The default is LLM-authored, screenshot-driven. A deterministic template can't capture "dark, bold, dramatic brand with a green accent and uppercase display type" — it reads colours by frequency and gets the mood wrong. So instead: a Python step assembles a brand kit, and you (the orchestrator) look at the brand and write the markup.

Execute-layer modules under common/runners/:

proposal/parse.py — offer text → structured skills.proposal.plan.v1 (items, prices, recomputed subtotal, mismatch + outlier flags).
proposal/brand.py — scrape brand tokens (accent / fonts / logo / name) + enrich each item with its real catalogue photo + description (og: tags).
proposal/kit.py — screenshot the brand site (headless Chrome), download the logo, write BRIEF.md bundling tokens + the item table + authoring rules.
proposal/render.py — the --quick deterministic themed template + HTML→PDF helper.

proposal-maker