v1.2.1

A gap left by 1.2.0, caught while writing the announcement for it.

1.2.0 changed how every challenge opens — from what's wrong with your idea to what's right in it, then the constraint that breaks it. But it left the summary line alone. So the single most quotable sentence in every teardown still led with an obituary:

"[Original framing] is dead on arrival; [reframed wedge] is a real business because [reason]."

That phrasing is exactly what prompted the tone work in the first place, which made the change cosmetic in the place it was most visible.

The surviving thesis now leads with what survives. The skill can still tell you an idea is dead — it says it once, in the verdict, rather than as the headline of the summary. The instruction is explicit about why: it's the line most likely to make someone stop reading their own plan, and the summary's job is to point at what to build.

The worked example was rewritten to match, on the same reasoning as 1.2.0 — an in-context example anchors behaviour harder than an instruction does.

The archived A/B outputs in docs/examples/ keep the old phrasing on purpose. They're records of what v1.1.0 and v1.2.0 actually produced, not live guidance.

Install

Download vc-teardown.skill below, then Customize → Skills → Add → Upload a skill. Requires a Pro, Max, Team or Enterprise plan with code execution enabled.

For Claude Code, use the symlink route in the README.

---

The tone change itself, and the A/B behind it, are documented in v1.2.0 and docs/tone-comparison.md.

Earlier releases remain published and downloadable: v1.2.0 · v1.1.0 · v1.0.1

v1.2.0

A tone change, A/B tested before shipping.

What changed

Challenges used to open on what's wrong with your idea. They now open on what's right in that part of your thinking, then the constraint that breaks it.

This is a persuasion change, not a softer one. Granting the strongest version of an idea before dismantling it leaves nowhere to retreat to — opening on the defect invites you to defend your framing instead of hearing the evidence.

Three guards keep it from becoming a compliment sandwich:

• The concession has to be earned. "There's real demand here" is flattery and everyone can feel it. If it can't be specific about what's right, it doesn't open with it.
• Fatal flaws get no concession at all. A manufactured "yes" in front of a fatal objection just tells you you're being managed.
• The sentence shape varies, so twelve challenges don't read as a template.

Why you can believe it hasn't gone soft

Both tones were run against the same two ideas, chosen to sit at opposite ends of the merit spectrum.

On a near-dead idea — a landlord-rating site in a market with 1,800 homes available nationwide — the two outputs were near-identical in warmth: 3 of 13 concession-led before, 3 of 12 after. The skip rule fired in plain sight: "This is the fatal objection and it has no good answer in the original framing." The honest scale verdict came back harder, not softer.

On an idea with real merit — retired tradespeople selling independent advice — concession-led openers went from roughly 0 of 12 to 7 of 13, and the best insight moved out of a footnote into a headline: "Independence is the actual product, and you haven't named it yet." Challenges with nothing to grant — gas regulation, professional indemnity — still opened cold.

Near-zero change where the idea doesn't deserve it. Large change where it does.

Judge it yourself: the full side-by-side, with every challenge opener from all four runs and the complete outputs. The reasoning behind the change is in #4.

Also in this release

• The worked example's twelve challenge openers were rewritten to match. Testing during 1.1.0 showed the example anchors behaviour harder than the instructions do, so leaving it alone would have quietly defeated the change.
• scripts/package.sh can build a renamed, version-suffixed variant bundle, so two tone variants can be installed side by side and compared.

Install

Download vc-teardown.skill below, then Customize → Skills → Add → Upload a skill. Requires a Pro, Max, Team or Enterprise plan with code execution enabled.

For Claude Code, use the symlink route in the README.

---

Earlier releases remain published and downloadable: v1.1.0 · v1.0.1

v1.1.0

A content release that turned into a correctness one. If you installed via git clone into your skills directory, it never worked — see Fixed below.

Added

Four new challenge surfaces, taking the library from 16 to 20:

• Timing — the "why now?" question, and the prior-attempts graveyard most founders can't name
• Platform dependency — separating price risk (survivable, modellable) from existence risk (not)
• AI-native risk — whether a product rides model improvement or gets absorbed by it, plus model commoditisation and accuracy economics
• Atoms and capital intensity — hardware, inventory and long-R&D arithmetic, closing a gap the contributing notes had admitted to

Three new moat patterns, taking the set from 9 to 12: embedded distribution, proprietary model or eval set (the AI-era counterpart to "generate the data"), and regulatory or license position. "Sequence the moat" renumbers to 12 to stay last as the capstone.

A worked example — references/example-teardown.md. One idea run end to end, with 14 footnoted sources rather than illustrative figures. The research inverted the verdict: the idea as briefed turned out to be a competitor's core product funded at ~$85M, the "nobody has this data" moat was false, and the city that motivated it had already cut review times from 15 weeks to 3. Which is the point — that's what a real teardown does.

A conversational close. Delivery now names the one or two things the analysis is least sure about, then offers three or four specific next steps: expand a named challenge, a PDF, a one-page executive summary, a deck outline, re-run the economics on your numbers, or a teardown of the reframe itself.

Fixed

The Claude Code install instruction never worked. Personal skills resolve at ~/.claude/skills/<name>/SKILL.md and discovery does not recurse, so cloning this repo's root put SKILL.md one directory too deep. There was no error — the skill simply never appeared. Use the symlink route in the README instead, which also makes git pull the update path.

The Claude apps instruction described a flow that doesn't exist. The real path is Customize → Skills → Add → Upload a skill. The README now also explains that this is a cloud upload stored against your account — which is why no local folder appears after installing that way — along with the plan requirements and the fact that the two install routes don't sync.

Changed

• Version metadata now ships in SKILL.md under metadata, deliberately not as a top-level version: key — that would break .skill packaging and claude.ai upload, which validate against the six-field Agent Skills spec.
• scripts/validate.py and scripts/package.sh make the bundle reproducible, with validation that fails the build rather than the upload.
• Guards so the worked example can't shorten output: one section at full depth as a density reference, explicit length markers on the rest, and the word target restated in SKILL.md. Verified end to end — 5,933 words with every section intact, from a run given no hint that length was being measured.
• The word range is explicitly not a target to iterate against. Testing found the model spending ten minutes in compression passes without converging; a stall is worse than a plan that runs long.

Install

Download vc-teardown.skill below, then Customize → Skills → Add → Upload a skill. Requires a Pro, Max, Team or Enterprise plan with code execution enabled.

For Claude Code, use the symlink route in the README.

---

v1.0.1 remains published and downloadable here.

v1.0.1

First public release of the vc-teardown Claude Skill.

Install: download vc-teardown.skill below, then click Save skill
on the file card in a Claude conversation.

Includes the skill definition plus three reference files — a 16-item
challenge library, 9 moat patterns, and the plan output template.