A meta-skill for authoring and adapting Agent Skills that run inside Gemini Spark — Google's 24/7 cloud agent. It encodes hard-won knowledge about the Spark sandbox (ephemeral VM, the Drive tool quirks, chromium-in-Docker rendering, the 5-minute RPC limit, Workspace isolation, the state-file pattern) so a skill written for Spark works on the first try instead of after many painful debugging rounds.
- Builds new Spark skills from scratch — interviews you first, then authors a full skill directory (SKILL.md + references + assets + evals) and packages it as a
.zip. - Adapts existing Claude / Cowork / Manus skills to run reliably in Spark.
- Debugs skills that loop, dump files to Drive root, or time out in Spark.
Load the skill into your agent environment, then ask it to build, adapt, or debug a Spark skill. See SKILL.md for the full workflow and references/spark-environment.md for the Spark sandbox knowledge it's built on.
spark-skill-builder/
├── SKILL.md workflow, triggers, hardening checklist
├── references/
│ ├── spark-environment.md the Spark sandbox knowledge (the heart of the skill)
│ ├── spark-checklist.md hardening checklist with pass/fail criteria
│ ├── state-file-pattern.md the resumable state-file pattern
│ ├── authoring-rules.md condensed standard skill-authoring rules
│ └── spark-evals-guide.md how to test Spark failure modes
├── assets/
│ └── spark-skill-template.md starting skeleton for a new Spark skill
└── evals/
├── evals.json test cases
├── README.md how to run them
└── files/ input fixtures
Created by Vlad Kornyshev (NGI Academy).
- GitHub: vladkor97
- Telegram channel: t.me/NGI_ru
Skills built with this meta-skill carry an attribution line in their own README pointing back here.
Free to use and adapt. An attribution back to the author and the NGI channel is appreciated.