Tekmerion provides compiler-grounded coding-agent skills for programming languages that LLMs may not know reliably.

Tekmerion packages operational guidance for coding agents, but treats that guidance as a reviewed output—not a source of truth. Sparse training data, changing compilers, and unfamiliar standard libraries make plausible-looking imports and APIs especially dangerous. Static prompt collections cannot show whether a claim still compiles or whether an agent actually followed it.

Tekmerion connects small executable questions to versioned evidence, language-appropriate development patterns, readable skills, evaluations, and CI:

language question
    ↓
executable probe
    ↓
compiler evidence
    ↓
structured finding
    ↓
verified pattern
    ↓
installable skill
    ↓
agent evaluation

Tekmerion currently has executable advanced-example suites for all four languages. The Zig and Gleam suites deliberately contain five broad projects each rather than six thin projects: a project is included only when it adds a distinct architectural or runtime pressure that is not already exercised by a stronger example. All four skills remain experimental.

Verification currently covers Linux x86-64. Odin guidance is explicitly version-specific because Odin distributes monthly development snapshots. Newer observed versions are not silently treated as supported.

• languages/<language>/probes/ contains narrow, deterministic mini-projects.
• languages/<language>/findings/ records compiler version, platform, command, result, and primary sources.
• languages/<language>/patterns/ compares a preferred approach with a valid alternative and records context and tradeoffs.
• languages/<language>/examples/advanced/ composes verified primitives into realistic ownership, state, concurrency, networking, performance, and dependency boundaries.
• Every authored language source file carries native, colocated module/package documentation, and every exported or top-level declaration carries native declaration documentation. Repository validation rejects undocumented additions.
• skills/<language>-development/ synthesizes verified behavior into installable Agent Skills with progressive disclosure.
• languages/<language>/evals/ tests outcomes, hallucination resistance, process adherence, authenticity, and maintainability without requiring a proprietary agent API.
• Repository validation checks schemas, reciprocal links, evidence status, version consistency, and generated content.

Compilation is necessary but not sufficient. Pattern evidence adds standard-library practice, explicit tradeoffs, alternatives, and judgment-oriented evaluations so agents can produce maintainable language-appropriate code.

Advanced examples follow a longer feedback loop:

language question
    ↓
toolchain and source inspection
    ↓
executable probe
    ↓
advanced composition
    ↓
finding and pattern delta
    ↓
skill guidance and evaluation

The suite includes allocator-backed event replay, framed TCP and datagram boundaries, actor-owned state, bounded worker composition, measured SIMD/native boundaries, C ABI work, and runtime-specific Erlang adapters. It also records negative results rather than forcing parity: Zig 0.16's tested threaded network backend does not implement the requested connect timeout option, Gleam's direct SIMD story belongs behind an explicitly Erlang-specific native boundary, C3's manual sixteen-lane fold was slower than its scalar loop in the recorded run, and C3's Linux socket-pair example verifies AF_UNIX streams rather than an IP listener. See the advanced example model and four-language comparison.

The Agent Skills specification standardizes the skill directory and SKILL.md; it does not define one universal installation directory for every client.

1. Choose a directory under skills/, such as skills/zig-development or skills/gleam-development.
2. Copy the whole directory into the project-local or user-scope skills directory documented by your agent client. Preserve SKILL.md, references/, scripts/, and assets/ together.
3. Use the client's documented skill listing or inspection UI to confirm the skill name and description were discovered.
4. Ask the agent to report the detected compiler version and commands it executed. Compare support with catalog.json or run npm run compatibility in this repository.

No client-specific install command is claimed here because installation locations differ across Agent Skills implementations.

Requires Node.js 24 or newer.

npm install
npm run typecheck
npm run validate
npm run probes -- --language odin
npm run probes -- --language c3
npm run probes -- --language zig
npm run probes -- --language gleam
npm run probes -- --all
npm run patterns -- --language odin
npm run patterns -- --language c3
npm run patterns -- --language zig
npm run patterns -- --language gleam
npm run examples:validate
npm run examples -- --language odin
npm run examples -- --language c3
npm run examples -- --language zig
npm run examples -- --language gleam
npm run examples:report
npm run skills:delta
npm run generate
npm run check:generated
npm run compatibility

The probe runner streams compiler stdout and stderr rather than hiding diagnostics. Pinned Linux toolchain installers live under each language's toolchains/ directory and verify official artifact checksums.

Start with a narrow question, inspect primary sources and installed source, add or update the smallest probe, execute it, record the finding, then evaluate any proposed guidance change. Normative prose may cite only verified or deliberately scoped version-specific findings. See CONTRIBUTING.md and the evidence model.

Incorrect guidance is a bug. Report the language, exact compiler output, platform, affected reference, and a minimal reproduction. “The model says this is correct” is not evidence.

Tekmerion is an experimental open-source repository, not a claim of broad production maturity. Original Tekmerion material and fixtures are MIT-licensed. External documentation and source are summarized and linked; upstream licenses remain controlling for adapted material.