Typical Workflow

Skills and the CLI are designed to work together effectively. The lai CLI provides the generators and the evaluation loop, while the actionable skills drive that loop with you. In addition, the reference skills keep an agent grounded in how LAI and Langium work throughout. This guide goes over how this synergy works in practice, moving from initializing, through evaluating and refining, to getting a reusable language skill.

The combined loop

A typical agent-assisted run, mapping each step to the skill and/or CLI command that drives it:

Step	Skill	CLI command	Outcome
1. Init	lai (reference)	lai init	lai.config.jsonc, an evals/ directory with utils.ts and a starter basic.eval.ts
2. Descriptor	lai-gen-descriptor	lai gen descriptor	a refined <language>.descriptor.yml
3. System prompt	lai-gen-sysprompt	lai gen sysprompt	a <language>.sysprompt.md
4. Evals	lai-gen-evals	—	a suite of .eval.ts files with real coverage
5. Evaluate	lai (reference)	lai evaluate	scored results to review and iterate on
6. MCP server	lai-gen-mcp	—	an mcp/ server for DSL validation
7. Language skill	lai-gen-language-skill	—	a SKILL.md for your DSL

You can run any step by hand, but the value of pairing skills with the CLI is that an agent carries out the software-engineering work around each command, and improves on the baseline descriptor and system prompt the generators emit on their own.

Where reference skills plug in

The lai and langium reference skills aren't direct steps in the loop. An agent brings them in as context to assist in the process. The lai skill tells the agent what each command produces and how the loop fits together. The langium skill gives it the grounding to describe your grammar, validation, and scoping accurately in the descriptor, prompt, evals, and generated skill. Every actionable skill explicitly leans on both in one way or another.

Where actionable skills fire

Steps 2, 3, 4, 6, and 7 above are associated with actionable skills. Most bootstrap an artifact with a lai gen ... command when none exists, then guide refinement from there. So they fit naturally after the CLI has generated its output, and before you have a version that's dialed in for your language. See each skill's entry on the actionable skills page for what it produces and when to reach for it.

Iterate, iterate, iterate

The loop isn't one-directional. Step 5, lai evaluate, is where the useful information feeds back into the loop and drives subsequent iterations:

• Low pass rates or recurring failure categories feed into lai-gen-sysprompt. The skill helps to diagnose failure patterns and makes targeted prompt improvements.
• Failures that trace back to missing or wrong project information may feed into lai-gen-descriptor. This leads to correcting the descriptor, then updating (or regenerating) the prompt.
• Shallow coverage: everything passes but only trivial cases are tested. This feeds into lai-gen-evals to broaden the suite before you start trusting the results.

After any of these changes, you (or your agent) would re-run lai evaluate and repeat. Refine descriptor and prompt, re-evaluate, and observe the score delta. Based on the delta, you would make changes, improve your evaluation suite, and run again. The resulting effect is an increasingly refined evaluation suite, and an improved AI stack.

See also

• Usage: the full lai command reference behind steps 1–5.
• Examples: the evaluation workflow end to end.
• Reference skills & Actionable skills: per-skill details.