business-canvas

name	business-canvas
description	Draw an Ash-Maurya Lean Canvas (10 boxes) for an idea by reading its `#NNNN-<slug>.md` files from `workflow/ideas/`, `workflow/approach/`, and `workflow/marketing/`, synthesizing each cell, then rendering a PNG into `docs/` via the `tldraw` MCP server. Use whenever the user asks to "draw the canvas", "render a lean canvas", "make a business canvas", "visualize idea

Render a Lean Canvas (Ash Maurya's 10-box variant — better than Osterwalder's 9-box for pre-launch ideas because it surfaces problem, unfair advantage, and key metrics explicitly) for one specific idea ID by pulling its evidence from the project's stage folders, building a tldraw shape document, and saving a PNG plus the editable source into docs/.

The canvas isn't art. It's a one-page compression of everything we already know. The point is to expose the weak cells — the ones with nothing to put in them — so the user can see what evidence is still missing.

Step 1 — Get the idea ID

Ask the user which idea to canvas. Accept any of: 0001, 1, #0001, #1. Normalize to a 4-digit zero-padded form like 0001. If they don't give you one and you can't infer it from the conversation, ask — don't pick one.

Step 2 — Find and read context

Look in each of these folders for a file matching #<NNNN>-*.md:

• workflow/ideas/#NNNN-*.md — problem, user, pain, alternatives, "why you" (from validate-idea)
• workflow/approach/#NNNN-*.md — unsolicited-demand evidence and verbatim complaints (from pain-hunt)
• workflow/marketing/#NNNN-*.md — channel pick, budget, success/kill criteria, named persona (from marketing-plan)

A glob like workflow/ideas/#0001-*.md is the way — slug isn't known in advance. The slug will be the same across all three folders for a given ID; capture it from whichever file you find first.

If a folder has no match, that's fine — proceed and warn the user once at the end ("Customer Segments cell is thin because no workflow/marketing/#NNNN-*.md exists yet — run marketing-plan to fill it"). Don't refuse; partial canvases still teach.

If a folder has multiple matches for the same #NNNN- prefix (shouldn't happen under the project's numbering rules, but be defensive), use the first one alphabetically and note it.

Step 3 — Synthesize the 10 cells

Lean Canvas cells map to the source files like this. Keep each cell to a short header line plus 2–4 punchy bullets — the canvas is read at a glance, not in paragraphs.

Cell	Pull from	What to write
Problem	workflow/ideas/ Problem + workflow/approach/ verbatim quotes	Top 1–3 problems, ideally in the sufferer's own words
Customer Segments	workflow/ideas/ User + workflow/marketing/ named persona	Who has the problem; mark early adopters distinctly
Unique Value Proposition	workflow/ideas/ core thesis + workflow/marketing/ positioning	One-sentence "single, clear, compelling" message — what + for whom + why different
Solution	workflow/ideas/ solution sketch + mvp/ if it exists	Top 3 features only. Resist listing the roadmap.
Channels	workflow/marketing/ "Single starting channel" + "Where they hang out"	The one channel that's been chosen; list the named venues
Cost Structure	workflow/marketing/ Budget + a short interview if it's thin	Early-stage spend only. Most idea-starter users are software engineers building solo, so build-time is effectively free until proven otherwise — don't list "engineering salary" or production AWS bills. List the actual money flowing out: ad spend, influencer fees, tool subscriptions (Buffer, Mailchimp, basic analytics), domain, landing-page service. Example: "$300/mo Meta ads test + $40/mo Buffer + $15/yr domain".
Revenue Streams	workflow/ideas/ willingness-to-pay + workflow/marketing/ pricing implications + interview if blank	Pricing model (one-time / subscription / freemium), price point ($5, $20, $99/mo — pick one), and rough LTV if guessable. Don't write "TBD" — push for a number you can both stress-test.
Key Metrics	workflow/marketing/ Success criteria	The 1–3 numbers that will say "this is working" by day 30 / 90
Unfair Advantage	workflow/ideas/ "Why you"	Something a competitor can't easily copy. If empty, write "⚠️ none yet — riskiest cell" rather than fluffing it.
Early Adopters	workflow/marketing/ named persona + workflow/approach/ complainants	Specific people/segments who feel the pain acutely. Names, handles, or roles — not "small business owners".

Interview before you give up on a cell

Most idea-starter users are software engineers building solo. That means Cost Structure, Revenue Streams, Key Metrics, and sometimes Channels won't be in the source files in numeric form — and the user is the engineer, so their build labor is effectively free. Don't just write "(no evidence yet)"; ask them. Loop on vague answers the same way validate-idea and marketing-plan do — keep asking until the number is concrete enough to put on a canvas.

Sample questions you can ask in any order, one at a time:

• Cost / budget split — "You're the engineer here, so let's treat your build time as free. What can you spend per month on getting users while you test this? Rough is fine — $100? $500? $2000?"
• Where the money goes — "Of that, how would you split it? Paid ads vs an influencer payment vs tool subscriptions (Buffer, Mailchimp, analytics) vs landing-page service?"
• One-time costs — "Anything you'll pay once instead of monthly? Domain (~$15/yr), logo, Stripe fees, ConvertKit annual?"
• Pricing — "Are you charging from day 1, or running free → paid later? What price point — $5, $20, $99/mo? Pick one to stress-test; we can change it later."
• Success metric — "What single number at day 30 tells you 'this is working'? Signups, paid conversions, replies to outreach, demo bookings? Real number, not 'good engagement'."

"$some money for ads" doesn't pass. "$300/mo for 8 weeks then re-evaluate" passes. Push for that level of specificity.

Typical early-stage budget tiers (for sanity-checking)

Use this to spot mismatches between the user's stated budget and the channel they picked in workflow/marketing/. If a mismatch exists, surface it in the final report (Step 7) — it's a more useful insight than a polished canvas.

Budget tier	What it can fund	Mismatch warnings
$0/mo (pure sweat)	Founder-led DMs, posts in subreddits/Discords, free tier of every tool	Picking "Meta ads" or "influencer" here is a mismatch — push them back toward community/founder-led
$50–300/mo	Tool stack (Buffer/Mailchimp/Plausible) + one tiny ad test ($50–100)	Not enough for an honest paid-ads test or any influencer fee
$500–1500/mo	Real ad test ($300–800), tool stack ($50–150), one nano/micro influencer ($100–500)	Enough to learn something; not enough to scale
$2000+/mo	Sustained ad spend, ongoing micro-influencer relationship, paid analytics	If the product can't fund $40+ CAC, this is still a money fire

When a cell is truly empty after the interview

If after asking the user still has no answer for a cell — they don't know their price, they don't know their metric — write "⚠️ unknown — open question" in that cell rather than fluffing it. Open questions on a Lean Canvas are the whole point; they're the user's homework. A fluffy guess is worse than an honest blank, because it hides the gap.

Step 4 — Learn the tldraw shape format

Call the tldraw MCP server's tldraw_read_me tool once. It returns the shape format reference (allowed shape types, the 13 named colors, fill modes, auto-conversion rules). Read it before building shapes.

Two constraints shape everything below:

• No hex colors. tldraw accepts only 13 named colors (orange, yellow, light-green, light-blue, red, grey, …) and four fills (none|semi|solid|pattern). So the canvas's "design system" is a fixed mapping onto those names — you can't drop in this product's #FAF3E7 / #C8791A tokens directly. The mapping below is chosen to read warm and light anyway.
• Standalone text shapes are unreliable in headless export. Their width is auto-measured before the web font loads, the measurement comes back ~0, and the text collapses into a tiny clipped box — so you silently lose your title or labels. Keep every piece of text inside a geo shape, which has a fixed w so text always wraps. The build script (Step 5) already does this; if you ever hand-edit shapes, hold that line.

Step 5 — Build the canvas (design system + build script)

A Lean Canvas that looks like a bare wireframe gets glanced at and closed. A little design — warm color, a clear header/body split, a flagged weak cell — makes the user actually read it and see the gaps, which is the entire job of this artifact. So render it through a small design system rather than freehand boxes.

Don't hand-build the ~30 shapes. Use the bundled script — it owns the coordinates, the color system, the header/body card split, and the all-text-inside-geo rule, so you spend your attention on the content of each cell (Step 3) instead of pixel math. You only feed it the bullets.

python .claude/skills/business-canvas/scripts/build_canvas.py <input.json> <out.tldr.json> [--theme color|mono]

input.json carries only what you synthesized — headers and layout are fixed:

{
  "id": "0001",
  "name": "Myanmar Serial Reader",
  "cells": {
    "problem":   {"body": "• one\n• two\n• three"},
    "solution":  {"body": "• ..."},
    "metrics":   {"body": "• ..."},
    "uvp":       {"body": "..."},
    "advantage": {"body": "..."},
    "channels":  {"body": "..."},
    "segments":  {"body": "..."},
    "cost":      {"body": "..."},
    "revenue":   {"body": "..."},
    "adopters":  {"body": "..."}
  }
}

A cell that is missing, blank, or marked "weak": true renders red (⚠ unknown — open question). Lean into that — an honest red cell is the user's homework and the most valuable thing on the page. Don't fluff a guess just to avoid the red.

The script writes the editable shapes JSON to <out.tldr.json> and prints the same JSON compactly to stdout — capture stdout and pass it straight to the render call in Step 6.

The design system ("warm-light Lantern Canvas")

This product reads in a warm light "lantern-dawn" register (paper + amber + literary serif), so the canvas matches: serif title and headers, sans body, soft fills, an amber accent. The script offers two themes:

• color (default) — region color carries meaning, so the business model is legible at a glance:

  Region	Cells	Color
  How it works	Problem, Solution, Key Metrics, Channels	light-blue
  The heart	Unique Value Proposition	orange (amber)
  Who it's for	Customer Segments, Unfair Advantage, Early Adopters	yellow
  The money	Cost Structure, Revenue Streams	light-green
  Weak / empty (overrides region)	any cell	red header + light-red body

• mono — every section in a single amber accent on white cards. Calmer, the most literal match to lantern-dawn. Less information (no region meaning); only weak cells still go red.

Render color by default. After it lands, tell the user the calmer mono version exists and re-render with --theme mono if they prefer it — this is the deliberate "user picks at render" step.

What the design system draws (so you can sanity-check or override)

• Cards, not cells. Each of the 10 boxes is a solid-filled header strip (serif, the cell name in caps) stacked on a semi-filled body card (sans, the bullets). The strip gives hierarchy without needing bold.
• Floating layout. Cells are inset ~7px and sit on a soft grey board, so visible gutters make it read as a designed dashboard, not a tiled grid.
• Decoration. A full-width serif title band (LEAN CANVAS · <name> · (#NNNN)), an amber accent rule beneath it, and a color-key legend of chips along the bottom that documents the palette.
• Readable sizes. The canvas is ~1560px wide, where tldraw size:"s" text is unreadable — body/headers are m, the title is l. Keep bullets short (Step 3 already enforces this) so m text fits the 286px columns.

The fixed grid the script lays out (tldraw px, ~1500×900 before the title band):

Cell	x	y	w	h
Problem	0	0	300	600
Solution	300	0	300	300
Key Metrics	300	300	300	300
Unique Value Proposition	600	0	300	600
Unfair Advantage	900	0	300	300
Channels	900	300	300	300
Customer Segments	1200	0	300	600
Cost Structure	0	600	750	300
Revenue Streams	750	600	450	300
Early Adopters	1200	600	300	300

Maurya's original tucks "Early Adopters" inside Customer Segments. Splitting it into its own bottom-right cell makes empty-vs-filled obvious at a glance — the point of the canvas for our flow. Keep this layout unless the user asks for the classic.

Step 6 — Render and save

1. Run the build script (Step 5), pointing its <out.tldr.json> at docs/#NNNN-<slug>-canvas.tldr.json (create docs/ if missing). That file is the editable source — the MCP takes shapes JSON directly, so the user can re-render or tweak it later.
2. Pass the JSON the script printed to create_tldraw_diagram:
   • shapes: the compact JSON string from stdout.
   • outputPath: an absolute path like <project-root>/docs/#NNNN-<slug>-canvas.png.
   • format: "png" by default — it embeds cleanly in slides and the pitch summary. Also offer/render "svg" when the canvas is text-dense or headed for print: vector text stays crisp at any zoom.

Step 7 — Report

Tell the user three things:

1. Where the PNG landed (docs/#NNNN-<slug>-canvas.png).
2. Which cells were thin or empty (the red ones), and which source file would fill them (e.g., "Channels is empty — run marketing-plan for #0001").
3. The single weakest cell — usually Unfair Advantage or Key Metrics for early-stage ideas — and the cheapest experiment to strengthen it.
4. A one-liner offering the alternatives: the calmer mono theme, or an svg for print/slides — re-render on request.

If a file with the same name already exists in docs/, append a numeric suffix (-canvas-v2.png, -v3, etc.) — canvases evolve and old versions are useful to keep for diffing.