@skills: Attention is all you have
Li Yin, Zhi Li, Zhan Shi, Haoran Zhang, Haebin Seong, Zhangyang, Wang
Atlas
cs.AI
2026-08-13
@skills unbundles skill delivery: paths reference any skill with zero resident tokens; only deliberate auto-trigger costs one .gitignore-style prompt line
Agent skills, the SKILL.md packages of procedural knowledge, have grown into a real ecosystem: a July 2026 crawl found 56,804 public skills across 1,133 GitHub repositories, with first-party catalogs from AWS (138 skills), Google (92), Stripe, Cloudflare, and Sentry. Skills spread faster than MCP because they are lighter: no server to maintain, just a folder of text an agent reads. Anthropic itself has been steering integrations from tool calls toward code execution over the filesystem, reporting a drop from 150,000 to 2,000 tokens on an equivalent task.
Delivery, though, is still installation. Once installed, a skill's description sits permanently in the system prompt, and the model has to match incoming requests against that one line probabilistically. The central claim: reliable auto-trigger capacity is under 100 slots per agent, and 56,804 skills are bidding for them. The constraint is physical, not a design mistake by any vendor. Resident descriptions sit at a fixed spot at the top of the context while every turn pushes the current request further away, and attention is weakest in the middle. Every resident description is paid on every message, relevant or not. More descriptions dilute each other. The more the ecosystem installs, the worse each installed skill works.
Installation bundles three separable functions (fetching content, persistence, auto-triggering), and only the last needs prompt residency. The @skills protocol splits them, with the path as identity:
Supporting decisions: a path with no SKILL.md loads as a menu of the skills beneath it instead of erroring; there is no manifest, lockfile, or registry; SKILL.md itself is unchanged. Multiple references in one message each load at their point of use, which makes multi-skill workflows deterministic, whereas under installation N descriptions must each win a probabilistic match at once. Adoption is one npm package plus a single SKILLS.md instruction file; any agent that can read files and run shell commands becomes a client.
No benchmarks; the evidence is ecosystem measurement plus the attention literature:
| Measurement | Result |
| Public corpus | 56,804 skills across 1,133 repos (full crawl, July 2026) |
| Resident cost per skill | 50 to 280 tokens; an 18-skill setup carries about 1,500 to 2,000 |
| Skill body size | median 921 words (about 1.2k to 3.7k tokens), P90 2,207 |
| First-party provider skills | 21 orgs, 958 skills; of 569 skills in 15 provider repos, 79% name their own CLI or API and 60% ship shell blocks |
| Install-location fragmentation | 75 agents spread across 54 distinct project-level skill directories |
From the literature: at 500 concurrent constraints the best frontier model reaches 68% instruction adherence, and long multi-turn settings lose an average of 39% against single-turn. Symptoms in the wild match the theory. Stripe's flagship skill pads roughly 150 words of trigger conditions into its description, and such trigger-engineered descriptions cost about 20 times the tokens of quiet ones. Two aggregator registries alone hold 21% of all fetched skills, so the long tail sits indexed and uninstalled. In the classified sample, 42% of skills require an external service account, which marks them as integration artifacts.
For anyone running Claude Code daily, this turns a vague feeling (more installed skills, dumber agent) into a mechanism with numbers. Distance decay, the standing tax, and dilution are physical properties of resident context, and they do not disappear with a better model. Teams pasting workflows into a monolithic AGENTS.md are on resident context too, with the same decay; the paper's reading is that the crudest mechanism wins on the only axis users feel, simplicity. The protocol itself is subtraction: no new fields, no mandatory hub, gh: and local paths resolve without it, and existing installs keep working. The transferable idea is the tier model, which turns prompt residency from a default into a decision.
Stated by the authors: the central quantity, the sub-100 reliable slot count, is bounded by argument and prior literature rather than measured, and no controlled experiment varies trigger reliability against install count. Corpus figures come from one crawl and one measured working setup. A referenced skill is remote text an agent will act on, an indirect prompt-injection channel answered only by provenance and review. The design trades cut both ways: no version pinning and no update lifecycle, so teams that need stability must :save and manage the copy themselves, and a reference puts a network round trip on the critical path. The discount that matters most: the author founded SylphAI, which builds the AdaL CLI and operates the companion hub, so the paper argues for infrastructure its author runs.