The End of Hand-Written Skills — From SKILL.md to Self-Evolving and Latent Skills · leejk

TL;DR. In the design unit called the skill, the human is pushed out one notch at a time. A hand-written SKILL.md assumes two things — the author is human, and the skill lives in context. The last month of papers pries off both assumptions in turn. ① The form leaves text (structured and multimodal skills), ② the author leaves the human (self-evolution), ③ the residence leaves the context (latent skills in the weights). But what's actually in your hands today reaches only partway into ①; the departing author and residence of ②③ are still a research track. The trajectory is sharp, but what you deploy in operation is still text a person wrote.

1. Hand-Written Skills — Where We Stand Now

The Agent Skill overview laid it down: "the skill is the design unit after prompts and context." In experts we watched Karpathy, Matt Pocock, and Addy Osmani package their entire way of working into SKILL.md; in anti-slop we watched AI-ness removal harden into a catalog a human wrote. All three share one premise — a skill is a text file a person writes by hand, and it works by loading into the model's context.

That premise is today's standard. And to the degree it's become standard, the scale has piled up too. The survey Agent Skills for LLMs: Architecture, Acquisition, Security, and the Path Forward organizes the field along four axes — architecture, acquisition, deployment, security — and reports that 26.1% of community-contributed skills carry vulnerabilities: the hand-written skill corpus is already large enough to need auditing. The hand-written era has matured; it hasn't ended.

But two nails are driven into this premise. Pull them out and lay them down:

Nail	Today's answer
Who writes it (author)	a human
Where it lives (residence)	a file inside the context window

The three acts below pull these nails in order. Form → author → residence.

2. The Form Leaves First — From Text to Structure and Multimodal

What shakes first isn't the author but the form. A person still makes the skill, but the artifact shifts from prose you read to an object a machine parses.

Structuring. From Skill Text to Skill Structure decomposes an unstructured SKILL.md into a typed three-layer JSON graph (Scheduling-Structural-Logical). Borrowing from Schank and Abelson's cognitive-representation theory (Memory Organization Packets, Script Theory), it disentangles a skill into scheduling, structural, logical layers — and the point is that the effect is measured: in Skill Discovery, MRR@50 goes 0.649 → 0.729; in Risk Assessment, macro F1 goes 0.409 → 0.509. Turn prose into types and both retrieval and auditing improve.

Multimodal. Some domains can't be captured in text alone — they need where to look. Agent Skills Should Go Beyond Text: The Case for Visual Skills points out that in GUI and visual-centric tasks reusable knowledge depends on spatial layout, visual grounding, localized state changes, and adds static priors (stable spatial conventions), dynamic priors (in-situ visual working memory), and interleaved skills (binding text steps to the source frames that justify them). MMSkills sits in the same spot — each skill becomes a state-conditioned package of a textual procedure + runtime state cards + multi-view keyframes.

Up to here the author is still human. But the moment the form leaves the human's handwriting, having a person fill that form by hand becomes steadily less efficient. There's no reason to hand-place every typed graph and keyframe. The next act fills that vacancy.

3. The Author Leaves — The Agent Grows the Skill a Human Used to Write

The second nail comes out. The premise that the author is human drops. The agent writes, refines, and verifies its own skills.

Skills-Coach — uses training-free GRPO to iteratively optimize a skill's instruction and code. It generates diverse tasks (Diverse Task Generation), runs the original and optimized skills against each other (Comparative Execution), and scores against criteria (Traceable Evaluation). It never touches the model weights — it evolves the skill text itself as if it were a policy.
EvoSkills (CoEvoSkills) — co-evolves a skill generator and a surrogate verifier. Without access to ground-truth tests, the verifier co-evolves to give actionable feedback. It lifts SkillsBench pass rate 32% → 75% and ranks first on both Claude Code and Codex. It takes direct aim at the bottleneck of human authoring — label cost and human–machine cognitive misalignment.
SkillOS — a frozen executor + a trainable curator. The executor only retrieves and applies skills; the curator updates an external SkillRepo from accumulated experience. Over time, higher-order meta-skills grow inside the SkillRepo. Where the human dropped out, curation itself is learned.
OpenSkill — goes furthest. After deployment it builds both skills and verification signals from scratch — no curated skills, no successful trajectories, no verifier, only a task prompt in an open world. It scrapes grounded knowledge and verification anchors from docs, repos, and the web, synthesizes them into skills, and hones those skills against virtual tasks grounded in the anchors rather than in target answers. Code is still coming soon.

In this act the human steps back from the person who writes the skill to the person who sets up the environment in which skills grow. It's the exact opposite face of the picture in experts — "the famous practitioner packages their workflow into SKILL.md." There, the human moved themselves into the skill; here, the agent writes its own skill by itself.

4. The Residence Leaves — From File to Weights

The last nail. The premise that a skill lives in context drops.

LatentSkill: From In-Context Textual Skills to In-Weight Latent Skills uses a pretrained hypernetwork to convert text skills into plug-and-play LoRA adapters. Because skill knowledge is stored in weight space rather than context, there's no need to prefill skill tokens every step — on ALFWorld it raises seen/unseen by +21.4/+13.4 points while cutting prefill tokens by 64.1%. Moreover the generated skill LoRAs form a structured semantic geometry, are precisely controllable via the LoRA scaling coefficient, and compose through parameter-space arithmetic — the "calling other skills, versioning, composition" the overview listed as a skill's strengths now happens on weights, not text files.

This is act four's punchline — the skill stops being a file inside the context window. It drops one rung below the table the overview defined — prompt (volatile), context (session-scoped), skill (persists as a file) — to persists as weights. The residence has left text.

5. Synthesis — The Human Is Pushed Out of the Design Unit, One Notch at a Time

Line the three acts up and the nails come out in order.

Act	Nail removed	Representatives	The human's place
1 → 2	form (text)	SSL · Visual Skill · MMSkills	still author, but no longer handwriting
2 → 3	author (human)	Skills-Coach · EvoSkills · SkillOS · OpenSkill	retreats to setting up the environment
3 → 4	residence (context)	LatentSkill	—

We just saw the same curve in Where Are We Going — what survives the collapse of measurement is judgment, and judgment may not be the human's permanent seat either. There it was an argument; here it's a research trajectory saying the same thing. The skill was the design unit after prompts and context, and from that unit the human — as author, then as the manager of its residence — steps back one notch at a time.

But there's a line you mustn't overstate. Exactly as noted in the overview's academic track — these are not tools you can grab right now. What you put in .claude/skills today is still text a person wrote. Act one (handwriting) is 100% of operation; act two's structuring has started arriving as retrieval and auditing middleware, but it's not yet the form a user writes in. Acts three and four are results on benchmarks, not products shipped to production — OpenSkill hasn't even released code, and LatentSkill's numbers are on closed tasks like ALFWorld and Search-QA.

So the value of this piece isn't what to install now but which way the push is heading. The ability to use hand-written skills well — the one experts and anti-slop covered — stays valid for a while. But where that ability's shelf life ends is what these three acts point to — form first, author next, residence last. A human writing skills by hand is the first chapter of skill history, not the last.

References

Academic (primary sources)

Related