Skill Overview

cascade-script-writer is a Claude Code skill for generating and validating Python scripts that use cascade-cms-rest. It is designed for agentic workflows in Claude Code, not chat-only use: the agent needs read and write access to the directory where the script runs. It earns its keep on multi-step business logic such as create-then-edit pipelines, batch processing with callbacks, and workflow orchestration. For a simple one-off script, writing directly against the library is often the better choice.

The core workflow runs in a fixed order. The agent breaks the request down into operations, targets, and data sources, then picks the matching template from templates/INDEX.md rather than writing from scratch. It then looks up exact payload names and aliases in the bundled schema and in references/asset_api.md. Before a script writes any field, the field's path is confirmed on a real asset, using the MCP server when it is connected, because Cascade silently ignores fields it does not recognise. Only then is the script edited, validated, and presented. Scripts that write are delivered and verified in stages.

Every invocation starts with Step 0, a session log. The agent creates a new timestamped file, reports/cascade-script-YYYYMMDD-HHMMSS.txt, and fills in the header with the task quoted verbatim. It then appends each section as the matching step finishes, so an interrupted session still leaves a partial record. The sections cover the task breakdown, references read, MCP tool calls, template selected, design decisions, validation runs, staged writes, and what was delivered. The log is the agent's own working record, separate from the runtime log the library writes.

A script is complete only when validate_script.py exits 0. The validator checks the script against the bundled library source and enforces four rules: Cascade is the only entry point, Asset fields are set by attribute, everything is type-hinted, and no line exceeds 60 characters. The 60-character limit is a hard requirement. Format with ruff format --line-length 60 --isolated before validating, and wrap by hand only the lines the validator still reports.

For the full rules and workflow, see AGENTS.md and the skill's references/ folder.