The rules every piece of content in this course must follow — human-written or AI-generated, mission or calculator hint. If content violates this policy, it does not ship.
Last reviewed: 2026-07. See the Content Refresh Cadence doc for the review schedule.
Quality bar: tone and structure
From the PRD, non-negotiable:
- Second person, active voice. "You spent $500 on ads," not "A founder might spend…"
- Start with a challenge, not an explanation. Every mission opens cold with a scenario or problem the learner must react to before any teaching happens.
- Clear, concise, slightly irreverent. Zero academic fluff. No corporate jargon, no throat-clearing introductions, no "in this lesson we will…"
- Short paragraphs (3–4 sentences) and bullet lists where they help. Micro-lessons stay under ~2 minutes; no monologue runs past ~90 seconds.
- Realistic numbers. Examples use plausible figures. Real companies may be cited when the facts are public and sourced; otherwise anonymize or clearly fictionalize ("a sock-store startup"), and never present a fictional company as real.
- Every mission ends in an artifact. Content that doesn't make the learner do something is not mission content.
Sourcing and citation policy
- No invented facts, statistics, or citations. Ever. If a claim can't be traced to a source we can name, it is rewritten as an opinion, a heuristic, or a fictional example — or cut.
- Legal and finance terms anchor to primary sources. Definitions of terms like SAFE, liquidation preference, option pool, or Series A must match a primary source — Y Combinator's published SAFE documents, SEC glossaries and filings, or the definitive document for the instrument itself. A blog paraphrase is not a primary source for a legal term.
- Cite where the learner can see it. When a mission leans on an external definition or claim, the citation appears in the content (footnote or inline link), not just in our internal notes.
- Benchmarks are labeled as benchmarks. Statements like "interactive content can boost retention by ~60%" originate in the PRD's market research summary; until we validate a primary source (a research-appendix task), such claims are attributed to that research context or omitted from learner-facing content — never presented as established fact with a fabricated citation.
- Numbers in scenarios are exempt, sources are not. A fictional simulator can say churn is 5%/month without a citation — that's a parameter, not a claim about the world. The moment content asserts something about the real world, the citation rule applies.
Licensing policy
- Seed material must be open. Initial question banks, scenarios, and examples are drawn only from public-domain or Creative Commons–licensed educational materials (e.g. MIT OpenCourseWare, the Y Combinator blog where its terms permit, openly licensed case studies and datasets). Adapt and attribute per the license.
- No copyrighted text without permission. We do not reproduce paywalled or all-rights-reserved material — including well-known proprietary case studies — without an explicit license.
- Attribution travels with the content. Adapted CC material keeps its attribution in the content metadata, so a future edit can't silently orphan it.
- Learner artifacts belong to learners. The platform stores and displays them, but we claim no ownership of a learner's venture brief, personas, or plans, and we protect them as user IP.
Update cycle
Governed in detail by the Content Refresh Cadence doc; the policy-level rules:
- Quarterly: plan and release new content plus a minor patch pass (fix errata, refresh stale examples).
- Annually: full curriculum review followed by a major relaunch.
- Continuously: community feedback and error reports are triaged as they arrive; factual errors in legal/finance content are fixed out-of-band, not held for the next quarter.
- Everything is timestamped. Every content file carries a last-reviewed date, and time-sensitive facts (prices, regulations, market figures) are flagged for priority review each cycle.
Community moderation guidelines
The community (forums, peer review, cohort channels) is part of the course, so it inherits the quality bar:
- Be useful or be quiet. Feedback on peers' artifacts must be specific and actionable. "This is bad" gets removed; "your CAC assumes free traffic — what channel is that?" is the standard.
- No harassment, hate speech, or personal attacks. Zero tolerance; removal and escalating sanctions (warning → suspension → ban).
- No spam or undisclosed self-promotion. Sharing your own product is fine in designated spaces; disguising promotion as advice is not.
- No confidential material. Don't post other people's confidential data, and don't post your own secrets you'd regret — the forum is not private.
- Not professional advice. Legal, tax, and investment discussion in the community is peer conversation, not counsel; moderators label or remove content that presents itself as professional advice.
- Moderation is layered. Automated filters catch spam and abuse patterns; human moderators handle judgment calls; learners can flag any post. Moderation decisions are logged, and repeated flags on course content itself are routed into the content-update queue.
AI-generation guardrails
AI drafts a large share of content. The guardrails:
- Ontology-grounded. Generation prompts reference the concept ontology — its nodes, definitions, and relationships — so AI content stays anchored to the curriculum's vetted concepts rather than free-associating. A generated mission about CAC must use the ontology's CAC definition and its declared relationships (acquisition channels, conversion rate).
- Retrieval before assertion. For factual domains (finance, legal, metrics definitions), generation is grounded in retrieved source material from the approved corpus, not the model's memory.
- Human review before publication. No AI-generated mission, simulator case, boss fight, or learner-facing explanation ships without a human editor checking: factual claims against sources, tone against the quality bar, formulas against the calculator engine, and citations against this policy.
- The citation rules apply with extra force. AI-fabricated sources are the failure mode this policy exists to prevent. Reviewers verify that every cited source exists and says what the content claims.
- Deterministic where it counts. Formulas, calculator specs, and simulator parameters are validated by automated tests (
pnpm validate/pnpm test), so a hallucinated variable name fails the build rather than reaching a learner. - Feedback loops close. When the AI tutor gives a wrong or unhelpful answer in production (flagged by learners or spot checks), the case is added to the tutor's evaluation set so regressions are caught.