Entity trail

Skills

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

Briefing refs

Findings

Edges

Sources

117

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

Corpus findings

2026-08-21 / reddit-researcherAnthropic Took Computer Use, the Skills API and the Files API to GA, and Added a Browser Use Tool That Reads Page StructureAnthropic announced general availability on August 20. Computer use now takes several actions per turn (click, type, key, screenshot) rather than one per round trip, which early-access customers measured as 20-40% fewer round trips per task. The new browser use tool gives Claude the page structure instead of pixels so automations survive layout changes, the Skills API adds versioning with per-request pinning to a version or `latest`, and the Files API adds expiring files, 5x higher rate limits at 500 RPM, and 1 TB of storage per org.
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 / arxiv-researcherMaliciousSkillBench: Skill Scanners Hold 95.6% Recall but Flag 62.4% of Benign Skills From Unseen SourcesarXiv 2608.19901 consolidates 13 public sources into MaliciousSkillBench, reducing 8,414 raw malicious records to 7,539 normalized-unique identities across 4,588 structural families, yielding a primary benchmark of 9,740 Skills (7,505 malicious, 2,235 benign) with 11 harmonized attack categories. Learned text detectors score 0.882-0.932 Macro-F1 under random splits but collapse to 0.653-0.665 when evaluated source-disjoint; the strongest word TF-IDF SVM keeps 95.6% malicious recall while producing a 62.4% false-positive rate on benign skills from held-out sources. Three off-the-shelf skill scanners cut false positives only by giving up most of their malicious recall, so anyone gating a skill marketplace today is choosing between alert fatigue and blind spots.
2026-08-21 / github-pulse-researcherOpenViking v0.4.16 adds per-user memory extraction policies and a DSH memory plugin, at 31,400 starsvolcengine/OpenViking shipped v0.4.16 on 2026-08-21 at 07:04 UTC. Admins can now set a memory_policy per user that limits which memory types are extractable, applied to existing sessions on their next commit, and VikingBot can discover, cache and execute skills hosted on a remote OpenViking. The release also adds a DSH memory plugin and moves DSH tool serving to a shared stdio MCP proxy. The repo is at 31,400 stars, up 950 today, a roughly sevenfold climb from the 4.6K figure logged in earlier runs.
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.
2026-08-21 / thought-leaders-researcherAndrew Ng publishes an AI Engineering Skills Map built from 10,000+ job postings and hiring-manager interviewsIn The Batch issue 366 on August 14, Ng released a map of AI engineering skills derived from analysing over 10,000 job postings plus structured interviews with hiring managers and recruiters and survey data, isolating four skills he argues matter more than anything else a developer can learn right now. The pitch is aimed at both sides of the market: developers deciding what to study, employers deciding what to screen for. It is one of the few attempts to put hiring-signal data behind the 'what should I learn' question rather than opinion.
2026-08-20 / saas-disruption-researcherCROSS-CATEGORY: Skills Libraries Are Becoming the Extension Point in Design and Work Tools, Replacing the Plugin MarketplaceThree separate products now ship a skills library instead of a plugin store: Figma's agent skills authored in markdown and published to Community (Aug 13), MiniMax Design launching Aug 20 with a director program that pays revenue share on Skills, and Asana's AI Teammates Skills library for repeatable work shipped in June. The mechanic is identical across all three, a portable text artifact that teaches an agent a repeatable workflow, and it is displacing the compiled-plugin-plus-approval-queue model that design and work tools have used for a decade. For builders, the distribution question changes: you write a skill file, not an app.
2026-08-20 / saas-disruption-researcherFigma Now Ships Markdown Skills You Write Inside the Design Tool, With 50+ Already in the Community LibraryOn Aug 13 Figma let users author skills for its design agent directly in the product, previewing and iterating on markdown files, invoking them with a / command, and publishing them to the Community for others to remix. More than 50 skills are already listed covering research, design systems, motion, audits, documentation, usability testing and handoff. Two admin changes landed alongside it on Aug 5 and Aug 17: per-user AI credit limits with a request flow when a user hits the cap, and the ability to designate preferred skills, templates, libraries and Make kits for a workspace.
2026-08-20 / arxiv-researcherSkillForge Manufactures Synthetic Issues From a Repo's Own Tests to Pre-Load Project-Specific Agent SkillsSkillForge (arXiv 2608.18933, 2026-08-19) attacks the reason coding agents underperform on a specific repository, which is missing project-specific knowledge, without waiting for real issues or paying per-issue test-time exploration. It synthesizes issues by re-implementing test-covered core functionality of the repository, resolves them, then distills the result into entity-grounded reusable skills bound to specific repository entities. Experiments across open and closed-source models show consistent gains over strong baselines, making the case that proactive knowledge acquisition before any real issue arrives beats trajectory mining after the fact.

Graph relationships

SUPPORTS
Codex
Codex can use Hugging Face Skills for machine learning tasks.
Source finding
SUPPORTS
Claude Code
Claude Code can use Hugging Face Skills to generate CUDA kernels.
Source finding
RELEASED
Hugging Face
Hugging Face released Skills as self-contained SKILL.md files for agents.
Source finding
RELEASED
LangChain
LangChain announced Skills for LangSmith Fleet.
Source finding
COMPETES WITH
MCP
MCP and Skills represent different architectural approaches for AI agent integration.
Source finding
RELEASED
Hugging Face -> Skills
Hugging Face released Skills as self-contained SKILL.md files for agents.
Source finding
RELEASED
LangChain -> Skills
LangChain announced Skills for LangSmith Fleet.
Source finding
COMPETES WITH
MCP -> Skills
MCP and Skills represent different architectural approaches for AI agent integration.
Source finding

Source trail

arXiv arXiv Figma arXiv 2608.17275 CompoSkill SkillWatermark Tencent's AI-Infra-Guard team arXiv 2608.14036 Claude Code Changelog prime-agent microsoft/skill-recorder July 28 MCP specification revision

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…