@skills: Attention Is All You Have (2026-08-12)

The first paper in this corpus to count the published skills ecosystem and then argue that its delivery model cannot scale. Li Yin and four colleagues at SylphAI, with Zhangyang Wang (UT Austin), crawled the public corpus in July 2026 and found 56,804 SKILL.md directories across 1,133 GitHub repositories, 56,245 of the bodies fetched and analysed. Their claim about that corpus is a single sentence:

56,804 indexed skills compete for fewer than 100 reliable auto-trigger slots per agent.

The proposed fix is @skills, a delivery protocol the same authors implement and ship. Read the measurements and the protocol separately — the first is the contribution to this spoke, and the second is a product the authors sell around.

The argument: installing bundles three things that are separable

Installing a skill does three jobs at once — fetch the content, keep it for next time (persistence), and put its description in the system prompt so it fires unasked (auto-triggering). Only the third needs prompt space, and it is the one in short supply. So the paper's move is to unbundle them into three tiers: reference a path (nothing stored, 0 resident tokens), :save a copy into the project's git tree (0 resident tokens), :install one line that makes it fire on its own (frontmatter resident, 50–100 tokens).

The unbundling is the whole idea, and the mechanics of it are on atskills-protocol.

Why the budget is small (the part built from other people's measurements)

Three properties of resident context, each cited rather than measured here:

Distance decay — a description sits at the top of the context and the request drifts away from it. Attention is U-shaped over position (lost in the middle, Liu et al. 2024), system-message adherence decays over turns (SysBench), and multi-turn performance falls an average of 39% against single-turn on the same tasks (Laban et al. 2025).
The standing tax — measured on one working setup, 50–280 tokens per installed skill, so a session with 18 installed skills carries 1,500–2,000 tokens on every message. Input length alone degrades reasoning below the nominal window (Levy et al., RULER).
Dilution — instruction-following degrades as concurrent instructions accumulate: at 500 concurrent constraints the best frontier model reaches 68% adherence, with degradation visible far below that (Jaroslawicz et al. 2025).

The conclusion the authors draw is stated as a bound, not a result: "We do not claim a precise constant; no controlled measurement of trigger reliability versus installed-skill count yet exists." They propose that measurement as future work. So the headline ratio — 56,804 against fewer than 100 — has a measured numerator and an argued denominator, and the paper says so in the open rather than in a footnote. That is the honest version of a claim this corpus usually has to reconstruct.

The measurement arrived 2026-10-07. skill-shadowing (Databricks, arXiv 2605.24050) runs the sweep with every skill in context: pass rate drops 8 points at 52 skills, 14 at 102 and 21 at 202. The denominator is now measured for two Anthropic models, and the decline starts below the argued hundred.

The mechanism worth keeping is the one that runs backwards from adoption: the better the ecosystem gets at installing skills, the worse each installed skill works. context-rot has been the spoke's name for the general form; this is the version specific to skill delivery, and it makes a 2,003-skill catalog a self-defeating artifact if its contents are meant to be installed.

Four findings about what the world actually publishes

Skills have become a provider integration channel. 21 major-provider organizations publish 958 first-party skills — AWS (138), Anthropic's own plugin collections (379), Google (92), Sentry (80), Elastic (70), Grafana and LaunchDarkly (49 each), Hugging Face, Expo, HashiCorp, Cloudflare, Stripe — and the naming convention <org>/agent-skills is hardening into the address for how agents drive our product. Of 569 skills in fifteen provider repositories, 79% name the provider's own CLI (aws, wrangler, terraform, sentry-cli) and 60% contain executable shell blocks. The replacement being described is explicit: instead of an MCP server per agent, a provider ships one CLI plus one skill and reaches every agent that has a shell. This spoke recorded that migration qualitatively at perplexity-api-platform-skills; here it is counted.

Skill bodies are too big to be resident and small enough to fetch. Median body 921 words, 90th percentile 2,207 — one to four thousand tokens. Which is why install-only delivery keeps just the descriptions resident and gambles on the trigger.

54 directories for one format. The community skills CLI supports 75 coding agents using 54 distinct project-level skills directories, and 58 distinct home-directory locations. Nineteen have converged on .agents/skills/; the other 56 sit in vendor dotdirs (.claude/skills/, .windsurf/skills/, .devin/skills/, .goose/skills/ — and the authors list their own .adal/ among them). The CLI copes by installing once and symlinking everywhere. This is the sharpest number against agentskills-spec's portability story: the format is universal and the location is not, so a team running two agents maintains two installations.

The long tail already lives in registries. Two aggregator registries alone hold 11,894 of the 56,245 fetched skills (21%), and the top fifteen repositories hold 51%. Hosting and indexing are solved; delivery is what stays gated on per-user installation.

Two pathologies with names

Trigger engineering. A skill fires only if the model matches the user's message against its one-line description, and a miss is silent. So authors stop writing descriptions and start writing bids — read this BEFORE opening the file; don't skip because it looks like a one-liner; trigger whenever… — and Stripe's flagship skill carries roughly 150 words of them, costing about 20× the tokens of a quiet description. Every bid deepens the tax it is competing in. The paper names a major provider practising it, which is more useful than the general observation.

Teams defect to AGENTS.md. Rather than fight the trigger lottery, teams paste their workflows into one always-loaded file: "The crudest mechanism wins on the only axis users feel — simplicity." Set this beside agents-md, documented here through its own adoption numbers, and beside context-files-ablation-paper, which found no correctness effect from that file at all. The three sources together say something none says alone: the convention that won on delivery simplicity is also the one measured to do nothing for correctness, and it won anyway.

What is the vendor's interest here

SylphAI ships the protocol (npm i atskills), the agent that implements it (AdaL), and the hub the protocol pairs with (atskills.one) — search, hosting, private collections, one-screen authoring. The paper argues at length that the hub can never be required (gh: paths and local folders resolve with zero hub involvement) and cites git-versus-GitHub as the precedent. That architectural commitment is real and checkable in the specification. It is also, exactly, the argument a company makes when its protocol is free and its hub is the business, so the parts to trust least are the ones where the hub is the answer.

Tier T2: an original full-corpus measurement, published as a preprint (not peer-reviewed), by authors who sell the remedy. The crawl figures are the load-bearing contribution. The classification shares (30% software engineering, 25% creative production, 16% writing, 15% design; 42% requiring an external account) come from a 100-skill uniform sample at roughly ±10pp, which the authors flag at every use — cite those as sampled, not as corpus facts.

Limits the authors state

The attention budget is bounded by argument and literature, not measured by them. The corpus figures are one crawl. On security, a referenced skill is remote text an agent will act on, which is the indirect prompt-injection channel: their answer is provenance and review rather than filtering, with the tiers forming a trust ladder — a throwaway session, a git diff on save, a one-line diff before anything becomes resident. They note the risk is not new (an installed skill's body also loads invisibly) and not removed.

atskills-protocol · agentskills-spec · agent-skills · sylphai · li-yin · context-rot · cca-f-lost-in-the-middle · agents-md · context-files-ablation-paper · agentic-awesome-skills · perplexity-api-platform-skills · model-context-protocol · agent-skills-can-be-harmful · synthesis