Agent skills enablement kit
A small, tool-agnostic kit for rolling out AI assistants, coding agents, Agent Skills and MCP connectors inside an organisation: the counting is deterministic Python, the teaching and the judgment stay with people. Part of grok-bot-local-first (lib/enablement/), standard-library Python, MIT.
What is in it
- Skill lint for a shared library of Agent Skills (
SKILL.md): name and folder, a description that says when to use the skill, referenced files present, no personal paths, no credentials. The same file format is read by several coding agents, so a clean skill can be shared across tools.
- Adoption KPIs from a usage export: weekly active users, adoption rate, power users and a 0–4 maturity level per function, cost per active user, functions below target, spend out of line with usage.
- Governance intake: plain rules sort proposed AI tools and MCP connectors into fast-track, Tech/Legal forum review or reject, and order the agenda.
- Playbook: rollout waves, standards, a three-level output verification protocol, KPIs, governance.
- Workshops: three hands-on 90-minute sessions for development teams (prompting and context, Agent Skills, AI in testing and review), one also in Swedish.
Example run (synthetic data)
| action | result |
skill-lint | 3 example skills: 2 clean, 1 with a non-portable name, no trigger, a missing file and a personal path |
adoption | 2026-W42: overall 49 %; three functions below the 50 % target; one function flagged for spend out of line with usage |
intake | 5 proposals: 1 fast-track, 3 to the forum (Legal items first), 1 rejected (restricted data, vendor outside the EU/EEA) |
Architecture view: retrieval, skills and agents
For AI search and agent-driven flows over unstructured data, the kit adds a small strict-TypeScript module, lib/enablement/ts/skill-router.ts: selectCards() picks which Agent Skills and MCP tools to load for a request within a cost budget, and packContext() packs retrieved chunks into a context budget behind an interface that a vector or hybrid search can replace. ARCHITECTURE.md shows the full path: ingest and chunk → budgeted context → agent step → verification (schemas, grounding, secrets, trace evals) → human approval for anything that writes, spends or sends → adoption and cost KPIs.
cd lib/enablement/ts && npm install && npm test # tsc --strict + node --test
Run it
python3 lib/enablement/enablement-kit.py --action skill-lint --dir path/to/skills
python3 lib/enablement/enablement-kit.py --action adoption --events usage.csv --headcount headcount.json
python3 lib/enablement/enablement-kit.py --action intake --proposals proposals.json
python3 -m unittest discover -s tests
Built with AI coding agents; deterministic tests gate the code. Markus Isaksson · GitHub