# Toloka agent skills

This is the public bootstrap for Toloka workflows. Choose the path that matches the user's goal, then read the matching runbook before acting.

## Choose a workflow

- First labeling project or pipeline: follow the [data-labeling runbook](https://platform.toloka.ai/agents/data-labeling.md). Install `toloka-build` for pipeline authoring, tests, plans, and apply.
- Supervised fine-tuning: follow the [fine-tuning runbook](https://platform.toloka.ai/agents/fine-tuning.md). Install `toloka-fine-tuning`.
- Prompt gisting: follow the [gisting runbook](https://platform.toloka.ai/agents/gisting.md). Install `toloka-gisting`.

## Prepare local agent entry points when needed

Reuse a compatible workspace and current installed skill. Run `toloka init` only when the workspace's Agent Bootstrap is missing or stale; Training-only workflows use the runtime-inclusive standalone CLI and need no package manifest, SDK, TypeScript, or test-package dependencies.

Search the bundled catalog and install the skill for the selected workflow when it is absent:

```sh
toloka skill search <workflow>
toloka skill install <skill-id>
```

Read `.agents/skills/<skill-id>/SKILL.md` before acting. `toloka-build` owns pipeline source, local checks, validation, plan, apply, and run semantics. Training skills own Training, Training Run, Model, and Model Version workflows.

## Safety gates

- Never ask the user to paste an API key into chat. Use `toloka login`.
- Never invent project, Dataset, Training, Model, Model Version, price, or run IDs.
- Show the user the exact plan or maximum charge before a paid mutation.
- Keep waits bounded and report observed statuses honestly.
- Do not claim deployment or inference readiness from a completed Training Run alone.