English | 简体中文
A library of real, sourced corporate failure case studies — usable with any AI tool, or with no AI tool at all.
Ask "does my company have any of the failure patterns that killed [Company X]?" and get an answer grounded in what actually happened — not generic business-book advice.
- Have an AI tool? Point it at this repo and ask about your company. See Platform support.
- Don't want to use AI? Just browse the cases yourself at dotfei.github.io/Learn-From-Failure.
- Want a quick self-check? Use CHECKLIST.md — no AI needed.
- Know a case that should be added? Suggest it here — no writing required.
Every case follows the same four-part structure:
- What happened — the facts, briefly.
- Root causes — the real reason, not just "they went bankrupt."
- Warning signs — what was visible before the collapse.
- Questions this raises — the same questions, rewritten so you can ask them about your company.
Cases are tagged by failure mechanism, not industry — fraud, governance, cash burn,
disruption denial, product-market fit, scaling too fast, culture — because most of these
patterns aren't specific to one industry. See the full tagged list in cases/_index.md.
Most "startup failure lessons" content is a listicle you skim once and forget. This is built to be used conversationally: describe your situation, and get pointed at the 3-6 cases that match the actual mechanism of your risk — plus the exact questions that, in hindsight, would have caught the problem at the real company.
You don't have to ask for it, either. If you're just talking through a normal business decision — cutting costs, expanding fast, a hiring plan — and it strongly matches a known failure pattern, the AI will flag it on its own: "Before that — [Company] did almost exactly this and here's what happened." You still get the warning even if checking against failure cases wasn't what you were asking about.
You: We're a 12-person SaaS startup. We just closed a seed round and plan to grow from 12 to 35 people in the next two months — before our renewal numbers next quarter come in. Our investors want to see "aggressive growth" in the deck. Is this a mistake?
Claude: This looks like
scaling_too_fast— closer to Webvan and Kenny Rogers Roasters than WeWork. Smaller scale, same core error: committing to fixed cost (headcount) before validating the metric that justifies it (renewal rate). Webvan built $1B of warehouses before proving the model worked in one city.A few questions from those cases, applied to you:
- If next quarter's renewals come in soft, can you still afford the 35-person team?
- Is the hiring pace set by your own customer data, or by what looks good in a board deck?
- Of the 23 new hires, how many are reversible (contractors) vs. locked-in fixed cost?
This isn't "don't hire" — it's about sequencing: validate the signal first, then hire.
Try it with your own situation, or read the full write-up in dist/learn-from-failure-bundle.md.
The content is plain Markdown, so there's no lock-in to one AI vendor:
| If you use... | Do this |
|---|---|
| Claude Code | Install as a plugin (.claude-plugin/plugin.json), or copy .claude/skills/learn-from-failure/ into your own project. |
| Cursor, Windsurf, OpenAI Codex CLI, Amp, Continue, or anything that reads AGENTS.md | Clone the repo and point your tool at it — AGENTS.md has the full instructions. |
| ChatGPT Custom GPT, Gemini Gem, or any tool with a "knowledge upload" | Upload dist/learn-from-failure-bundle.md — one file with everything, sources included. |
| A raw API call, or any chatbot with no file upload | Paste the contents of the same bundle file into your prompt or system message. |
| No AI tool at all | Browse dotfei.github.io/Learn-From-Failure — searchable, filterable, nothing to install. |
Everything above reads from the same files in cases/. Nothing is copy-pasted by hand, so the
different versions can't drift out of sync — the bundle and the website are both regenerated by
scripts that CI checks on every change.
- This is a reflection tool, not a prediction engine. Many companies show one or two of these same warning signs and turn out fine (survivorship bias is real). A match means "worth investigating" — not "you will fail."
- Check the sources. Every case has a
## Sourcessection. Verify a number before repeating it publicly. Found a mistake? Report it. - Don't paste in confidential company info. If you're analyzing your own company with an AI tool, keep that in your own private chat — don't commit it here. New cases should only cover companies with real public reporting (news, court filings, Wikipedia, etc.).
- This is commentary, not a legal claim of guilt. Allegations, indictments, and civil/ regulatory findings are described as such — not as criminal convictions — unless a court actually convicted someone. See LEGAL.md for the full explanation and how to request a correction.
CHECKLIST.md (简体中文) is a static list of every case's "Questions this raises," grouped by failure mechanism — good for a quarterly review without needing a conversation.
Three ways, from easiest to most involved:
- Just tell us. Suggest a case, add a source, or report an error — fill in a form, no writing or Markdown required.
- Edit on GitHub. Open any file in
cases/, click the ✏️ pencil icon, make your change, and submit. GitHub creates the pull request for you — no local setup needed. - Write a full case locally. See CONTRIBUTING.md for the format and required sections.
AGENTS.md — instructions for AI tools other than Claude Code
.claude/skills/ — the Claude Code Skill (delegates to AGENTS.md)
.cursor/rules/ — the Cursor rule (also delegates to AGENTS.md)
cases/
_index.md — tagged index of every case
_template.md — template for adding a new case
<company>.md — one file per case, each with sources
contrasts/ — cases where a company faced the same pressure and adapted well
scripts/ — regenerates the bundle and the website from cases/
dist/ — single-file bundle for tools with no file access
docs/ — the searchable website (GitHub Pages)
CHECKLIST.md — static self-audit checklist
CHANGELOG.md — what's changed, release by release
CODE_OF_CONDUCT.md — expectations for contributors and discussion
LEGAL.md — sourcing standards, trademarks, corrections
Case files are written in English so the ## Sources links map cleanly to the English-language
reporting they cite. That doesn't mean the conversation has to be in English — ask in Chinese
(or any language) and the analysis will come back translated. README.zh-CN.md and
CHECKLIST.zh-CN.md are ready-made Chinese references if you'd rather read them directly.
MIT — see LICENSE. Case write-ups are original analysis of publicly reported facts, not a copy of any single source; check the linked sources for authoritative detail. See LEGAL.md for how this project handles trademarks, allegations vs. convictions, and correction requests.