0.000 to 0.805: a 42-skill library rescues a model that can't solve a single hard RTL problem alone

Notes on LEGO: An LLM Skill-Based Front-End Design Generation Platform (arXiv:2604.23355) — Jincheng Lou, Ruohang Xu, Jie Ma, Runzhe Tao, Xinyu Qu, Yibo Lin · April 2026

Note published July 25, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

AI-assisted notes · reviewed by SkillFed Skill retrieval

LEGO breaks the digital front-end chip-design flow — spec generation, RTL coding, testbench generation, simulation, debug — into six fixed steps and represents every reusable capability within those steps as a circuit skill: a seven-field unit (name, function, constraints, entry command, IO spec, schema, completion criteria) that any code agent can invoke like a subroutine. An automated three-stage pipeline builds the library from there — summarize capabilities, map to a workflow step, standardize the fields — mining more than 100 papers and 11 open-source EDA-agent projects (VerilogCoder, MAGE, Spec2RTL-Agent, AutoBench, the hierarchy-verilog line of work, RTLFixer, and others) down to 42 skills across 24 functional groups. A companion retrieval layer, Agent Skill RAG, skips embedding models entirely: it loads only short skill descriptions up front and pulls the full entry — symptom, root cause, fix — once a match is found, keeping lookup at sub-millisecond speed while letting new fixes get appended as plain text.

The test: VerilogEval v2's 156 specification-to-RTL tasks, run end to end through gpt-5.2-codex at its highest ('extra-high') reasoning setting, filtered down to the 41 problems the model failed outright — a hard 0.000 Pass@1 floor. Handing that same model a single RTL-generation skill lifts Pass@1 to 0.56–0.61; adding a debugging skill with a retry loop pushes it to 0.805, 33 of 41 solved, an 80.5-point swing from zero. Cross-project skill composition does just as well: a pipeline built from a hierarchy-style spec skill, a hierarchical RTL generator, and an RTLFixer-derived repair skill also reaches 0.805 — beating the hierarchy-verilog baseline by 14.6 points and VerilogCoder by 2.5, tying MAGE, and beating standalone RTLFixer by 39 points.

Key numbers

Circuit skills in the library	42 (11 open-source projects, 24 functional groups)
Baseline Pass@1 (gpt-5.2-codex, xhigh, hard subset)	0.000 (0/41 solved)
Best single/composed skill result	0.805 (33/41 solved, +80.5 pts vs baseline)
Gain vs hierarchy-verilog / VerilogCoder baselines	+14.6 pts / +2.5 pts (ties MAGE)
Agent Skill RAG retrieval speed	sub-millisecond, no embedding model

What this means for skill authors

The number worth sitting with is the floor, not the ceiling: gpt-5.2-codex, run at its most expensive reasoning setting, solves zero of these 41 problems unassisted. The gap between 0.000 and 0.805 isn't a model-capability story — the same weights sit under both numbers. A validated circuit skill hands the model a procedure it wouldn't otherwise assemble on its own. That's about as clean a data point as you'll find for skill-based tooling generally: in a narrow, tool-mediated domain, the skill artifact — not the base model — is doing the work.

The more useful result for anyone building a skill library is what happens when skills get mixed across sources. RTLFixer alone solves 17 of 41 problems; folded into a LEGO pipeline next to a hierarchy-style spec skill and a hierarchical generator, the same repair logic contributes to a 33-of-41 result. The paper's explanation matters here: standalone RTLFixer works from incomplete task context and produces patches that are syntactically valid but behaviorally wrong, while inside LEGO it inherits upstream context and gets checked against skills from unrelated projects before a fix is accepted. That's the seven-field skill contract earning its keep — once name, IO spec, and completion criteria are standardized, a debugging skill built for one project slots into someone else's pipeline unmodified. It's a concrete version of the argument behind independent skill vetting: a skill's value comes from a checkable boundary, not from which repository it shipped in.

The retrieval design is a smaller but sharp point. Agent Skill RAG deliberately skips embeddings because the corpus is curated and small — 42 skills, not tens of thousands — so exact matching on short descriptions plus lazy-loaded detail can outrun semantic search and stay easy to audit. The paper notes that skill marketplaces already list thousands of daily-updated skills; that's the scale at which this shortcut stops working, and the point where retrieval quality, not just skill quality, becomes something worth vetting on its own.

Skills related to this research

translate-craft Translate Craft delivers translation that reads naturally in the target language, not like a translation. It covers 8 languages with language-specific rules for formality, tone, and cultural nuance, plus i18n framework setup (react-intl, i18next, ICU), RTL layout implementation, and locale-aware formatting for dates, currencies, and numbers.★ 110 internationalization-i18n This skill guides you through implementing internationalization across web and mobile applications using industry-standard libraries like i18next and expo-localization. Learn to organize translation files, handle locale-aware formatting for dates and currency, support right-to-left layouts, and manage language preferences without page reloads.★ 202 localization-l10n Localization-l10n guides you through adapting internationalized applications for specific regions, languages, and cultural contexts. It covers translation management with namespace organization, locale-specific formatting for dates and currency, RTL support, and regional feature handling like address formats and measurement units. The skill includes testing strategies, content delivery optimization, and integration patterns to ensure your app works seamlessly across global markets.★ 202 Solve Solve delivers methodical problem-solving for mathematical and algebraic challenges. It clarifies your specific problem, analyzes root causes, and provides executable step-by-step solutions with attention to edge cases and best practices.★ 1,538

Related notes

References

LEGO: An LLM Skill-Based Front-End Design Generation Platform — Lou, Xu, Ma, Tao, Qu & Lin, arXiv:2604.23355 (2026)
MAGE: A Multi-Agent Engine for Automated RTL Code Generation — Zhao et al., DAC 2025
VerilogCoder: Autonomous Verilog Coding Agents with Graph-Based Planning and AST-Based Waveform Tracing — Ho, Ren & Khailany, AAAI 2025
HiVeGen: Hierarchical LLM-Based Verilog Generation for Scalable Chip Design (the hierarchy-verilog baseline) — Tang et al., ICLAD 2025
RTLFixer: Automatically Fixing RTL Syntax Errors with Large Language Models — Tsai, Liu & Ren, DAC 2024