market-teardown: research in, a sourced verdict out

Most market write-ups survey a space and end with "so what?". market-teardown works the other way round. It decides the verdict first, then makes every section build toward it, and every number has to trace back to research you have already done.

Ask Claude for "teardown of on-device inference economics" or "turn my notes on vertical SaaS in construction into an article". You don't need to describe the format; the skill fires on the request alone.

Research first, never invented

The skill reads your project files, uploads and earlier chats before it writes a word. It searches with the topic's own nouns and numbers, not with words like "research". A piece this long needs 15–25 sourced figures. If it finds fewer than about ten, it stops and tells you, rather than padding the piece with guesses.

Every figure carries its date, such as (FY26) or (Q1 2026). Each comparison uses one currency. Your own estimates are labelled "Author's model, directional only".

“Two things ruin a teardown: made-up statistics and a hedged conclusion. Most of the skill exists to prevent those two.”

Verdict first

Before drafting, the skill commits to exactly one verdict:

• No-go names the constraint that breaks the model.
• Conditional go names the condition and the number attached to it.
• Go is rare, and even then it names the risk most likely to kill the idea.

The four middle sections each do a fixed job, whatever their headings:

• What's actually real: the demand, sized and dated. It also says where no published figure exists, because that gap is often the finding.
• Why the obvious version fails: the structural blocker, or the companies that already tried it.
• The economics, and who pays: unit economics and the named counterparty whose budget the money comes from.
• What would have to be true: the conditions, and who owns each risk.

The companion post holds something back

The short post exists to get the teardown read, not to summarise it. It opens with the hardest collision in under 140 characters, for example "$110M raised. $11M sold." It then uses two or three figures to set up the contradiction and says exactly what the teardown answers: "The verdict, and the four conditions attached to it, are in the piece." It never states the verdict.

Guardrails

• Criticise business models, not people. Published financials and shutdowns are fair game. Character judgments about named founders are not.
• Admit uncertainty; don't hedge. "No published figure exists, and that gap is why nobody owns this" is a finding. "This could be interesting, but there are challenges" says nothing.
• Your own plans stay private. The skill writes about something you are building only when you ask, and only at the level of detail you set.

See it in use

The repo includes a full worked example: a teardown of mid-term furnished rentals in India, with around 30 dated figures, 15 references and a Conditional go verdict with four conditions. Read it in references/example-teardown.md.

To adapt the skill, tell it where your research lives (a Claude Project, a folder of PDFs, an Obsidian vault over MCP). You can also loosen the formatting rules if your platform renders full Markdown, including tables.