HEAT · Climate & Cooling Engineering · Proposal Studio Learning session · 2026-06-01 · worked on the Olympus DC/UHI RFI

NotebookLM × Claude Code — the Grounded Proposal Workflow

From a first draft to a HEAT-designed, evidence-grounded proposal — one terminal, six moves.
Claude Code — orchestrates & renders NotebookLM — grounds & selects evidence HEAT design skills — draw the figures Proposal (.docx) — the deliverable
01 · Why this workflow

Manual drafting vs. a grounded pipeline

MANUAL DRAFTING GROUNDED WORKFLOW Evidence trail Evidence trail Figure quality Figure quality Re-use speed Re-use speed scattered · generic stock visuals one source-of-truth HEAT-designed figures upgrade

Same expertise, but every claim and every figure becomes traceable, on-brand and reusable across the next tender.

Take-away: grounding is what turns a good draft into a defensible one.
02 · The pipeline

Six moves, one terminal

1 · Draft /proposal-write → first-draft.md 2 · Inject draft + ToR as sources 3 · Ground NotebookLM Q&A + gaps 4 · Bridge Claude Code via MCP 5 · Render HEAT design skills → figures 6 · Integrate into .docx + QA → ship ↑ the MCP bridge is the new piece

Steps 1–3 establish a grounded source-of-truth; steps 4–6 turn evidence into HEAT-branded figures inside the document.

Take-away: it all runs from one Claude Code session — no copy-paste between tabs.
03 · Step 1

First draft from a skill

RFI / ToR + KB evidence 3 pillars · refs · experts · donor stack /proposal-write (tender mode) USP first · section-critic loop first-draft.md methodology · workplan · team figures-tailored / (placeholders) marked “to render in step 5” inputs skill output deferred

The right HEAT skill (here /proposal-write) produces a structured Markdown draft; figures are flagged but not yet drawn.

Take-away: draft in Markdown — it is what NotebookLM and the design skills both read.
04 · Steps 2–3

Ground it in NotebookLM

NotebookLM grounding add_source draft.md + ToR ask_question cited answers select what matters claims · gaps · key numbers audio team briefing

NotebookLM answers from your sources only, surfaces gaps, and tells you which evidence matters — it does not draw the charts.

Take-away: grounding chooses the figures; it never invents them.
05 · Steps 4–5

Bridge & render in HEAT design

selected evidence “ΔT vs distance” Claude Code terminal · MCP picks renderer HEAT design skills ish · paper-banana · geojson-map HEAT-branded figure 0m 10km ΔT decays with distance · 95% CI

Claude Code reads the grounded evidence and calls the right HEAT skill to draw the figure on-brand — Open Sans, HEAT palette, contained labels.

Take-away: the figure is born in HEAT design — no restyling later.
06 · Step 6 + roles

Integrate, QA, ship

figures into draft.md DOCX build _build_v2.py QA nano-banana check ship Who owns what Claude Code — orchestration, bridge, render, build NotebookLM — grounding, citations, gap-finding Design skills — ish / paper-banana / geojson-map Claudia (EIA lead) — steers evidence, reviews fit

Figures drop into the Markdown, the DOCX rebuilds, a visual QA pass checks fidelity, and the proposal ships — fully grounded.

Take-away: the same six moves replay on the next tender — this is a reusable HEAT capability.