Assembly hallucinations
Invented mnemonics that do not exist (movqad), AT&T/Intel operand inversion, zeroed immediates.
Works with Claude Code · Cursor · GitHub Copilot · VS Code · Gemini CLI · OpenCode
Every skill is a compact SKILL.md: when to use, what the agent often gets wrong, how to reason correctly, and how to verify.
Named failure classes, catalogued from real incidents — not vibes.
Invented mnemonics that do not exist (movqad), AT&T/Intel operand inversion, zeroed immediates.
Thread-safe-looking code running on one thread. The gate measures real thread counts and wall-clock scaling.
Nonexistent crates that resemble real ones — a typosquatting risk (5.2–21.7% of generated packages).
A "passing" harness that does not test the target. Fixed-shape oracles certify buggy GPU kernels.
Three source-traced surveys turned real AI-agent failures into an engineering playbook.
~55 documented failures in low-level code: CONCUR, RustEvo², Ghostty VM leaks, crypto-Rust, ISO-Bench.
agent-failures-surveyInvented mnemonics, AT&T inversion, Thumb-2 errors, plausible-but-wrong decompilation.
asm-failures-survey18+ external skill repositories audited: the Zig gap, new domains, licensing and methodology lessons.
external-repos-auditgit clone https://github.com/TrothByte/low-level-skills-trothbyte
npx skills add TrothByte/low-level-skills-trothbyte
/plugin marketplace add TrothByte/low-level-skills-trothbyte
Then run /plugin install low-level-skills@trothbyte-low-level-skills
65 of 124 skills were verified by executing examples on real toolchains: GCC 16.1, rustc 1.97, GDB, objdump, CMake/Ninja.
Every normative claim is traced in registry/claims.yaml, backed by 177 primary sources.
Where a toolchain is unavailable, skills are marked researched with the exact verification command.
python tools/validate.py checks all 124 skills, the registry, and every source trace in seconds.