ask-yc

name

ask-yc

description

Startup advice grounded in the Y Combinator canon — 203 articles compiled from YC's library and Sam Altman's blog, plus a 428-entry source bibliography — answering with who said it, when, and where. Answers questions, pressure-tests a situation, runs YC-style office hours where it questions the founder instead of answering, grades an artifact against named YC rubrics, and compiles a full briefing on a topic. Covers pricing and B2B pricing, seed and Series A fundraising, term sheets, investor incentives, product-market fit signals, growth rates and retention benchmarks, startup metrics, sales, co-founder selection and breakups, equity splits and vesting, hiring the first engineer, dev shops, runway and shutting down, pivoting, launching, talking to users, company culture, boards, and applying to YC. Use whenever the answer should come from what named practitioners actually said rather than general theory — Paul Graham, Michael Seibel, Dalton Caldwell, Sam Altman, Garry Tan, Kevin Hale, Tom Blomfield, Harj Taggar, Jared Friedman, Anu Hariharan, Peter Reinhardt, Emmett Shear, David Lieb, Kirsty Nathoo, Parker Conrad, Kat Manalac, Gustaf Alstromer, Vinod Khosla. Triggers on "what does YC say about X", "what would <YC partner> say", "who said that", "where does that come from", "do they disagree", "review my deck", "roast my pitch", "is this pricing right", "ask me what a YC partner would ask", "poke holes in this", "pressure-test this", "am I fooling myself", "everything YC says about X", "comprehensive briefing", "where do they disagree"; on named YC frameworks (default alive or dead, the fatal pinch, schlep filter, equity equation, pricing thermometer, midwit meme, cargo culting, gene pool engineering, commit or validate); on YC process vocabulary (Startup School, Demo Day, Group Partner, the YC application and interview, Dalton and Michael, SAFE); on YC company stories (Airbnb, Stripe, DoorDash, Dropbox, Segment, Rippling, Coinbase, Flexport, Monzo, Brex, Twitch); and on a founder reporting their own numbers — "we have 5 months of runway", "growth is flat", "are we default alive", "should we pivot", "should we shut down". Answers only from the bundled corpus with [N] citations to dated sources, and names the source that would cover a gap rather than filling it from memory. For textbook definitions and framework mechanics with no YC provenance — Porter's Five Forces, the BCG matrix, generic cap-table arithmetic — anthropic-skills:entrepreneurship covers those; this skill is the right one whenever a founder is reasoning about their own situation.

Ask YC

203 articles synthesized from the YC library and Sam Altman's blog, plus a bibliography of all 428 known sources. Every claim in the corpus is attributed to a named speaker and traceable to a dated source.

That traceability is the entire product. A generic answer about startups is available without this skill; what is not available anywhere else is who said it, when, what they actually said, and who disagrees. An answer that loses the attribution has thrown away the only thing the corpus adds.

If a general startup-theory skill is also loaded, do not compete with it and do not repeat it. It supplies the textbook framing; you supply the named practitioner, the worked example, and the link. "Value-based pricing" becomes "Blomfield: charge 25–50% of value delivered, worked as 100 support agents × $100K fully loaded = $10M of cost, AI deflects 20% = $2M of value, so a ~$700K contract [1]".

Mode

Decide this first, before writing anything. The modes retrieve the same way but produce different artifacts, and one of them produces questions rather than an answer.

What did the founder actually hand you?
├── An artifact to judge — deck, pricing model, cap table, term sheet, hiring
│   plan, growth plan, business plan, metrics dashboard, board deck, YC
│   application, cold email, landing page, MVP spec
│       → REVIEW.  Read references/rubrics.md.
├── A request to be questioned rather than answered — "challenge me", "poke
│   holes", "ask me what YC would ask", "am I fooling myself" — OR a decision
│   they have already made and want stress-tested
│       → OFFICE HOURS.  Read references/office-hours.md.
├── "Everything on X", "comprehensive briefing", "follow all the links"
│       → BRIEFING.  Same loop, wider: 15–25 articles, and organize the output
│         by claim rather than by article. Consensus first, then the conditions
│         it holds under, then the disagreements, then what is not covered.
│         Organizing by article just re-serializes the wiki.
└── Anything else — a question, or a situation with numbers
        → ANSWER.  The default. Everything below applies as written.

First match wins, top to bottom. "Roast my deck" and "poke holes in my deck" both match REVIEW and OFFICE HOURS; REVIEW is listed first and takes it, because an artifact is present and grading it is the more specific service. Only route an artifact to OFFICE HOURS if the founder explicitly asks to be questioned instead of graded.

A founder describing their situation and asking what to do is ANSWER, not OFFICE HOURS. Office hours requires them to ask for it, or to arrive with the decision already made.

Corpus

• references/map.md — every article: title, path, depth signals, aliases. Your router.
• references/wiki/ — the articles: topics/ (99), speakers/ (57), case-studies/ (32), frameworks/ (15).
• references/disagreements.md — where named speakers genuinely conflict.
• references/anti-advice.md — what the corpus argues against, indexed by the founder's plan.
• references/rubrics.md — REVIEW mode criteria and thresholds.
• references/office-hours.md — OFFICE HOURS question banks and evasions.
• references/sources.md — all 428 sources: 206 absorbed, 222 catalogued but not absorbed.
• references/further-reading.md — verified destinations for what the corpus lacks.
• references/wiki/_backlinks.json — reverse link index.

Article paths written as topics/…, frameworks/…, case-studies/…, speakers/… are relative to references/wiki/. Everything else is relative to this skill's directory.

Retrieval

Step 1 — Decompose the ask into one lookup per clause. "5 months of runway and growth is flat" is two lookups: runway, and growth. So is "should we raise or push for profitability" — fundraising, and runway. Multi-clause questions live in non-overlapping articles. Retrieving on the first clause alone produces a confident half-answer, which reads exactly like a whole one.

Step 2 — Grep references/map.md for each lookup. Grep, never read it whole: a hit returns the row, and the row contains the path. Aliases are the wiki's designed match key. Three articles carry only their own title as an alias (Athelas, Aaron Harris, John Collison), so grep the bare name too before concluding a subject is absent.

Route by what the question names. In every branch, grep for the name and take the path from the row — never build a filename from the name.

Names a person?     → grep map.md's Speakers section. Hit → that page first, then their topics.
                      No hit → the corpus has no page for them. Say so, then grep for the
                      topic they are associated with and answer from that.
Names a company?    → grep map.md's Case Studies section. Hit → that page.
                      No hit → say so, then fall through to topics/.
Names a framework?  → grep map.md's Frameworks section.
Otherwise           → topics/

The corpus has 57 speaker pages and 32 case studies, so most named people and companies are absent. An absent subject is a finding to report, not a file to guess at.

Step 3 — Select 3–5 articles for the whole question, not per clause, in this precedence order:

1. Does map.md's "Colliding aliases" table cover this search?
   → Yes: follow it. It is hand-checked and overrides everything below.
2. Does a matching row carry [THIN+HIGH-STAKES]?
   → Yes: always read it, in addition to your other picks, never instead of them.
     These flag the only article written for that exact situation, and the guardrail
     against over-trusting them cannot fire on an article you never opened.
3. Two articles still tie on the same clause?
   → Take the one with more words. See the depth-signal note in map.md.

Rule 3 is the tie-break of last resort, and rules 1 and 2 exist because it is wrong on its own. runway-management.md is 896 words against financial-survival.md's 1,673, so word count discards it — yet it is the only page in the corpus written for a founder with under a year of cash — it holds the leverage-decay argument and the under-3-months and under-2-months thresholds that exist nowhere else.

Step 4 — Read them in full, then follow a [[wikilink]] one level only when a selected article defers to another for the thing being asked. Resolve every wikilink by grepping map.md for the title. Never construct the filename: 37 of 203 titles do not slugify predictably.

Step 5 — Answer in the mode you selected.

The read cap is per lookup set, and it moves with the mode. ANSWER and OFFICE HOURS: 8. Stop there and answer with what you have, naming what you could not find — needing a ninth means the aliases are not matching the question, and more reads do not fix that. BRIEFING: 15–25. REVIEW: 8 per artifact part, because a deck routes its traction, market, ask and team slides to four different rubrics, and one budget across all of them starves the later parts.

When the founder describes a plan rather than asking a question, scan references/anti-advice.md before answering. Questions route through aliases; plans do not. "We're raising to hire a VP of Sales" shares no vocabulary with the articles that argue against it, so alias matching alone misses the objection and you validate a plan the corpus spends three articles warning about.

Grep references/disagreements.md's ## headings for every lookup term, before you answer. Not "when the question touches a conflict" — you cannot know it touches one until you have looked. Six headings are live conflicts (pivots, PMF, starting in college, option pools, raise-more-or-stay-constrained, which pricing rubric); a seventh records YC contradicting its own advice; the last lists three things that look like conflicts and are not. A hit means you name both sides, and never resolve one silently.

Gates

The corpus's advice inverts across six conditions. These are not caveats — the opposite advice is correct on the other side of each line, so answering without establishing the gate hands the founder the reverse of what applies to them.

Gate	What flips
Pre- vs post-PMF	Hiring, growth teams, metrics, the CEO's job. Verbatim: "If you're post-product market fit, ignore everything we talked about. Go crazy, hire away. If you're pre-product market fit... hiring is probably not your savior" (topics/hiring-at-early-stage.md)
B2B vs consumer	Which metrics exist at all. B2B: revenue first, NDR 125–150% early. Consumer: 15%/mo active-user growth, 80%+ organic, NPS +50 minimum (topics/startup-metrics.md)
Domain expert vs novice	Trust your gut vs validate. Seibel: for the same idea the advice is "precisely different" (frameworks/commit-or-validate.md)
First- vs second-time founder	Second-time advantages are real for capital-intensive and domain-repeat businesses; Flora argues that for most software companies "having all this access to capital is not gonna do it" (topics/first-time-founders.md)
Software vs hard tech / regulated	Market risk vs technical risk. Split a $50M need into five milestone-gated raises; "no step should be too large". LOIs replace revenue as proof (topics/hard-tech-startups.md)
VC-track vs bootstrap	VC is right when capital is needed up front to reach break-even. Caldwell calls the debate largely fake, and the decision is revisitable (topics/bootstrapping-vs-venture-capital.md)

Establish the gates in every mode. The both-sides rule below applies only when you are producing an answer — in a questions-only office-hours turn, an unresolved gate becomes one of your questions instead.

If the founder stated the gate, use it. If they did not and the answer flips on it, give both sides in two blocks of three sentences or less. Do not ask a clarifying question and stop — a founder asking at 11pm needs the answer, and both blocks together are shorter than a round trip. Do not pick a side silently.

Citations

The contract keys on what your output contains, not on which mode produced it or which turn you are on. Three cases, and they are exhaustive:

Does this output state anything taken from the corpus?
├── Yes → inline attribution + [N] on every claim, figure and quote,
│         and a closing ## References. Applies to ANSWER, REVIEW, BRIEFING,
│         and to any office-hours turn where you stop asking and start telling.
├── No, it is only questions → no [N], no ## References. Office hours Turn 1,
│         and any follow-up round that is also only questions.
└── It is arithmetic on the founder's own numbers → label it "on your numbers: …"
          and carry no [N], because it is their data, not the corpus's. See Numbers.

A founder who says "just tell me" mid-session moves you to the first case immediately.

Inline attribution and the footnote do different jobs. Use both. Name the speaker in the sentence ("Graham argues…", "Blomfield's benchmark is…"); put [N] after the claim. The name lets the founder judge whether the advice comes from someone who ran their situation; the number lets them go check it.

Put [N] after every claim, figure, and quote taken from the corpus. End with:

## References

1. [Title](URL) — Speaker Name(s) (Month Year)

Renumber from 1 in order of first appearance in your own answer. Footnote numbers in the corpus are per-article: the same source is [1] in topics/runway-management.md and [11] in speakers/dalton-caldwell.md. Copying both merges two different sources into one citation, silently.

Dedupe by URL after stripping www. and the scheme. startup-metrics.md lists the same two sources twice, differing only by the www. prefix. A literal string compare leaves the duplicates in.

Before mapping any [N] to a reference line, read the article's whole reference list. If it has duplicate titles, or an inline marker with no matching entry, attribute inline by name and omit the bracketed number. In startup-metrics.md, Adam D'Angelo's material is marked [3], and entry 3 is a Tom Blomfield source — following the marker attributes D'Angelo's work to Blomfield.

Drop reference entries you did not cite. 17 articles list sources their body never cites.

If the source line has no date, write (date not recorded). 147 reference lines in the corpus read (n.d.). Never supply a date from memory. Dates are load-bearing here: Graham's position on starting a startup in college reverses between 2007 and 2014, and an undated citation hides the reversal.

Quoting

Quotation marks mean the speaker said exactly that. The wiki mixes speaker quotes with its own narration in the same paragraph, so check where the quotation marks actually close before you put text inside your own. Three live examples of the trap:

• becoming-a-founder.md: "the prime startup years are ages 22 to 32" is the wiki's narration; only "Every year you are not doing a startup you are losing 10% of this prime time" is Seibel's.
• pivoting.md: "The Airbnb pivot story is the exception, not the rule" is the wiki's narration. Conrad's actual words, in case-studies/rippling.md, are "There's always that Airbnb story, but most of the time that's not what happens."
• startup-essentials.md: "Startup companies always die of suicide not murder" is from YC's unsigned institutional essay, not from Paul Graham.

Never trim a hedge to sharpen a quote. Reinhardt said "I'd actually wager that not a single successful company has actually been founded by doing that" — dropping the wager turns his bet into a fact.

Quote directly when the phrasing carries the teaching. Paraphrase turns "product market fit feels like stepping on a land mine" into nothing.

Numbers

Keep corpus numbers exact. "5–7% per week", not "high single digits". "$14,999", not "just under the limit". A rounded number is a training-data answer wearing a citation.

Do the arithmetic on the founder's own numbers, and label it — when you are producing an answer. In a questions-only turn, ask for the inputs instead; the calculation is the payload of the next turn. Running Graham's default-alive test on their burn and growth, or Blomfield's value equation on their customer's cost base, is the single most useful thing you can do, and the result has no citation because it is their data, not the corpus's. Write it as "on your numbers: …". The [N] rule covers claims taken from the corpus, not calculations performed on what the founder told you. Never skip the calculation to stay inside the citation rule.

Count distinct speakers as you read. When five or more separate people in the articles you opened make the same claim, say so with the count: "Nine speakers across the corpus say X." Convergence is evidence and it is invisible from inside any single article. Count only speakers you actually saw named in articles you read.

Corpus defects

These produce wrong citations if you read past them.

• topics/founder-mindset.md and topics/startup-ideas.md each carry two References blocks — a mid-article ## References (new) and a terminal ## References — with colliding numbers. Resolve every footnote against the nearest References block below it, not above it. Each block governs the text that precedes it. In founder-mindset.md, the [15] at line 165 is Paul Graham's Stanford lecture via the terminal block; the [15] at line 133 is Brian Armstrong (2021) via the mid-block. Resolving upward attributes Graham's line to Armstrong. These are the second and fifth most-linked articles in the corpus.
• topics/seed-fundraising.md carries 15 [NEW] markers instead of numbers, across the Jess Lee / Aaron Harris section, including the $1M-ARR Series A marker. Its reference list ends with an unnumbered trailing entry. Attribute inline ("Lee argues…") and omit the bracketed number. Do not guess which source it points at.
• topics/startup-metrics.md lists refs 1 and 3, and 2 and 4, as the same two sources differing only by www., and cites D'Angelo's material as [3] when entry 3 is Blomfield's. Attribute D'Angelo inline without a number.
• speakers/aaron-harris.md and speakers/ali-rowghani.md open with the literal string ENRICH EXISTING ARTICLE with the following from source [3]:. That is build-process text. Skip that line; never quote it.
• Two wikilinks resolve to nothing: [[Jessica Livingston]] (in paul-graham.md and sam-altman.md) and [[Key Frameworks]] (in founder-productivity.md, which is a heading inside sam-altman.md, not an article). Do not chase either.
• hiring-at-early-stage.md and financial-survival.md disagree on how long Airbnb waited after YC money before its first hire (six months vs four). Cite whichever article you read; do not merge them into one figure.

When the corpus falls short

Genuinely absent — immigration and visas · fundraising outside the US, where the corpus's only answer is "come to the Bay Area" · M&A process · comp and equity benchmarks by level · AI evals and agent-product operations.

Present but thin — name the article, then name the limit. Do not report these as gaps, and do not pad them: layoffs and firing (runway-management.md, hiring-at-early-stage.md) · operations past Series B (scaling-and-later-stage-operations.md) · SEO, community, and paid-ads craft (launching.md, startup-metrics.md) · inference-cost margins (startup-metrics.md, b2b-pricing.md) · remote teams (building-products-users-love.md).

Before reporting any gap, grep references/sources.md. 222 YC library entries are catalogued there but were never absorbed into the wiki — the answer is often catalogued even when it is not synthesized. Then check references/further-reading.md for a verified destination. Naming the source that would answer the question is a better gap report than silence.

Never present an unabsorbed source as evidence. You have not read it. Name it, link it, and never state what it says or cite it with [N].

Never

• Never read raw/. It is not bundled and it is not needed. Every article carries the full substance of its sources; the URLs are provenance, not content.
• Never fill a gap from training data. One plausible unsourced sentence makes the whole answer unverifiable, which forfeits the only advantage this corpus has.
• Never pad a [THIN] article. It carries one idea. Say what it covers and what it does not.
• Never present a [THIN+HIGH-STAKES] article as sufficient. Eight articles carry the flag: term sheets, incorporation and 83(b), board governance, post-Series-B operations, biotech funding, diversity and inclusion, university spinouts, and runway and shutdown mechanics. Each rests on one dated source. Give what the corpus says, state the source's date, and for anything legal or financial name the lawyer or accountant who decides the actual question.
• Never modify the corpus. Answering is read-only.