report-builder

# Report Builder The job is almost always the same shape: **data in → one HTML file out**, rendered by a template, viewable by double-clicking with no server. Keep that shape and the work stays small and reproducible. The two failure modes to design against are (1) non-determinism (charts that fetch live, timestamps that drift) and (2) unsafe data interpolation (raw values dropped into HTML or `<script>`). Both are solved below. ## The pattern ``` data (CSV / DataFrame / JSON) └─ your prep code (pandas, optional) → a plain JSON "context" └─ Jinja2 template (.html.j2) → scripts/render.py └─ one self-contained report.html ``` Separate **prep** (project-specific, may use pandas) from **render** (generic, reusable). `scripts/render.py` is the generic half — point it at a template and a JSON context file and it emits HTML. Don't rewrite it per project. Run `render.py` by its **absolute path** — it lives in this skill's base directory (announced when the skill loads, usually `~/.claude/skills/report-builder`); a relative `report-builder/scripts/render.py` won't resolve from another repo. It has PEP 723 inline deps (jinja2), so it **must** run under `uv` — never bare `python`. Define an `rbuild` function per block (a function passes arguments correctly under bash and zsh, where a plain `$VAR` doesn't word-split; re-declare it per block since shell state doesn't persist between tool calls): ```bash # uv is required here (jinja2); resolve it even wh...