AI Agent Skills and Tools Ecosystem

Open-source agent skills launched across design, video, finance, and portable coding. - Impeccable, a frontend design skill at 66K stars, extends Anthropic's frontend-design skill to cure 'AI Slop' - AI Slop attributed to all models training on the same SaaS templates: Inter fonts, purple-blue gradients, nested cards, gray text on colored backgrounds, rounded icon blocks - Layer 1: one /impeccable entry with 23 subcommands (init, critique, audit, polish, bolder, quieter, distill, live) as a shared design vocabulary - Layer 2: 61 deterministic rules in Rust, zero LLM and zero API key: 32 'slop' and 29 'quality' checks - Detector runs standalone via 'npx impeccable detect src/', exits with code 2 on issues, CI-ready - Layer 3: hard workflow contracts - hooks auto-run detectors after UI edits, critique uses two isolated subagents, finish reviewer runs in fresh context - concept-seed script must randomly assign design direction before artifact code; Claude's 'cream background serif' preference logged as pre-spent budget - Comp-led build gates: comps, spec, plates at 1.5x resolution, hero at 72% pixel-diff pass, sections, motion, responsive; binaries measure, not model self-assessment - Ships as a static Rust binary with no Node runtime; one core compiled to static, CDP browser, and wasm engines; consistent across 17 tools - PRODUCT.md stores non-negotiable product facts; DESIGN.md is back-written from code after build - ffmpeg-skill adds 21 local FFmpeg tools to Claude Code, Cursor, Codex: probe first, stream-copy over re-encode, re-probe results, thumbnails - Tools cover cutting, silence removal, aspect fitting, karaoke subtitles, watermarks, HDR-to-SDR, noise reduction, multi-camera alignment, loudness normalization; originals never overwritten; MCP-compatible - Awesome Finance Skills: 8 skills, 10+ news sources (Cailian Press, Wallstreetcn, Weibo, Polymarket), A-share/HK/US quotes with fundamentals - FinBERT sentiment from -1 to +1, Kronos forecasting adjusted by news sentiment, signal validation tracking, transmission chain diagrams, four-step research reports - Installs via one command on Claude Code, Codex, OpenCode, OpenClaw; framed as a research assistant, money decisions left to the user - OpenClaude Portable packs the coding agent into a USB-ready folder, ~30 MB, no host files; 9 providers including DeepSeek, OpenRouter, Gemini, Claude, OpenAI, Ollama, LM Studio; offline via Ollama - Local proxy compresses system prompts from 10,000 to 300 tokens; CPU first-token latency down from 1-2 minutes to 5-20 seconds; sessions resumable by ID; cross-OS shared data directory - ego lite team clarified: overly broad skill triggering to be fixed next version; privacy terms were a mix-up, browser product keeps data local while the 'ego' agent uses their servers - Skills described as manuals covering only what models don't know; tools execute
Key insights
Impeccable's Three-Layer Architecture Targets AI Slop
Impeccable, a frontend design skill at 66K stars built by pbakaus from Anthropic's frontend-design skill, targets 'AI Slop': convergent design outputs attributed to all models training on the same batch of SaaS templates - Inter fonts, purple-blue gradients, cards within cards, gray text on colored backgrounds, rounded icon blocks above headings. The architecture stacks three layers: a prompt-layer skill entry /impeccable with 23 subcommands (init, critique, audit, polish, bolder, quieter, distill, live) forming a shared design vocabulary; a deterministic Rust detection layer; and a process contract layer. Distribution is dependency-free: each skill is a few-hundred-line launcher that downloads a static Rust binary to ~/.impeccable/bin/, needs no Node at runtime, and behaves identically across 17 tools; the core rule logic compiles three times into static HTML/CSS parsing, CDP-driven Chrome inspection, and wasm bundles.
61 Deterministic Rust Rules Detect Slop Without Any LLM
The core innovation of Impeccable is 61 detection rules implemented entirely in Rust with zero LLM calls and zero API keys, statically analyzing HTML/CSS or scanning rendered URLs. Rules split into 32 'slop' checks (side-tab colored edges, bounce easing, radial glow, blinking cursors, gradient text, em-dash overuse) and 29 'quality' checks (line width, touch targets, skipped heading levels, low contrast, overflowing text). The detector runs standalone via 'npx impeccable detect src/', exits with code 2 when issues are found, and integrates directly into CI independent of any AI tool.
Hard Process Contracts Mechanize Resistance to Model Inertia
Impeccable's process contract layer enforces behavior mechanically: design hooks auto-run the detector after UI file edits and feed findings back to the agent, with Cursor intercepting before write; critique forces two isolated subagents whose LLM judgments and detector evidence remain unpolluted before synthesis; the finish reviewer must run in a fresh context because 'a reviewer who inherits your transcript inherits your optimism'. A concept-seed script must randomly assign design direction and shuffle challengers before artifact code, countering models' convergence to category defaults, with Claude's 'cream background serif' preference logged as pre-spent budget. The comp-led build path gates comps, spec (grid-measured, Google Fonts fingerprint matching), plates (1.5x resolution), hero (72% pixel-diff pass line), sections, motion, and responsive stages by binary measurement, since documentation states models systematically overestimate their HTML reproduction of images. PRODUCT.md holds non-negotiable product facts while DESIGN.md is back-written from code after build.
ffmpeg-skill Turns Coding Agents Into Local Video Editors
ffmpeg-skill equips Claude Code, Cursor, and Codex with 21 video tools that all run FFmpeg locally, with no cloud and no API key. The workflow is fixed: probe the file before acting, prefer lossless stream copying over re-encoding, re-probe results after completion, and generate a thumbnail when the picture changed. Coverage includes cutting and concatenation, silence removal and jump cuts, duration and aspect fitting, subtitles and karaoke subtitles, watermarks and subtitle bars, HDR-to-SDR conversion, audio noise reduction, multi-camera alignment, and loudness normalization. Tools can print commands without touching files, originals are never overwritten, and the same toolset works as MCP. A sample task - keeping two segments of interview.mp4 and cutting them to exactly 60-second Reels - triggers cutting, 9:16 cropping, export, and platform-spec checking with a final duration and resolution report.
Awesome Finance Skills Adds Live Data, Sentiment, and Forecasting
Awesome Finance Skills is a set of eight financial-analysis agent skills that let agents pull news, check quotes, score sentiment, and draw transmission chains instead of giving generic answers without real-time data. News aggregation covers 10+ sources including Cailian Press, Wallstreetcn, Weibo, and Polymarket; quotes span A-shares, Hong Kong stocks, and US stocks with fundamentals. Sentiment analysis uses FinBERT scoring from -1 to +1, and time-series forecasting uses the Kronos model adjusted by news sentiment, with tracking of whether investment signals were later strengthened or falsified. Market transmission chains render as diagrams, and output follows a four-step pipeline of planning, writing, editing, and illustration into a research report. The skills install with one command on Claude Code, Codex, OpenCode, and OpenClaw; the framing is a research assistant, with real-money decisions left to the user.
OpenClaude Portable Runs a Coding Agent From a USB Drive
OpenClaude Portable packages the open-source coding agent OpenClaude into a single folder that runs from a USB drive on any computer without leaving files on the host. First run auto-downloads Node and the engine for about 30 MB total, then a one-time provider selection leads to a launch menu. Nine providers are supported, including DeepSeek, OpenRouter, Gemini, Claude, OpenAI, Ollama, and LM Studio, with local Ollama usable offline. A local-model acceleration proxy compresses the system prompt from 10,000 tokens to 300, cutting first-token latency on CPU-only machines from one to two minutes down to 5-20 seconds. A web panel resumes interrupted sessions by ID, and Windows, Linux, and macOS share one data directory, avoiding reconfiguration across systems. The noted use case includes company computers that prohibit software installation.
ego lite Clarifies Privacy Mix-Up and Overly Broad Skill Triggering
The ego lite team reached out to clarify two issues. First, skill triggering was too broad; after user feedback requested default operation without specifying a skill, the current version's description is inaccurate, and a fix is promised for the next version - a move characterized as the common product mistake of changing products directly on user feedback. Second, the data privacy concern was a mix-up: the team's in-development agent product 'ego' processes context through their servers, while the released browser product 'ego lite' keeps data local to the agent and does not upload it; the published privacy agreement mistakenly used ego's terms, and future data collection would require a consent popup. The team offered dedicated staff to answer and fix user-reported issues.
Skills Defined as Manuals Covering Only What Models Don't Know
Skills are characterized as instruction manuals for agents that need only state what the model does not know, while execution is performed by tools - scripts, CLIs, apps - that the model calls directly. When a model does not know how to use a specific tool or execute a specific workflow efficiently, the skill is the component that bridges that gap.