google/adk-docs
6 skills16.1k installs8.4k starsGitHub
Install
npx skills add https://github.com/google/adk-docsSkills in this repo
1Adk Dev Guideadk-dev-guide is Google’s always-on playbook for Agent Development Kit (ADK) projects. Solo and indie builders shipping ADK agents should load it at the start of every ADK session so work stays spec-driven: read DESIGN_SPEC.md first for requirements, success criteria, tool boundaries, and things the agent must not do, then implement in the agent directory with preservation rules and model choices that match the spec. The guide frames a repeatable lifecycle—understand the spec, build, run evaluations, and deploy—while pointing you to companion skills when context windows drop earlier instructions. It is aimed at builders using Claude Code, Cursor, or Codex who want agent behavior to be auditable against a written contract instead of drifting in chat. Use it whenever you extend an ADK repo, onboard a new contributor, or resume a long session after compaction. It does not replace the cheatsheet, eval, deploy, or scaffold skills; it orchestrates when to re-invoke them.2.8kinstalls2Adk CheatsheetADK Cheatsheet is a must-read reference skill for solo builders who already have a Google Agent Development Kit (ADK) Python project and need correct API shapes before the agent edits agent, tool, or workflow code. It consolidates core primitives—Agent (LlmAgent vs BaseAgent), Tool families, Session events and state, orchestration, auth, and advanced concerns like context caching and session rewind—into a single installable SKILL.md workflow with a deeper references/python.md companion. The skill is intentionally not a project generator; scaffolding stays on adk-scaffold, which keeps cheatsheet usage focused on modification and extension paths where subtle SDK mistakes are costly. For Claude Code, Cursor, or Codex users shipping agent backends or multi-agent orchestration, invoking this skill early in a coding task reduces rework from wrong class names, callback signatures, or state handling. Pair it with adk-scaffold when starting greenfield, then keep cheatsheet on the critical path for every ADK touch.2.7kinstalls3Adk Scaffoldadk-scaffold is Google’s procedural skill for solo and indie builders working in the Agent Development Kit ecosystem. It instructs the agent to treat agent-starter-pack as the default factory: gather the agent’s purpose, external tools, safety guardrails, and hosting preference, then pick the right template flags for retrieval, search, or A2A serving. The readme positions prototyping as the recommended first path while still documenting production routes on Vertex Agent Engine, Cloud Run, and GKE. Because enhancements like CI/CD and RAG attach to an existing repo, the skill spans greenfield agent creation and brownfield hardening. Use it whenever someone asks to “build me a search agent” or to add deployment plumbing without improvising folder layout. Apache-2.0 metadata signals enterprise-friendly reuse. It complements raw ADK API docs by encoding opinionated scaffolding steps agents can follow end to end.2.7kinstalls4Adk Eval GuideADK Eval Guide is a Google-authored agent skill that teaches the evaluation methodology for Agent Development Kit projects before you trust a green CI badge. Solo builders shipping ADK agents use it when scores dip below threshold and they need to know whether the failure is bad prompts, tool trajectory mismatch, judge configuration, or evalset design—not random flakiness. It explains the eval-fix loop, how to run evaluations via scaffolded make eval or raw adk eval, and points to deep references for eight evaluation criteria, match types, custom metrics, conversation scenarios with user simulators, and caveats for google_search and multimodal inputs. The guide assumes you may already have tests/eval/evalsets and eval_config.json from adk-scaffold, but still works for hand-rolled repos. It deliberately separates evaluation from deployment and API patterns so you read the right doc at the right time. Compatible with Claude Code and Codex when your repo follows ADK layout and you are iterating agent quality pre-release.2.6kinstalls5Adk Deploy Guideadk-deploy-guide is Google’s mandatory deployment playbook for Agent Development Kit (ADK) agents on Google Cloud. Solo and indie builders shipping conversational or tool-using agents should open it before the first production push, whether they used `/adk-scaffold` and its tested Make/Terraform pipeline or are wiring a minimal ADK CLI deploy. The skill routes you through Agent Engine, Cloud Run, and GKE Autopilot with reference depth on Dockerfiles, scaling defaults, deploy metadata, Workload Identity, stateful session choices, and custom FastAPI hooks for event-driven workloads. It also steers infrastructure work through terraform-patterns for IAM and imports, and it tells you when to stop reading—API patterns, eval harnesses, and greenfield project layout belong in sibling skills. For Prism’s audience running Claude Code or Cursor against a repo that already contains ADK code, this skill turns vague “deploy to GCP” requests into ordered checklists aligned with official ADK deployment docs, reducing misconfigured networking, wrong session modes, and missing secrets in CI.2.6kinstalls6Adk Observability GuideADK Observability Guide is Google’s procedural reference for making Agent Development Kit deployments observable in production. Solo builders shipping ADK-based agents should read it before wiring monitoring—not after incidents stack up. It explains tiered observability choices, links scaffolded Terraform projects where Cloud Trace and prompt-response logging are already provisioned, and points no-scaffold users toward official ADK manual setup and the adk-scaffold workflow for production-grade infrastructure. Deep reference files document cloud trace and logging env vars plus BigQuery Agent Analytics for queryable agent sessions, GCS offloading, and tool provenance. Use it when configuring tracing and logging, analyzing production conversations and tool calls, or debugging why an agent misbehaves under real traffic. The skill sits at intermediate complexity: you need a deployed or scaffolded ADK service and comfort with GCP concepts. It complements Ship security and launch readiness by making post-launch iteration data-driven rather than guesswork.2.6kinstalls