Claude Code skills that teach the agent to write idiomatic BAML. They auto-trigger when Claude opens a .baml file or is asked "how do I install / use BAML?". Two variants of the same describe-first skill ship: core (discover each name as needed) and overview (lead with baml describe baml for the whole stdlib).

The premise: BAML is two things in one file — a statically-typed, expression-oriented language (basically TypeScript with snake_case methods and name: type, class fields: types, classes, generics, closures, optional chaining, control flow, throws/catch/catch_all, a stdlib) and a small declarative DSL for LLM calls (client<llm>, function … { client: prompt: }, generator, test) that desugars into that language.

The skill is deliberately minimal and describe-first: it names the handful of traps that trip up agents, shows one minimum-viable LLM function + test, and otherwise points the agent at baml describe for every stdlib name, type, and signature rather than padding the prompt with examples it would guess wrong. The CLI is the reference. This is the arena-winning micro-describe variant.

Every code example in the skill is verified to compile against the baml_language CLI.

Without it, agents invent stdlib methods (baml.json.decode_str, s.contains()), forget {{ ctx.output_format }} on prompts, and probe for things that don't exist. With it, they reach for baml.json.from_string<T>, snake_case names, and client: "openai/gpt-4o-mini" on the first try.

/plugin marketplace add BoundaryML/baml-skill
/plugin install baml@boundaryml-baml

The skill auto-loads in any Claude Code session.

Two describe-first skills that share the same content and differ only in how they enter it:

• core — minimal and describe-first, the arena-winning micro-describe variant. Discover each name as you need it (baml describe <name>).
• overview — same skill, full-picture-first. Leads with baml describe baml, which dumps the ENTIRE stdlib (every namespace, type, and function) in one listing, then drills into any name. Use it when you'd rather see the whole map before writing.

Both cover:

• Install + the baml describe / run / test / fmt agent loop — baml describe <name> is the reference for any module/type/method/signature
• The five fatal traps the language model otherwise gets wrong: name: type, fields (baml fmt writes the colon + comma), trailing-expression blocks + -> null unit, for (let x in xs) over values, no implicit string coercion / panicking index / (x: T) -> R { ... } closures (not =>), type-only non-exhaustive catch
• One minimum-viable LLM function + test — client:/prompt: blocks with {{ ctx.output_format }}, the return type as schema, the name$parse companion, testset/test with assert.*
• The throughline: anything not shown → baml describe <name>; check it with baml run -e

baml-skill/
├── .claude-plugin/
│   └── marketplace.json              # marketplace catalog (one plugin: baml at ./plugins/baml)
├── README.md                         # this file
└── plugins/baml/
    ├── .claude-plugin/plugin.json    # plugin manifest
    └── skills/
        ├── core/SKILL.md             # describe-first, discover names as needed
        └── overview/SKILL.md         # same skill, full-picture-first (`baml describe baml`)

# Symlink the work-in-progress copy into your global skills dir for fast iteration.
ln -sf "$PWD/plugins/baml/skills/core" ~/.claude/skills/baml-core-dev
ln -sf "$PWD/plugins/baml/skills/overview" ~/.claude/skills/baml-overview-dev
# Edit plugins/baml/skills/*/SKILL.md, then restart your session; skills auto-reload.

When happy, bump version in plugins/baml/.claude-plugin/plugin.json and commit.

• 0.1.x — content corrections, no structural changes.
• 0.2.x — added or removed sub-skills, restructured.
• 0.3.x — aligned to the canonical BAML agent guide.
• 0.4.x — collapsed the five skills into one lean, example-dense skill.
• 0.5.x — restructured around the language/DSL split; expanded language coverage; every example verified against the CLI.
• 0.6.x — aligned to the baml_language rewrite: name type fields, -> null unit, generics, optional chaining, catch_all, C-style for, richer stdlib; reconciled against a team syntax reference + the compiler.
• 0.7.x — split into four skills (core, bridges, serving, testing), each verified against the shipped CLI.
• 0.8.x — collapsed back to one minimal, describe-first core skill; the arena-winning micro-describe variant.
• 0.9.x — added the overview skill: the same content, full-picture-first via baml describe baml; corrected the closures/.collect() trap against baml-cli 0.12.1 (current).
• 1.0.0 — stable.

Apache-2.0.