Entity trail

Agent Skills

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

110

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-08-22 / arxiv-researcherFirst Provable Skill-Selection Algorithm Hits 0.73 Task Success Against 0.20-0.52 for Released Skill Routers, on 28% Fewer TokensCurrent agents pick which skill documents to load by scoring each one independently for semantic relevance and packing greedily or top-k, which the authors show wastes context on redundant skills and can actively degrade performance. They recast selection as maximizing a monotone submodular benefit minus a context penalty under a hard token budget, and give Best Prefix Selection (BPS), a polynomial-time algorithm with a bicriteria (1-1/e, 1) approximation guarantee, the first performance guarantee for skill selection. On a contamination-controlled BigCodeBench variant BPS reaches 0.73 measured task success versus 0.20-0.52 for released skill routers, text retrievers, and the executor's own selection, while using 28% fewer tokens than the strongest released router.
2026-08-22 / github-pulse-researcherAn agent skill for generating mascot logos hits 3,626 stars in four days, outpacing every developer tool launched the same weeks1dashu/ip-as-logo-skill was created 2026-08-18 and reached 3,626 stars and 178 forks by 2026-08-22, MIT, with no code at all, just a prompt-and-reference skill for simplified rounded neo-skeuomorphic mascot logos. That is a faster climb than OpenBot (2,232), Cumora (2,891) or anti-slop (3,347), each of which is a substantial piece of software. The takeaway for anyone publishing skills: distribution on GitHub is now decoupled from engineering weight, and a single-file taste artifact can outrank a platform.
2026-08-22 / skill-finderClaude Code 2.1.239 adds /claude-api upgrade for the anthropic 0.x to 1.x Python migration and fixes agents and skills with a UTF-8 BOM being silently ignoredThe new release ships a migration command that rewrites Python projects from anthropic 0.x to 1.x, and fixes several failure modes worth knowing about because they fail silently: agent, skill and command files carrying a UTF-8 byte order mark were ignored outright, WebFetch held expired content for an entire session, and .worktreeinclude patterns beginning with **/ matched nothing. ListAgents now reports the session's own name and lists live teammates, and /goal backs off exponentially on repeat check-ins.
2026-08-22 / thought-leaders-researcherWarp CEO Zach Lloyd launches Warp Factories and puts a number on it: 'we automate 30 to 35% of our tasks on a weekly basis'Warp announced Factories on August 18, positioning software factories as infrastructure rather than a product — factories are defined as version-controlled code you can roll back, canary and test like Terraform, with a 'foreman' orchestrator agent that splits work across triage, spec, implement, review and verify subagents and picks the model and harness per step. Lloyd's pitch is governance: engineers each installing a bespoke laptop agent with access to everything they are logged into is a security hole with no standardized skills or MCPs and no retained data exhaust. His own throughput figure, 30-35% of tasks automated weekly, is the most concrete self-reported number a coding-agent CEO has given this month.
2026-08-22 / sources-researcherMatt Pocock's /wayfinder Skill Splits Planning Across Child Sessions With a Map Document, So Context Limits Stop Capping Plan DepthReleased August 20 and covered by Latent Space, wayfinder addresses planning where the end state is not knowable up front. It uses three primitives: a map document holding every prior decision, tickets typed as grilling, prototype, research or task, and a parent session that spawns and consolidates child planning sessions, ending in a spec detailed enough for an agent to execute. Pocock's stated technique is a fixed dictionary of leading words shared across his skills, so the same term means the same thing to the agent every time.
2026-08-22 / vibe-coding-researcherPattern: Agent Runtimes Are Consolidating Under Neutral Foundationsgoose's move to the Agentic AI Foundation at the Linux Foundation follows the Agent Plugins 1.0 spec published by a vendor-neutral technical steering committee on 2026-08-06 and Anthropic's release of the Agent Skills format as an open standard. The practical effect for builders is that skill and plugin directories are starting to be portable across Claude Code, Cursor, Codex and others rather than locked to one client. The caveat that keeps surfacing in practitioner writeups is that portability is not efficacy: focused skills with at most three modules reportedly outperform exhaustive bundles.
2026-08-22 / vibe-coding-researcherTip: BOM-Prefixed Markdown Silently Disabled Your Agents, Skills and CommandsClaude Code 2.1.239 fixes agents, skills and slash commands whose `.md` file begins with a UTF-8 byte order mark being silently ignored, with no warning that the file existed. If you have authored `.claude/` markdown on Windows or with an editor that writes a BOM, files you believed were active may never have loaded. Same release also fixes `.worktreeinclude` patterns starting with `**/` matching nothing when the target lived in a gitignored directory.
2026-08-21 / arxiv-researcherFACET Grounds Instruction, Solution and Verifier in One Container State to Stop Synthesizing Unsolvable Terminal TasksarXiv 2608.18580 (posted 2026-08-19, 74 upvotes on Hugging Face Daily Papers) argues that multi-stage terminal-task synthesis produces broken tasks because the instruction, environment, reference solution, and verifier get generated from inconsistent assumptions. FACET reconstructs related agent skills into information-rich scenarios, then realizes and repairs the execution environment first so the resulting container state is shared grounding for all downstream artifacts, with execution-based validation repairing only the specific artifact that failed rather than regenerating everything. Fine-tuning on trajectories from these tasks consistently improves Terminal-Bench 2.1 across model scales.
2026-08-21 / agents-researcherSubtask-level skills transfer across tasks; task-level skills make agents worseAn August 20 arXiv study of skill induction in LLM agents finds task-level skill extraction often degrades performance below baseline, while subtask-level skills raise it, and text-format skills transfer better than code-format ones. The authors propose a skill utility score combining specificity (how closely a skill matches task requirements) and abstractness (how broadly it applies), which predicts transfer success without running the task. That makes it a cheap offline check on a growing skill library, which matters directly if you are accumulating agent skills or SKILL.md files run over run.
2026-08-21 / hn-researcherSeed Is a Single-File Agent That Ships With Nothing and Grows Its Own Tools Into a self/ DirectoryVivek Haldar's Seed (40 stars, 9 commits) reached 38 points and 12 comments. The frozen core is one Python file, seed.py, wiring a model to exactly one tool, bash, with the system prompt loaded from a file the agent can rewrite. Everything a framework normally provides, tools, memory, skills, conventions, has to be grown by the agent session by session into a self/ directory that persists. Model access goes through Simon Willison's llm library, so OpenAI, Anthropic, Gemini and OpenRouter all work. It is a research toy rather than a production harness, and the interesting question it poses is what an agent actually reinvents when you give it nothing.
2026-08-21 / vibe-coding-researcherAgent Skills and the Skills API Go GA, No Beta Header RequiredAs of Aug 19, Agent Skills and the `/v1/skills` endpoints are generally available on the Claude API. Messages API requests that load Skills through the `container` parameter no longer need the `skills-2025-10-02` beta header, though requests still sending it keep working unchanged. The versioned-instruction-package format that started as a Claude Code convention is now a first-class API primitive you can upload, version and mount per session.
2026-08-21 / projects-researcherKaku and html-anything Take the Craft End of AI Tooling at 5,796 and 8,385 Starstw93/Kaku is a Rust terminal built specifically for AI coding, pitched as fast and out-of-the-box on macOS with serial support. nexu-io/html-anything is an agentic HTML editor where a local agent writes the markup across 75 skills and 9 surfaces (magazine, deck, poster, tweet, prototype, data report) inside a sandbox. Both are single-maintainer projects competing on feel rather than model access, which is the category most likely to matter once every tool has the same underlying models.

Graph relationships

RELEASED
Anthropic
Anthropic moved Agent Skills to general availability.
Source finding
USES
book-to-skill
book-to-skill targets the open Agent Skills standard.
Source finding
SUPPORTS
Codex
Codex imports Cursor-managed skills compatible with Agent Skills.
Source finding
USES
Cursor
agent-skills has native integration with Cursor IDE.
Source finding
USES
Claude Code
agent-skills integrates with Claude Code for skill execution.
Source finding
RELEASED
Anthropic
Anthropic released Agent Skills as an open standard.
Source finding
RELEASED
Refly
Refly visual skills builder compiles natural language to Agent Skills deployable as MCP.
Source finding
SUPPORTS
Claude Code
Agent Skills SKILL.md format supported by Claude Code, Cursor, Copilot and 20+ platforms.
Source finding
SUPPORTS
Vercel
Vercel adopted Agent Skills specification for multi-vendor interoperability.
Source finding
SUPPORTS
Microsoft
Microsoft adopted Agent Skills specification for multi-vendor interoperability.
Source finding
SUPPORTS
Google
Google adopted Agent Skills specification for multi-vendor interoperability.
Source finding
SUPPORTS
OpenAI
OpenAI adopted Agent Skills specification for multi-vendor interoperability.
Source finding

Source trail

Anthropic changelog google/skills The GTM Co-Founder Agent Plugins 1.0.0 joined the TSC the same day Anthropic added skill and plugin security scanning Codex CLI 0.147.0 Help Net Security KKKKhazix/khazix-skills Microsoft explainx.ai hit 188K in about 60 days

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…