| name | extract-startup-ideas |
| description | Read the CS183B lecture transcripts in course/transcripts/ and extract the startup ideas / company case studies discussed, rendered into course/notes/ using the template at template/index.html. |
You read the CS183B "How to Start a Startup" lecture transcripts and extract the startup ideas / company case studies that the speakers (Sam Altman, YC partners, founders of Airbnb, Stripe, Facebook, a16z, Peter Thiel, etc.) use as teaching examples.
Goal
Turn the 20 lectures into a clean, browsable atlas of startup case studies. Each card is one company, with five fixed fields, rendered into a polished HTML page using the shared template.
Use the HTML template in:
template/index.html
Do not invent a new layout. Only replace the top-level placeholders and repeat the case card block.
INPUT
Source transcripts:
course/transcripts/Lecture01-.md … course/transcripts/Lecture20-.md
Twenty files, ~150 KB total. Plain Markdown, one lecture per file.
If the user passes a scope argument when invoking the skill (e.g. "Airbnb only", "Lecture 4 only", "just the manual-work pattern"), treat it as a filter on which companies to extract. With no scope, extract every company that gets a real narrative (not just a name-drop).
OUTPUT FILE
Write the rendered HTML to:
course/notes/startup-ideas-{scope_slug}.html
Rules:
scope_slug: short lowercased slug describing the scope. For the default "all companies, all lectures" output, use cs183b-all. For a filter, derive a short kebab-case slug (e.g. "Airbnb only" → airbnb, "Lecture 4" → lecture-04, "manual work pattern" → manual-work).- Create the
course/notes/ directory if it doesn't exist.
Example: course/notes/startup-ideas-cs183b-all.html
Do NOT paste HTML into chat. The chat response should only say:
Created course/notes/{filename} — {N} companies across {L} lectures.
CARD SELECTION
A company qualifies as a card when the speaker uses it as a story — describing what the company did early, why it worked, or what lesson it teaches.
Include:
- Companies whose founding story, early scrappiness, or specific tactic is described
- Companies used to illustrate a named principle (e.g. "do things that don't scale")
- Companies whose later outcome the speaker explicitly references
Skip:
- Generic name-drops ("...like Google") with no narrative behind them
- Companies mentioned only in passing as competitors or examples of a category
- Companies that are only mentioned as part of a list
Merge across lectures: if Airbnb is discussed in Lecture 4 and Lecture 8, produce one card and list both lectures.
CARD FIELD RULES
Every card has exactly these five fields, in this order:
Company Name
The company name, no qualifiers ("Inc.", "the company", etc.). Use the canonical capitalization (Airbnb, not airbnb; Y Combinator, not YC, unless that's the speaker's actual usage).
Mentioned in
Comma-separated list of lecture numbers, e.g. Lecture 4, Lecture 8. Use lecture numbers, not file names. Order them ascending.
Context at the time
2–4 sentences. What the company looked like when the speaker is describing them — usually early, scrappy, pre-traction. Capture the specific weird / manual / unscalable thing they were doing or the skepticism they faced. Do NOT describe the present-day company here. Stay faithful to the transcript — don't fabricate metrics or quotes.
What happened later
1–2 sentences. The outcome the speaker references — IPO, acquisition, scale, current valuation, "still operating at scale", etc. CS183B was filmed Fall 2014, so if the lecture predates a major outcome and you have widely-known public knowledge to fill it in (Airbnb's IPO, WhatsApp acquisition, etc.), state what happened after. If neither the transcript nor widely-known public knowledge supports a claim, write Outcome not stated in lectures.
Pattern
One sentence. The generalizable startup lesson the speaker is using this company to illustrate. Should read like a quotable principle, e.g. "Great startup ideas often begin as narrow, weird, manual workflows." If the same company illustrates different lessons in different lectures, pick the dominant one.
EXTRACTION STRATEGY
Don't read all 20 transcripts cover-to-cover. Work in passes:
-
Sweep for company names with grep. Seed query:
grep -nE "\b(Airbnb|Stripe|Facebook|Google|Twitter|Pinterest|Dropbox|Uber|Reddit|Instagram|WhatsApp|PayPal|LinkedIn|YouTube|Snapchat|Tesla|SpaceX|Palantir|Amazon|Apple|Microsoft|Netflix|Square|Slack|GitHub|Y Combinator|Wufoo|Justin\.tv|Twitch|Homejoy|Heroku|MongoDB|Cloudera|Mixpanel|Optimizely|Box)\b" course/transcripts/*.md
Expand the regex as you spot new companies in passing.
-
For broad "extract everything" runs, delegate the inventory to the Explore agent: "Scan all 20 transcripts in course/transcripts/ and return every company that gets a narrative (not just a name-drop), with the lecture number(s) and a one-line snippet of context." Then read only the cited passages to write the cards.
-
Read tight passages, not whole files. Once grep tells you Lecture 4 line 287 mentions Airbnb, read a window around that line — not the whole lecture.
-
Order cards by first appearance across the lectures (Lecture 1's stories first), unless the user's scope implies a different order.
HTML RENDERING RULES
Use the template file exactly. Replace these top-level placeholders:
{{title}} — page <title> text, e.g. Startup Ideas from CS183B{{eyebrow}} — short label above the heading, e.g. Knowledge atlas · CS183B{{heading}} — h1, e.g. Startup Ideas from CS183B (can be the same as title, or a longer phrasing){{lede}} — 1–2 sentence subtitle that frames the doc, e.g. Every startup the lecturers used as a teaching example, with the original context, the outcome, and the pattern they were illustrating.{{section_title}} — section label above the list, e.g. Companies or Case studies{{card_count}} — integer count of cards, e.g. 27{{footer}} — short footer line, e.g. Compiled from Stanford CS183B (Fall 2014).{{cards}} — concatenated card HTML (see below)
For each company, render one card using this exact structure:
<li>
<article class="case">
<div class="case-header">
<div class="case-name-row">
<span class="case-num" aria-hidden="true"></span>
<h3 class="case-name">{{company_name}}</h3>
</div>
<span class="case-meta">{{mentioned_in}}</span>
</div>
<div class="case-grid">
<div class="case-block">
<div class="label">Context at the time</div>
<p>{{context_at_the_time}}</p>
</div>
<div class="case-block">
<div class="label">What happened later</div>
<p>{{what_happened_later}}</p>
</div>
<div class="case-block pattern">
<div class="label">Pattern</div>
<p>{{pattern}}</p>
</div>
</div>
</article>
</li>
Per-card field rules:
{{company_name}}: company name only, no quotes, no prefix{{mentioned_in}}: e.g. Lecture 4 · Lecture 8 (use · between entries — matches the mono-uppercase meta styling). Just Lecture 4 if only one.{{context_at_the_time}}, {{what_happened_later}}, {{pattern}}: plain prose, no quote marks, no leading label- The card number (01, 02, …) is generated automatically by CSS — do not write numbers into the HTML.
- Escape HTML-sensitive characters in all rendered text:
& → &, < → <, > → >. For apostrophes in names like "Lenny's", use ’ for typographic quality.
DON'T
- Don't add new fields to the card beyond the five above.
- Don't group cards by pattern (e.g. don't add intermediate
<section> headers). The pattern is per-card; the page is a flat list.
- Don't include source quotes or transcript snippets in the rendered HTML — the Pattern field is the takeaway, not the evidence.
- Don't write the HTML to chat. Only write it to the output file.
