--- name: video-production-buddy/text-generation source: https://app.decimal.ai/s/video-production-buddy-text-generation@1/SKILL.md source_sha256: 8af3dc1cc12a --- # Text Generation Provider Calls Use this skill for standalone text-generation provider tools: - `qwen_chat` - `minimax_chat` These tools are optional provider calls for bounded text sub-tasks. They do not replace the coding assistant or become a Python pipeline orchestrator. ## Boundary Video Production Buddy is agent-first. The agent remains the primary reasoner and owner of pipeline decisions, artifacts, validation, and checkpoint writes. Use text-generation tools only when a specific bounded text job benefits from a billed provider call, such as: - bulk script polishing - translation variants - long-document summarization - style rewrites - structured copy alternatives Do not auto-wire these tools into pipeline stages. Do not use them to bypass stage director skills, artifact schemas, reviewer skills, checkpoint policy, or human approval gates. ## Provider Setup `qwen_chat` requires `DASHSCOPE_API_KEY` and calls Bailian / DashScope's OpenAI-compatible chat-completions endpoint. `minimax_chat` requires `MINIMAX_API_KEY`. The default API base is the China-mainland MiniMax host; use `MINIMAX_API_BASE` when an overseas host is required. Both providers are network API tools. Confirm setup through the tool status or `provider_menu_summary()` before proposing them as available. ## Model Choice Use Qwen when the task needs Qwen's long-context models, strong Chinese and multilingual handling, or Qwen Coder. Use MiniMax when the task needs MiniMax-M3/M2.7 behavior, MiniMax account routing, or multimodal message support exposed by the provider. Prefer the lower-cost or faster model for drafts and alternatives. Reserve the highest-quality model for final polish, difficult reasoning, or long-context review. ## Output Persistence When `output_path` is provided, it must be project-scoped and should point under `projects//artifacts/...`, `projects//assets/...`, or `projects//renders/...` as accepted by the tool. Persist outputs when the generated text becomes evidence, a draft artifact, or an input to later review. For one-off exploratory calls, returning the text in the `ToolResult` is enough. ## Cost Discipline Keep prompts narrow and explicit. Include only the source text and constraints needed for the sub-task. Avoid sending secrets, credentials, private keys, personal contact/payment data, or unrelated project files. Record provider, model, token counts, output path, and cost from the tool result when the call contributes to a production artifact.