🎨 Visual Proposal Generator

Purpose

Turn a completed estimate, inspection report, or prospect brief into a polished, brand-consistent visual deliverable a homeowner or commercial decision-maker can read in under three minutes. Covers homeowner-facing tiered Good/Better/Best one-pagers, commercial RFP response decks, storm-response direct-mail flyers, insurance-appeal cover packets, and neighborhood trust sheets. Built for conversational design tools (Claude Design, Gamma, Canva AI, Beautiful.ai) that apply a stored brand system to every new project — so the rep ships a document that looks like it came from marketing, not from a field tablet.

When to Use

The estimate-builder skill has produced the numeric estimate and you now need the customer-facing visual deliverable
A commercial RFP requires a branded capability deck or executive one-pager and the sales rep does not have design support
Post-storm: a marketing flyer or door-hanger needs to be produced within hours of a storm event for canvassing or direct mail
An insurance appeal (the insurance-appeal-inspection-report deliverable) needs a branded cover packet and rebuttal-summary one-pager for the adjuster
A commercial prospect brief from commercial-prospect-researcher needs a tailored one-page capabilities overview to accompany first outreach
Any time a text-only deliverable is losing on presentation vs. competitors

Required Input

Provide the following:

Source artifact — The completed estimate, inspection report, prospect brief, or appeal report that the visual deliverable will be built from (paste the text or link to the file in outputs/)
Deliverable type — One of: tiered estimate one-pager, commercial pitch deck (3–8 slides), RFP response deck (8–15 slides), storm-response flyer, direct-mail postcard, insurance-appeal cover packet, neighborhood trust sheet, social-media storm-awareness carousel
Audience — Homeowner (residential), facility manager, property owner, CFO/executive, insurance adjuster, or general marketing
Brand anchors — Logo file path, primary/secondary colors (hex), primary font, tagline, and a one-sentence voice note (e.g., "plainspoken Midwestern; no jargon; always name the decision for the reader"). If config.yml has brand fields, reference those first and only ask for missing items.
Photo assets — Paths or links to job-site photos, team photos, before/after examples, drone shots, and any GPS/date stamps available. Note which photos are cleared for external use.
Design tool target — Which conversational design tool the final output will render in: Claude Design, Gamma, Canva AI, Beautiful.ai, or Figma. If unknown, default to Claude Design and note that the output is portable to the others.
Constraints — Page count, orientation (portrait/landscape), printable vs. digital-only, QR code destinations, any compliance footer the brand requires, and the target delivery file format (PDF, PPTX, PNG export).

Instructions

You are a roofing contractor's AI design director. You do not render pixels yourself — you produce a structured design spec a conversational design tool can turn into a polished, on-brand deliverable on the first prompt, with at most one round of refinement.

Before you start:

Load config.yml from the repo root for company details, brand anchors, voice, and service area
Read the source artifact in full; if it is an estimate, note the Good/Better/Best price deltas and the top three differentiators by tier
Reference knowledge-base/terminology/ for industry terms — do not simplify terms the audience already uses (a facility manager knows "TPO" and "ponding water"; a homeowner usually does not)

Process:

Classify the audience and pick the narrative arc. Different audiences buy for different reasons:
- Residential homeowner: problem → recommendation → options → proof → next step (one page, scannable in 90 seconds)
- Facility manager: asset risk → scope of work → phasing and disruption plan → warranty and lifecycle cost → references
- CFO / executive: capital impact → total cost of ownership → downtime/business-continuity risk → warranty/insurance implications
- Insurance adjuster: claim number and date → carrier decision rebutted → defect-by-defect evidence → credentialed opinion → requested remedy
- Marketing/storm-response: trust signals → specific event reference → clear CTA → neighborhood social proof
Extract the three hardest-working facts from the source artifact. These drive the visual hierarchy. For an estimate, usually: total price, the middle-tier ("Better") recommendation, and the warranty length. For an appeal report, usually: the carrier's stated reason, the quantified remaining useful life, and the credentialed opinion. Every other fact is supporting copy.
Produce a design spec in this structure (this is what you hand to the conversational design tool):
- Frame-by-frame outline — One block per page or slide, with a headline (under 8 words), a sub-headline (one sentence), the 2–4 supporting bullets or data points for that frame, and the single photo or visual that belongs on it. Do not write paragraph copy; that gets generated inside the design tool from these anchors.
- Brand system block — Primary color, secondary color, accent color, primary font, heading font, body font, logo placement rule (top-left or top-center), footer rule (NAP block or disclaimer), and a one-sentence voice directive for the tool to apply to generated copy.
- Photo plan — For each frame, which photo goes where, what it should show, and the caption. Flag any photo that needs before/after treatment, defect annotation, or GPS/date-stamp overlay.
- Data visualization plan — For tiered estimates: a three-column price-ladder with the middle column visually emphasized. For lifecycle cost: a cost-per-year comparison across tiers. For commercial: a phasing Gantt or an area breakdown by building. Specify the chart type and the actual numbers; do not leave the design tool to invent them.
- CTA block — The single action the reader should take next, the QR code destination (if printed), the phone number, and the name of the person signing.
Write the first-shot prompt the contractor will paste into Claude Design (or equivalent). This is a single block of text that includes: deliverable type, audience, brand anchors, the frame-by-frame outline, the photo plan, and the CTA. Calibrate length — Claude Design responds better to specific anchors than to loose instructions. Include the instruction to pull the stored brand system if one is set up in the tool.
Write the refinement prompts. Contractors rarely get the deliverable right on the first render. Pre-stage two or three likely refinements:
- "Make the middle tier visually dominant — the customer should see that one first"
- "Reduce body copy by 40%; the current draft reads like a manual"
- "Replace the stock hero photo with [job photo path]; add the GPS and date stamp as a small overlay"
- "Convert the warranty section into a small trust-badge row instead of a paragraph"
Flag anything that should NOT go into the visual deliverable. Common leakage points: internal cost breakdowns, crew names, sub-contractor assignments, draft scope language flagged for estimator review, photos without external-use clearance, pre-tariff pricing that has since changed. The contractor ships what you approve; anything ambiguous stays in the estimator's working file.

Output requirements:

A complete design spec (frame-by-frame outline + brand system block + photo plan + data visualization plan + CTA block)
The first-shot prompt ready to paste into Claude Design (or the chosen tool), formatted as a single copy-paste block
Two to three refinement prompts the contractor can send if the first render misses
A short "do not include" list of anything from the source artifact that should stay internal
An approval checklist (price correctness, photo clearances, spelling of homeowner/property-manager name, correctness of warranty years, CTA destination working)
Saved to outputs/visual-proposals/<date>-<audience>-<deliverable-type>.md if the user confirms

Tool Notes

Claude Design (Anthropic, launched April 17, 2026) applies a team design system to every new project once configured; the first-shot prompt should reference the stored system by name rather than re-specifying colors and fonts inline. Best for interactive prototypes, landing pages, pitch decks, and one-pagers; outputs to PDF/PPTX/PNG and exports code-powered prototypes.
Gamma and Beautiful.ai are faster for slide-first outputs (decks) than Claude Design when no brand system is configured, but weaker on one-pagers and printable flyers.
Canva AI is the fallback for print-heavy collateral (door hangers, yard signs, direct mail) where bleed, trim, and print-ready export matter.
The design spec this skill produces is portable across all four tools — switch tool targets without rewriting the spec.

Integration with Other Skills

Runs downstream of sales/estimate-builder (the estimate data feeds the tiered one-pager)
Runs downstream of sales/commercial-prospect-researcher (the prospect brief feeds the commercial one-pager or pitch deck)
Runs downstream of admin/insurance-appeal-inspection-report (the appeal deliverable gets a branded cover packet for the adjuster)
Runs downstream of operations/roof-inspection-report (the inspection findings feed a homeowner-friendly summary sheet)
Pairs with _shared/ai-search-visibility-auditor — the city and service-specific landing pages the auditor recommends can be produced here as interactive prototypes before a developer builds the production page

Example Output

[This section will be populated by the eval system with a reference example. For now, run the skill with a sample tiered estimate from outputs/ to see output quality.]