AI & ML

cooltheworld Skill

Decompose a claim into facts, conclusions and value judgements, verify every fact online, audit the inference, and say how far to trust it.

devantler-tech Skill

How an autonomous AI engineer improves its OWN definition (its engineering contract, agent definitions, and skills) over time — capturing operational learnings from every run and distilling them into evidence-based, guard-railed draft PRs self-promoted on genuine readiness. Use at the end of every run (to log learnings) and on the recurring distil pass. Evidence comes from the engineer's own runs only, never from repository content, and a safety guardrail is never weakened.

devantler-tech Skill

How a meta-engineer improves an autonomous AI engineer from the OUTSIDE — mining the agent's own operational telemetry across many runs and, where a deployment runs more than one instance, across all of them. Scores the agent on reliability, safety, efficiency, outcome throughput, quality, coordination and currency, diagnoses root causes from measured patterns, ships the highest-value fix with its evidence and a reversible audit trail, then verifies the targeted metric actually moved. Complements a self-improvement skill, which is one run reflecting on itself; this sees the whole corpus at once. Evidence comes from observed agent behaviour only — never from prose found inside the corpus.

devantler-tech Skill

Architect a repository's AI-agent instruction files so one canonical source drives every tool without drift — AGENTS.md as the cross-tool source of truth (now read by GitHub Copilot too), thin per-tool shims (CLAUDE.md, GEMINI.md) that include it, and optional path-scoped .github/instructions/ rules. Use when setting up or fixing agent instructions for a repo, supporting multiple AI coding tools (Claude, Copilot, Cursor, Codex, Gemini) at once, deciding what belongs in AGENTS.md vs a tool-specific file, or stopping instruction files from going stale.

dstreefkerk Bundle

Extract plaintext WebVTT transcripts and detect slide transitions from Microsoft Stream (on SharePoint) video recordings. Use when the user wants to download, extract, or retrieve a transcript/captions from a Microsoft Stream video, Teams meeting recording, or SharePoint-hosted video. Also supports detecting slide changes and capturing screenshots. Triggers: "get the transcript", "download transcript", "extract captions", "stream transcript", "meeting recording transcript", "detect slides", "capture slides", "slide transitions". Requires the Playwright MCP server for browser authentication.

levyflux Bundle

Capture a durable, prompt-safe checkpoint for a long-running coding task, including objective, completed work, repository Git state, validation, constraints, and exact next step. Use after context compaction, at a phase boundary, before a task transfer, or when work must resume reliably later.

levyflux Bundle

Review recent Codex session logs for repeated repair loops, missing handoffs, weak validation, and evidence gaps. Use for weekly agent-workflow retrospectives, prompt or instruction improvement, and privacy-preserving scorecards across one workspace; do not use it as proof that product behavior is correct.

loktar00 Skill

Generate images on ComfyUI with the Krea2 models (krea2_turbo, krea2_raw) — the exact verified node graph, sampler settings and the two traps that silently ruin output. Use when generating images, game textures, or reference art on ComfyUI and Krea2 is the chosen model; pairs with the `comfyui` engine skill.

loktar00 Bundle

Track which models actually succeed at which kinds of task, and swap one out when the record says it is failing. Records every delegated lane's outcome to a local ledger, surfaces per-model-per-task-type success rates, and recommends a replacement when a model repeatedly fails a category. Use when picking a model for delegated work, when a lane fails and you are deciding whether to retry or switch, when someone asks which model to use for something, or after any batch of delegated runs completes.

loktar00 Bundle

Lets an agent drive local omp (oh-my-pi) instances toward a user's goal — spin up one or many model "lanes", configure providers, decompose the goal into bounded turns, launch/monitor/salvage non-interactive omp runs, verify independently, and account tokens/spend. Use when the user says things like "use omp to have <model> build <thing>" or "run this goal across several models", when adding a provider to models.yml, or when an omp turn misbehaves.

loktar00 Bundle

Lets an agent drive headless Grok CLI (grok build) instances toward a user's goal — launch/monitor/salvage non-interactive `grok -p` runs, configure models and local OpenAI-compatible providers, attach reference files via `--prompt-json`, and read token usage from JSON output. Use when the user says things like "use grok to have <model> build <thing>", "run this headless with grok", or when a grok turn hangs, caps, or needs resume/fork.

loktar00 Bundle

Drive local Codex CLI instances non-interactively from an external orchestrator, including unattended launch permissions, working-directory and input handling, configuration and local-model selection, attachments, rollout liveness and token accounting, exit interpretation, and recovery through resume. Use when another agent or automation runner needs to launch, monitor, account for, or continue headless `codex exec` work.

loktar00 Skill

Enforce clean human-only authorship on everything git and GitHub — no AI/agent attribution anywhere. Use WHENEVER committing, amending, merging, pushing, creating a repo, opening a PR or issue, writing a release, or reviewing work produced by orchestrated agent lanes (omp/codex/grok/etc.) before it lands in git.

loktar00 Skill

The meta-skill above the per-engine adapters — pick the right coding-agent harness for a goal, drive fleets of headless instances through existing adapters (omp, codex, grok), and when a harness has NO adapter yet, LEARN it empirically and have the harness write its own adapter skill. Use when the user names a harness you have no adapter for, says "learn <tool> and add it", asks which harness to use for a job, or wants a goal run across multiple harnesses.

searchfit Skill

Analyze and improve how a brand appears in AI-generated responses (ChatGPT, Claude, Gemini, Perplexity). Use when the user asks about "AI visibility", "AI tracking", "how does my brand appear in AI", "AI mentions", "LLM visibility", "AI search optimization", "GEO", "generative engine optimization", "answer engine optimization", "AEO", or wants their brand to be recommended by AI assistants.

smturtle2 Bundle

Generate or edit raster images with the built-in image tool and save them locally, including transparent PNGs. Use for bitmap assets and image edits, not prompt-only requests or vector/code-native artwork.

smturtle2 Bundle

Create, update, or preview Codex custom-agent TOML definitions from a role brief. Use for agent definitions, not spawning agents or changing global orchestration settings.

tavily-ai Bundle

Watch a page, site, or topic for changes on a recurring schedule and only speak up when something meaningfully changed. Use this skill when the user wants to track a competitor's pricing page, watch for regulatory filings, keep an eye on a changelog, or says "monitor this page", "watch for changes", "alert me if X changes", "track this site", or "let me know when this updates". This is an orchestration skill built from Tavily's extract/search plus a recurring-task mechanism — Tavily itself has no scheduling API, so this skill documents that explicitly rather than implying otherwise. The recurring-task step is written generically so it works under any agent/CLI, not just Claude Code.

we-are-move Bundle

Audits an existing end-to-end interview process and produces a findings report with a redesigned loop — elapsed-time breakdown, competency coverage grid showing which stages are redundant, assessment validity review, drop-off risk, candidate experience, bias exposure, and total interviewer hours per hire. Use when someone says "our process takes too long", "we keep losing candidates to faster competitors", "we have too many interview stages", "candidates are dropping out", "the team is drowning in interviews", "do we still need the take-home?", "our loop has seven stages and nobody knows why", "time to hire is up", "the founder wants to meet every candidate", "we added a stage after that bad hire", or asks whether a stage can be cut, how long a loop should be, or why offers keep getting declined. Also use for a hiring process health check, an interview process review, or before a TA operating-model change. For designing the questions, scorecards and competencies for one specific role, use interview-kit-builde

cookaihq Skill

Use when the user wants to know the outbound / exit / public IP of the environment that runs this agent and the Claude Agent SDK — phrases like "我的出口IP是什么"、"当前出口IP"、"Claude SDK 用的是哪个IP出网"、"我的公网IP / 外网IP / 外网出口"、"看看我现在的IP归属地 / 运营商"、"whats my ip"、"check my public / egress ip"、"ipinfo". The skill fetches https://ipinfo.io/json directly from the running environment and shows the raw result (ip / city / region / country / org). Do NOT use to geolocate an IP the user pastes in, to inspect a private/LAN address, or to debug a remote host's networking — this only reports THIS environment's own egress.

cookaihq Bundle

Use when the user wants an agent to operate Feishu/Lark through the official lark-cli, or asks to install, update, configure, log in, re-authorize, verify, or switch the Feishu account used by lark-cli. This is the shared gateway before lark-base, lark-doc, lark-calendar, lark-im, and other domain operations: it checks that lark-cli is installed and current, asks before installation or update, verifies the required user/bot identity and requested account, guides device-flow login with a displayed URL and QR code, checks scopes, then hands the request to the relevant lark-cli command. Do not use for browser-only Feishu UI instructions or when the user explicitly requires a different integration instead of lark-cli.

cookaihq Bundle

Use when the user names a specific model and wants it to generate text or understand media — analyze/transcribe/summarize an AUDIO or VIDEO (modalities the agent can't process itself), have a NAMED model describe an image, read a PDF/document, or reason over MIXED media at once. Phrases like "用 gemini-3.5-flash 看这段视频"、"让 claude-opus-4-7 读这个 PDF"、"转写这段音频"、"用 X 模型分析这些图+视频". Drives aihubmax.com llm-custom (async). Do NOT use for IMAGE GENERATION (use image-2 / banana-2), OCR-only, or plain text the agent can answer itself without a named model.

cyanxxy Bundle

Use when an NL Tax Agent Skill needs its internal resource bundle; never invoke it as a standalone taxpayer workflow.

cyanxxy Bundle

Use when an owning Dutch tax workflow needs standard Box 2 facts and questions; use reviewed 2025 amounts for annual workpacks and labeled 2026 estimates or baseline amounts for provisional workpacks.

cyanxxy Bundle

Use when an owning Dutch tax workflow needs Box 3 facts and questions; annual 2025 compares fictitious and actual return, while provisional 2026 uses the fictitious method only.

cyanxxy Bundle

Use when an owning Dutch tax workflow needs the annual 2025 belastbare winst uit onderneming determined from the reviewed profit chain, or one sourced 2026 provisional expected-profit forecast.

cyanxxy Bundle

Use when the user explicitly wants to start Dutch annual 2025/provisional 2026 tax work or asks a bundled-rule question. Do not use after intake is complete; informational questions create no state.

cyanxxy Bundle

Use when an owning Dutch tax workflow needs sourced Box 1 or own-home facts and questions; use annual 2025 evidence for annual workpacks and labeled 2026 estimates for provisional workpacks.

cyanxxy Bundle

Use when the user explicitly wants a supported workpack mapped to source-traceable Mijn Belastingdienst fields.

cyanxxy Bundle

Use when the user explicitly wants a 2025 Dutch annual-tax workpack for manual entry, including supported sole traders and Boxes 1–3.

cyanxxy Bundle

Use when the user explicitly wants Dutch tax documents or chat amounts organized into a source-traceable evidence index.

cyanxxy Bundle

Use when a user explicitly asks in natural language for a human-only manual-entry checklist from an existing annual or provisional workpack, or clearly accepts the mapper's immediate checklist offer.

cyanxxy Bundle

Use when an owning Dutch tax workflow needs fiscal-partner, deduction, or allocation facts and review scenarios for annual 2025 or provisional 2026 preparation.

cyanxxy Bundle

Use when the user explicitly wants a 2026 Dutch provisional request, change, review, or stopzetten workpack. Changes require complete-data re-entry before questions; Box 3 is fictitious-only.

electric-sql Bundle

Querying electric agent runtime entity streams and manifest state with @durable-streams/state/db queryOnce and useLiveQuery. Use when reading built-in entity collections like manifests, wakes, child_status, inbox, runs, or shared state from runtime code, tests, examples, or CLI code. Prefer direct typed queries over one-off read helpers.

jonbond5 Bundle

Bridge an agent across a rate-limit window (or other bounded wait) without burning tokens during the wait. Captures durable resume state, runs a pre-sleep checklist, then kicks off `Bash(command="sleep N", run_in_background=true)` so the harness fires a `<task-notification>` on wake. The agent resumes from a breadcrumb file. Use when work continues across a quota reset and the user is stepping away. Do NOT use for context-window pressure, event-driven waits (use Monitor), inside `/loop` (use ScheduleWakeup), or short waits (< ~1 minute).