SaschaHeyer

Use this skill when you want to treat agent system prompts like build artifacts instead of one giant hand-edited file. Covers authoring modular prompt fragments, transpiling them into a single rendered golden file with Jinja2, build-time validation for missing variables and circular imports, and CI drift checks that keep the committed prompt equal to source. Tools used, Python, Jinja2. Ships two runnable utilities, transpile.py and drift_check.py.

Use this skill when evaluating agent memory options, when running or reviewing the always-on memory agent sample (Google generative-ai repo, by Shubham Saboo), or when deciding between a self-hosted memory loop and the managed Vertex AI Memory Bank. Covers how the sample works and its rough edges, and a verified Memory Bank quickstart. SDKs used, Google ADK, google-genai, vertexai agent engines, Gemini 3.1 Flash Lite.

Use this skill when governing an AI agent's tool calls with a plain-English policy on Google's Gemini Enterprise Agent Platform, using Semantic Governance Policies (SGP). Covers deploying an ADK agent to Agent Runtime behind the Agent Gateway with agent identity, standing up the SGP engine and the VPC network path (PSC endpoint, network attachment, DNS peering), wiring the authz extension and policy, and creating a natural-language constraint that returns ALLOW or DENY on each proposed tool call. Tools, ADK (google-adk), vertexai agent_engines, gcloud network-services / compute / ai, Semantic Governance Policies.

Use this skill when adding on-device semantic search, FAQ matching, deduplication, clustering, or RAG reranking to a JavaScript or TypeScript app with no server, no API key, and no runtime model download. Covers the @ternlight/base and @ternlight/mini WASM sentence-embedding packages, the embed / cosineSim / similar API, indexing a corpus once and searching it, and the Vite and webpack bundler setup for the browser build. Works in Node, browsers, Cloudflare Workers, Vercel Edge, Deno, and Bun.

Use this skill when running untrusted or model-generated code safely on Google Cloud Run, building an AI code interpreter, or letting an agent execute Python it wrote without exposing your service credentials. Covers deploying a Cloud Run service with the sandbox launcher, running code in a sealed sandbox via the sandbox CLI, the deny-by-default network and per-call egress control, credential and metadata isolation, and the ADK CloudRunSandboxCodeExecutor one-liner. Tools and SDKs, gcloud beta run, the /usr/local/gcp/bin/sandbox binary, google-genai, google-adk.

Use this skill when running Google's AlphaEvolve, the Gemini-powered evolutionary coding agent, to discover or optimize an algorithm on Google Cloud. Covers wrapping a function in EVOLVE-BLOCK markers, writing a scoring function, driving the controller loop against a Gemini Enterprise engine, reading back the best evolved program, and the license, cost, and non-monotonic-progress gotchas. Uses the alpha_evolve client library, Discovery Engine, gemini-3.5-flash and gemini-3.1-pro-preview.

Use this skill to evaluate an ADK agent with Google's Agent Quality Flywheel, run an evalset, choose the right metric, and read the results. Covers the ADK metric menu and which metrics run locally versus which call an AI judge, how the LLM judge scores meaning rather than words, how a rubric judge writes a plain-English reason for each verdict graded against trusted evidence, and how to close the eval fix loop. SDKs and tools used, agents-cli, adk eval, google-adk[eval], the Gemini API for the agent and the judge.

Use this skill when working with TabFM, Google's zero-shot tabular foundation model, running classification or regression on a CSV or DataFrame without training, benchmarking TabFM against XGBoost or TabICL, or debugging TabFM installation and weight loading. Covers the tabfm Python package, TabFMClassifier and TabFMRegressor, Hugging Face weight download, the PyPI loader bug, CPU dtype performance, hard model limits, and the weight license.

Prepare a finished feature for review — clean the worktree, rebase on the base branch, resolve conflicts, and open or update a GitHub pull request linked to the Linear issue. Use as the final shipping step once implementation is complete and the user wants to raise or update a PR.

Autonomously execute an approved implementation plan from a Linear issue — write code, run tests, commit changes, and check off tasks in Linear. Use when the plan is ready and the user wants to build the feature or fix end to end, isolating the work in a dedicated git worktree.

Refine a rough feature request or bug report into a structured PRD or Bug Brief inside a Linear issue. Use at the start of the engineering workflow to define and clarify what to build before any planning or coding — e.g. when the user wants to write a PRD, draft requirements, or turn an idea or ticket into a spec in Linear.

Generate a detailed implementation plan with parent tasks and atomic sub-tasks from the PRD in a Linear issue, syncing it to a

Generate a narrative summary and visual storyboard (browser screenshots) of completed changes and sync it to the Linear issue. Use after implementation to produce a human-readable walkthrough and visual proof of what was built for reviewers and stakeholders.

Manages Git worktrees for isolated parallel development. Creates branches (linking ticket IDs if available), copies local environment files, and pushes changes to the remote.

Lightweight skill to capture solved problems and learnings into simple markdown files

The full Loop Runner lifecycle, interactively — GCP infra setup, first canary run, authoring new loops (interview-driven), local dry-runs, deploys and triggers (cron / GitHub event / manual), cost tracking, and querying past runs. Use when the user wants to set up or get started with loop-runner, run agents headless/unattended/on a schedule in the cloud, create/add/scaffold a new loop, deploy or dry-run a loop, wire a trigger, check what a run cost, or inspect/debug/replay a loop run.

A template example of a per-loop skill. Use this to describe custom scripts, tools, or runbooks specific to this loop that the agent can execute.

Authoritative reference for the Agentic Resource Discovery (ARD) specification — the open, federated, domain-anchored standard (by Google, Microsoft, Hugging Face and others) for cataloging, discovering, and searching agentic resources (MCP servers, A2A agents, Skills, APIs, datasets, registries) across the web via ai-catalog.json manifests and a REST search interface. Use this skill whenever the user mentions ARD, "Agentic Resource Discovery", agenticresourcediscovery.org, ai-catalog.json, the urn:air: identifier scheme, Agent Finder, an "agent registry" or "discovery service", publishing/discovering/searching agents or tools, MCP/A2A catalog manifests, trustManifest / attestations / provenance for agents, the .well-known/ai-catalog.json path, federated agent search, or how agents/LLMs dynamically find capabilities at scale instead of hardcoding them — even if they don't name "ARD" explicitly. Also use when asked to author or validate an ai-catalog manifest, build an ARD client, stand up a registry, or conne