A skill-specific LoRA beats prompting the full SKILL.md by 5.2 points and cuts token cost 6.6%

Notes on Skill-to-LoRA: From Using Skills to Learning Behaviors for Token-Efficient LLM Agents (arXiv:2606.16769) — Tianyi Zhang, Zhong-Kuo Qi · June 2026

Note published July 23, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

AI-assisted notes · reviewed by SkillFed Frontier & other

Skill-to-LoRA (S2L) treats a SKILL.md file as training data, not runtime cargo. Offline, a teacher model reads the full skill document and generates synthetic task-response pairs that demonstrate the behavior it describes; those pairs train a small LoRA adapter for that one skill, with the base model frozen the whole time. At inference, the adapter loads and the SKILL.md text disappears from the prompt entirely — what's left is a lightweight skill-id tag, and the model acts on a learned behavioral prior instead of re-reading procedure on every step.

The setup: Qwen3.6-27B on a 21-skill, 210-task slice of SWE-Skills-Bench. S2L passed 65/210 tasks against 59/210 for no-skill prompting and 54/210 for stuffing the full skill text into context — a 2.9-point gain over the baseline and 5.2 points over full-text prompting. Token cost moved the same direction: S2L ran 4.89% below the no-skill baseline while full-text prompting added 13.39%, leaving S2L's per-step cost 6.6% below full-text prompting's in absolute terms. Per-skill, S2L matched or beat full-text prompting on 18 of 21 skills and the no-skill baseline on 15 of 21 skills. The controls rule out a generic fine-tuning explanation: loading the wrong skill's adapter, or one adapter shared across all skills, both hurt performance relative to S2L's skill-specific adapters. Retrieval mismatch is where the method's edge shows most: full-text prompting collapsed from 35.6% to 27.8% pass rate when the wrong skill got retrieved, while the LoRA adapters degraded more gracefully, from 46.7% to 35.6%. A configuration sweep settled on a 64-example, rank-16 adapter (about 6.03M parameters) that matched larger, higher-rank variants.

Key numbers

S2L pass rate vs. Full Skill Text	65/210 vs. 54/210 (+5.2 pp)
S2L pass rate vs. no-skill baseline	65/210 vs. 59/210 (+2.9 pp)
Token cost vs. Full Skill Text prompting	-6.6% per step
Skills where wrong/shared adapters hurt performance	both controls degrade vs. S2L
Selected adapter size	rank-16, ~6.03M params (~0.022% of base model)

What this means for skill authors

This is the clearest empirical case yet that a procedural skill and its SKILL.md file are not the same object. The document turns out to be training signal, not runtime cargo — once a model has been shown what a skill changes about its behavior, it can carry that change as a small adapter instead of a paragraph that gets re-parsed on every step. For skill authors, that reframes the deliverable. A well-written SKILL.md isn't just an instruction set; it's raw material for distillation. And the ablations rule out a lazy explanation — a Shared-LoRA trained across pooled skills underperforms one adapter per skill on every control tested, so the gain is skill-specific, not a generic fine-tuning bump.

The catch is scope, and the paper is explicit about it. S2L works best on skills with a stable workflow and a fixed output schema, where success has a clear check built in — the profile of a well-structured procedural skill, and exactly what a vetted catalog can screen for. Skills that lean on concrete code examples or open-ended reasoning are named as weaker fits: compressing them into a fixed adapter risks losing the long tail of edge cases the original text still carries. That's an argument for keeping dual-tool scanning and a human-readable SKILL.md as the source of truth even in a LoRA-serving future — the text stays the auditable spec, the adapter is a derived, disposable artifact trained from it, and a catalog that already tracks which skills are stable-workflow versus example-heavy knows which ones are worth distilling in the first place.

One implication for anyone running many skills at once: the paper's setup assumes one skill and one adapter per task, and leaves multi-skill composition and adapter routing as open problems. That's exactly the terrain a federated skill discovery layer has to reason about — not just which single skill to fetch, but how several selected skills interact once none of them are readable at runtime.

Skills related to this research

seedance-recipes Seedance-recipes provides production-ready recipe patterns for video content across genres: product, lifestyle, drama, music video, landscape, commercial, animation, and more. Each recipe preserves core creative constraints while inviting customization of subject, camera, lighting, and sound. Use recipes as proven starting shapes, not rigid templates.★ 5,445 fal-text-to-image fal-text-to-image provides access to multiple text-to-image generation models including FLUX variants, SDXL, and specialized options like Recraft for design assets. Configure parameters like guidance scale, inference steps, image size presets, and batch generation to control output quality and speed.★ 49 project-maintainer This skill performs a complete health check on projects following the Open Source Framework pipeline. It verifies phase gates, syncs task tracking across multiple sources, identifies stale content and broken links, and surfaces exactly what needs attention next. Use it for maintenance checks, progress updates, or whenever you need clarity on project status.★ 351 10 Music Video This skill transforms your music into precise visual prompts for Seedance 2.0 on Higgsfield, a platform that synchronizes generated video to your uploaded audio track. Learn to reference specific beats, describe genre-matched visuals, and apply proven hooks—from particle explosions on drops to color shifts on harmonic changes—that lock your visuals to every moment of your song.★ 712

Related notes

References

Zhang, T. & Qi, Z. (2026). Skill-to-LoRA: From Using Skills to Learning Behaviors for Token-Efficient LLM Agents. arXiv:2606.16769.
Han, T. et al. (2026). SWE-Skills-Bench: Do Agent Skills Actually Help in Real-World Software Engineering? arXiv:2603.15401.
Hu, E. J. et al. (2022). LoRA: Low-Rank Adaptation of Large Language Models. ICLR 2022.
Dettmers, T. et al. (2023). QLoRA: Efficient Finetuning of Quantized LLMs. NeurIPS 2023.
Charakorn, R. et al. (2025). Text-to-LoRA: Instant Transformer Adaption. ICML 2025.