads that text.
  3. The refusals are specific: a recursive delete hidden behind find -exec, a force push without a lease, a credential on its way into a file. Each one names its rule and the environment variable that turns it off for the session.
  4. The skills do the repeatable work: changelogs, PR descriptions, reviews, migrations. Each was run against a fixture repository before it shipped, and the unedited transcript of that run is in the download.

Writing

Eval results, guard cases and what broke: the three pieces so far are in the guide, first published on skillkeel.substack.com (subscribe, RSS); new versions are announced to Gumroad followers. Northbridge is the developer behind Skillkeel; mail to [email protected].

Contact

[email protected]