# ask-yc
[](https://skills.sh/HKTITAN/ask-yc)
**Startup advice from the Y Combinator canon — with who said it, when, and where.**
Ask a startup question and get an answer grounded in 203 articles compiled from YC's
library and Sam Altman's blog, every claim attributed to a named speaker and cited to a
dated source. Not a generic composite of startup advice: Kevin Hale's pricing
thermometer, Tom Blomfield's value equation with the worked numbers, Paul Graham's
default-alive test, and the places where they disagree with each other.
The corpus is bundled. Nothing to clone, no API key, no setup.
## Install
```bash
npx skills add HKTITAN/ask-yc
```
Install for a specific agent, or globally:
```bash
npx skills add HKTITAN/ask-yc -a claude-code -g
```
Try it without installing:
```bash
npx skills use HKTITAN/ask-yc | claude
```
Works with Claude Code, Codex, Cursor, OpenCode, and 70+ other agents.
## Use
It fires on startup questions automatically. Four modes, chosen from what you send.
**Ask** — a question, or your situation with numbers.
```
We're 4 months post-launch, growing 8% week over week, 5 months of runway.
Should we raise now or push for profitability?
```
It runs Graham's default-alive test on *your* numbers, puts your 8% against his 5–7%
good / 10% exceptional / 1% "you have not figured it out" benchmarks, and cites every
one of them.
**Office hours** — it questions you instead of answering.
```
I'm thinking about pivoting. Don't tell me what to do — ask me the questions
a YC partner would ask.
```
Two to four questions, each answerable with a number, a date, or a name. It also names
your evasions: answer a question about last week with a plan for next month and it will
say so.
**Review** — it grades an artifact Pass / Fail / Can't-tell against named rubrics.
```
Here's my seed deck outline: [paste]. Roast it.
```
Hale's Legible-Simple-Obvious for the design, and separately the content: a waitlist is
a form, free pilots are not traction, press is not traction, a top-down analyst TAM
fails, and a VP of Sales hire pre-PMF is a Fail.
**Briefing** — everything the corpus has, organized by claim.
```
Everything YC says about fundraising. Where's the consensus, where do they disagree?
```
## Why attribution instead of a summary
- **Positions change.** Graham says start a startup in college in 2007 and "do not start
a startup in college" in 2014. An undated answer hides the reversal.
- **The canon is split.** Parker Conrad: "If it's not working, give up immediately." The
rest of the pivot canon says persist. Six live conflicts are catalogued, and the skill
names both sides rather than picking one.
- **Advice inverts by stage.** "Go crazy, hire away" and "do not hire" are both correct
YC advice, on opposite sides of product-market fit. Six such gates are checked before
answering.
- **Gaps are named, not filled.** When the corpus does not cover something, the skill
greps 222 catalogued-but-unabsorbed YC sources and points you at the one that does,
instead of quietly falling back to training data.
## What's inside
```
SKILL.md mode router, retrieval loop, gates, citation rules, corpus defects
references/
map.md 203 articles: title, path, source count, word count, aliases
office-hours.md question banks by situation, evasion table
rubrics.md review procedure + criteria with thresholds, by artifact type
disagreements.md live conflicts, both sides named, plus three that only look like conflicts
anti-advice.md what the corpus argues against, indexed by plan not topic
sources.md all 428 sources — 206 absorbed, 222 catalogued but not absorbed
further-reading.md verified destinations for what the corpus lacks
build-map.py regenerates map.md from corpus frontmatter
build-sources.py regenerates sources.md from the manifest
wiki/ the corpus
topics/ (99) speakers/ (57) case-studies/ (32) frameworks/ (15)
```
`map.md` is the router, meant to be grepped rather than read — a hit returns the row,
and the row contains the path. It exists because `[[wikilinks]]` in the corpus cannot be
resolved to filenames reliably: 37 of 203 titles do not slugify predictably
(`[[The Midwit Meme]]` → `frameworks/midwit-meme.md`).
## Refreshing the corpus
Replace `references/wiki/` with a newer copy of upstream's `wiki/` directory and
`references/_sources.json` with upstream's `raw/_sources.json`, then:
```bash
python references/build-map.py > references/map.md
python references/build-sources.py > references/sources.md
```
Both preserve their hand-authored headers and reproduce their output byte-for-byte on an
unchanged corpus, so a non-empty diff means the corpus actually changed. No article can
silently drop out of reach.
## Notes
- `raw/` (37MB of transcripts, 25MB of it images) is deliberately excluded. The articles
carry the full substance; source URLs are provenance. The skill is told never to read raw.
- The 67 image references were rewritten to `raw.githubusercontent.com` URLs so they
resolve without the image directory.
- Corpus defects the skill routes around, all verified against the files:
`founder-mindset.md` and `startup-ideas.md` each carry two References blocks with
colliding footnote numbers; `seed-fundraising.md` uses 15 `[NEW]` markers instead of
numbers; `startup-metrics.md` lists two sources twice and misroutes D'Angelo's material
to a Blomfield reference; two speaker pages open with leaked build-process text; 147
reference lines read `(n.d.)`; two wikilinks are dead; one cited YC URL 404s upstream.
## Credit
Corpus: [VishalRohra/founder-wiki](https://github.com/VishalRohra/founder-wiki), which
implements [Karpathy's LLM-knowledge-base pattern](https://x.com/karpathy/status/2039805659525644595)
over the public YC canon. Source material is Y Combinator's library and Sam Altman's blog.
Skill packaging by [@HKTITAN](https://github.com/HKTITAN).
## License
MIT for the skill packaging. The underlying source material belongs to its authors.