Low-level skills TrothByte

Production-grade low-level engineering knowledge for AI coding agents — verified, source-backed, and cheap to load.

Note

60 skills · 23 domains · 47 source-backed · 62 primary sources · 17 traced claims. AI agents writing C, C++, Rust, assembly, kernels, or firmware fail in predictable ways — they trust "it compiles", guess ABIs, ignore memory ordering, and skip verification. TrothByte exists to fix exactly those failures.

---

🎯 What is TrothByte

Every skill answers the questions that matter before a single line is written:

1. When to use and when not to — so an agent loads the right tool and nothing else.
2. What the agent often gets wrong — the systematic mistakes, named and catalogued.
3. How to reason correctly — the positive reasoning process, not just "don't do X".
4. What to verify and how — executable gates, not vibes.
5. Where the knowledge comes from — every normative claim traces to a primary source.

SKILL.md          small, operational, loads first
  └─ references/  deep knowledge, loads on demand (progressive disclosure)
     └─ examples/ good|bad, compiled and run against real toolchains
        └─ evals/ synthetic · false-positive · historical CVE · adversarial

🚀 Quick start for an agent

1. Read AGENTS.md — engineering rules and resume protocol.
2. Route to the minimal skill set with meta-routing.
3. Open that skill's SKILL.md; load its references/ only as needed.
4. Verify like the skill says: warnings-as-errors, sanitizers, asm inspection, runtime asserts.

🗂️ Skills catalog

The complete index — what each skill does, its stability, and where it lives — is in docs/SKILLS.md. The area map below is your orientation.

Area	Domains
🔤 Languages & semantics	c · cpp · rust · concurrency
⚙️ Compilers & IR	compiler · llvm
🔩 Machine level	assembly · abi · ffi · elf · dwarf
🧠 Systems engineering	kernel · networking · embedded · bootloader · qemu
🔬 Analysis & performance	binary-analysis · reverse-engineering · performance · simd · gpu
🧰 Tooling & agent behavior	sanitizers · _meta

🏆 Flagships to start with

Skill	Solves
safe-low-level-from-scratch	writing new code that is safe by design, not by fix
compiler-ub-assumptions	"works at -O0, breaks at -O2"
memory-ordering-reasoning	races that compile and pass naive tests
abi-layout-reasoning	struct layout and calling conventions, verified with the compiler
ffi-boundary-cross-language	where one language's safety guarantees end

📂 Repository map

Path	Purpose
skills/	60 skills across 23 domains (SKILL.md + references/ + examples/ + evals/)
registry/	machine-readable state: skills, sources, claims, cross-links, tools, evals
roadmap/	coverage matrix, uniqueness analysis, priorities, live progress
research/	the original research documents this repository was built from
docs/	skill catalog, acknowledgments, architecture
tools/	shared scripts: validators, token measurement, generators
AGENTS.md	engineering rules and resume protocol
WORKLOG.md	development journal

🛡️ Quality & provenance

Important

Verification is executed, not asserted. Examples were compiled and run with GCC 16.1, rustc 1.97.1, GDB 17.2, GNU as/ld/objdump, and Python 3.11. Where a toolchain is unavailable (NVIDIA CUDA, Linux eBPF, LLVM, QEMU, sanitizer runtimes), the skill is honestly marked researched with the exact target commands documented.

• Every normative claim is traceable: claim → source → section → skill in registry/claims.yaml, backed by 62 primary sources in registry/sources.yaml.
• Quality gates run on every change: skill_lint.py · registry_check.py · source_check.py · token_measure.py.
• Token budget is measured. A typical SKILL.md costs ~1–2K tokens; deep knowledge stays in references/ and loads only when needed.

🧭 Stability levels

Level	Meaning
source-backed	claims verified by execution on a real toolchain
researched	grounded in primary sources; verification needs a toolchain not present here
evaluated / stable	reached after evals and full review (next phases)

🤝 Contributing

TrothByte is built as a product, not a pile of files. Before contributing, read AGENTS.md (engineering rules) and CONTRIBUTING.md.

📄 License

MIT — see LICENSE.md. Attribution policy for the repositories, standards, talks, and research this work builds on: docs/ACKNOWLEDGMENTS.md.