| name | yc-apply |
| description | Guides a founder through the full Y Combinator batch application end-to-end. A 10-phase workflow that captures the live YC form, profiles the founders, stress-tests the idea via an embedded grill loop, runs a mandatory 5-agent parallel external research pass on the startup, drafts every form field with anti-pattern and accepted-example checks, produces founder-video bullet notes (no script), runs a final adversarial gate, generates paste-ready submission answers, unlocks an interview-prep simulator after invite, and supports reapplicant delta tracking and post-decision post-mortems. Writes a documented markdown trail under a user-chosen workspace. Use when a founder wants to prepare a YC batch application, build their founder video, drill mock YC interview questions, or reapply with delta evidence. Don't use for pitch-deck design unrelated to YC, generic startup advice without applying, or post-funding work. |
YC Apply
A founder-facing workflow that drives a Y Combinator batch application from blank workspace to paste-ready answers. The skill is the engine; the founder's data lives in a separate workspace they own and version. The body of this file is a dispatcher — load-bearing detail lives in references/. Inline content here is a tripwire, not the contract.
Bundled Path Rule
Two distinct roots are in play and must never be confused:
<skill-dir> — the directory that contains this SKILL.md (the bundled engine: references/, assets/, scripts/). Resolve it once at the start of a session (e.g. the absolute path of the directory holding this file).
<workspace> — the founder's external application directory (resolved per the Invocation section). All founder data is written here.
Every references/…, assets/…, and scripts/… path in this file is relative to <skill-dir>, never to <workspace> or the current working directory. Because the agent often runs shell commands with the CWD set to <workspace>, always expand bundled helpers to an absolute path before invoking them: write <skill-dir>/scripts/bootstrap-workspace.sh <workspace>, not bare scripts/bootstrap-workspace.sh. The same applies to reading bundled files: read <skill-dir>/references/partner-signals.md. When a row or step below says scripts/X or references/X, treat it as <skill-dir>/scripts/X / <skill-dir>/references/X.
Required Reading Router
Match the phase you are running to the row. Read the listed file in full before producing any output for that phase. The reference is load-bearing; the inline summaries in this SKILL.md are tripwires only.
| Phase / Trigger | MUST read |
|---|
| Any phase (background context on what YC partners look for) | references/partner-signals.md |
| Phase 2 idea-stress-test, Phase 4 drafting any high-stakes field | references/seven-pitch-questions.md |
| Phase 4 drafts of any field, Phase 6 final adversarial gate | references/anti-patterns.md |
| Phase 4 drafting (every high-stakes field) | references/accepted-examples.md |
| Phase 5 founder-video | references/video-spec.md |
| Phase 8 interview-prep | references/interview-playbook.md |
| Phase 0 detects reapplicant, Phase 4 "what changed" draft, Phase 9 post-mortem | references/reapplicant-playbook.md |
Reference Index
references/partner-signals.md — what YC partners explicitly look for in written answers, the founder video, and the interview. The "20-second skim" reader model, the Dalton story-shape heuristic, Sam Altman's four evaluation questions.
references/seven-pitch-questions.md — Michael Seibel's Seven Pitch Questions, Jared Friedman's Idea Quality Score, the Email Test. The pre-flight checklist every founder must pass before drafting.
references/anti-patterns.md — buzzword list, Catherine Cross's 10-item "never say" list, behavioral red flags. Used by scripts/buzzword-scan.sh and by every grill loop.
references/accepted-examples.md — verbatim accepted answers and videos: Anja Health W22, Embark Trucks W16, Hera (4-attempt case), Static EV (reapp case), Nalify S25, Razorpay. Reference rubric, never invent.
references/video-spec.md — the founder-video rules (≤60s, all founders on camera, bullet notes not script, splice OK), full accepted-video transcripts.
references/interview-playbook.md — 10-minute interview signals, inverted-pyramid rule, "I don't like your idea" adversarial probe, mock-interview format, demo-readiness, co-founder choreography.
references/reapplicant-playbook.md — delta-doc pattern, how to surface "what changed since last application," explicit feedback-engagement rule.
Invocation
Founder calls /yc-apply with optional flags. The skill autodetects state and resumes where it left off.
/yc-apply # autodetect: workspace + phase
/yc-apply --new # bootstrap fresh workspace
/yc-apply --workspace ~/apps/acme # explicit path override
/yc-apply --phase 4 # jump to phase
/yc-apply --resume ~/apps/acme # resume specific workspace
/yc-apply --force-submit # bypass Phase 6 gate (logs warning)
Workspace resolution order:
--workspace flag.
$YC_APPLY_WORKSPACE env var.
- Current working directory if it contains
00_meta.md.
- Prompt the founder to choose: bootstrap new, resume by path, or cancel.
Workspace Layout
The skill is the engine; the workspace is the artifact. The founder owns and may version the workspace. Bootstrap creates this exact tree.
<workspace>/
├── 00_meta.md # batch, deadline, company-slug, status
├── 01_form-spec.md # canonical form (operated on, never bundled)
├── 02_founder-profile.md # founders, cool-things, hardest-built
├── 03_idea-narrative.md # Seven Pitch Questions + Idea Quality Score + Email Test
├── 04_research/ # 5-agent parallel external research
│ ├── 01_founder-footprint.md
│ ├── 02_competitive-landscape.md
│ ├── 03_market-tam.md
│ ├── 04_regulatory-context.md
│ ├── 05_traction-signals.md
│ └── summary.md
├── 04_reapplicant-delta.md # only if Phase 0 marked reapplicant
├── 05_drafts/ # one .md per form field, versioned
├── 06_video/
│ ├── notes.md # bullet talking notes (always)
│ └── script.md # only if founder explicitly opts in
├── 06_gate.md # Phase 6 result: the 10 checks + RESULT: PASS|FORCE-SUBMIT
├── 07_interview/ # only after invite (Phase 8)
├── 08_final/ # paste-ready locked answers
├── 09_journal.md # milestone log
└── 10_post-mortem.md # only after rejection (Phase 9)
Phase Map
Ten phases, idempotent. The skill detects the current phase from workspace state via <skill-dir>/scripts/phase-status.sh and offers to resume there.
| # | Phase | Definition of done |
|---|
| 0 | bootstrap | Workspace exists; 00_meta.md has real batch/company_slug/target_deadline/reapplicant values (no {{}}); 01_form-spec.md captured. |
| 1 | founder-profile | Every founder has cool-things + hardest-thing-built logged; no {{}} left; named versioned section; grill loops journaled. |
| 2 | idea-stress-test | Seven Pitch Questions answered; numeric Idea Quality Score ≥6/10 avg w/ weakest dim addressed; Email Test logged. |
| 3 | external-research | All 5 slice files + 04_research/summary.md present; each slice has a Findings section and ≥1 real URL. |
| 4 | drafts | A draft per high-stakes field in 05_drafts/ (count ≥ [HIGH] fields); fields grilled; coding-agent session staged. |
| 5 | founder-video | 06_video/notes.md exists; estimated ≤60s; founders named with credentials. |
| 6 | pre-submit-gate | 06_gate.md written with all 10 boxes [x] and RESULT: PASS (or FORCE-SUBMIT); final adversarial grill logged. |
| 7 | submission-pack | Gate passed; 08_final/SUBMIT.md populated with paste-ready answers. |
| 8 | interview-prep | Unlocked only after operator confirms interview invite. Mock drills journaled. |
| 9 | post-mortem | Auto-prompted after rejection; lessons + next-batch reapplicant scaffolding written. |
Embedded Grill Loop
Phases 1, 2, 4 (high-stakes fields), and 6 all use the same relentless interview format. Run it as discrete rounds, one question at a time. Do not batch questions; do not move on until the round resolves.
Each round:
- Question — one sharp, specific question targeting the weakest part of the current answer.
- Founder answer — wait for it. Never answer on the founder's behalf.
- Critique — name what is vague, unverifiable, buzzword-laden, or unsupported (cross-check
references/anti-patterns.md).
- Recommendation — propose a concrete sharper answer so the founder can react to a draft, not a void.
- Decision — founder accepts, edits, or rejects. Record the resolved answer.
Stop conditions for a field/topic (all must hold):
- The first sentence passes the 20-second skim test (a reader reproduces the core claim from it alone).
<skill-dir>/scripts/buzzword-scan.sh returns 0 hits on the draft, or each hit is journaled with a rationale.
- Every numeric or named-customer claim is defensible if Googled live.
- No adjective-only founder claims remain.
Cap a single field at ~6 rounds; if it has not converged, journal the sticking point and move on rather than looping forever. Log a one-line outcome per field via <skill-dir>/scripts/journal-append.sh (e.g. "Phase 4 competitors field resolved in 4 rounds").
Procedures
Step 1: Resolve workspace and detect phase
- Resolve
<skill-dir> (the directory containing this SKILL.md) per the Bundled Path Rule. Parse flags. If --new, run <skill-dir>/scripts/bootstrap-workspace.sh <path> (mutating helper) and proceed to Phase 0.
- Otherwise, resolve workspace per the order above. If no candidate exists, ask the founder whether to bootstrap new, resume by path, or cancel — do not invent a default.
- Run
<skill-dir>/scripts/phase-status.sh <workspace> (read-only helper) and parse the current phase + completion checklist.
- Announce the detected phase to the founder. Offer to resume there or override via
--phase.
Step 2: Phase 0 — Bootstrap
- If
00_meta.md does not exist, run <skill-dir>/scripts/bootstrap-workspace.sh <workspace> (mutating helper). It copies the <skill-dir>/assets/template-*.md files into the workspace and seeds 09_journal.md.
- Ask the founder for batch target (e.g.,
S26, F26), company-slug, target submission date. Write real values (no {{PLACEHOLDER}} left) into 00_meta.md — phase-status.sh treats any remaining {{ as "not done".
- Ask: "Paste the live YC application form text from apply.ycombinator.com, OR confirm the bundled form spec for your batch is current." The live paste is the source of truth; the bundled spec is a fallback only. If confirmed, copy
<skill-dir>/assets/form-spec-summer-2026.md (or the matching batch spec) to <workspace>/01_form-spec.md. If pasted, write the pasted text instead.
- Ask: "Have you applied to YC before?" If yes, set
reapplicant: true in 00_meta.md, force paste of prior interview/rejection feedback, and create 04_reapplicant-delta.md. STOP. Read references/reapplicant-playbook.md in full before structuring the delta doc.
- Append milestone via
<skill-dir>/scripts/journal-append.sh <workspace> "Phase 0 bootstrap complete" (mutating helper).
Step 3: Phase 1 — Founder profile (embedded grill loop)
STOP. Read references/partner-signals.md in full before opening the grill loop. That file is the contract for what YC reads in founder bios — the inline tripwires below are not the source of truth.
Gist tripwires for what the grill must extract per founder:
- One specific verifiable achievement (artifact, not adjective). Reject "dedicated / passionate / driven."
- The single hardest intellectually difficult thing the founder has built. Cannot be the current startup itself.
- The non-resume hack story (if not in the live form, capture it anyway — useful for the narrative).
- Concrete role on the team + 2-3 line backstory.
Procedure:
- For each founder, run a relentless one-question-at-a-time interview. After every answer, ask the next sharpest question. Always recommend an answer per question (do not leave the founder to guess your direction). Mirror the
grill-me pattern in this workspace.
- Reject any answer with only adjectives. Reject any "cool thing" that is the founder's role at a job — push for shipped artifacts.
- Write the grilled output to
02_founder-profile.md, one section per founder, versioned (### Alex Doe — v3).
- Append milestone via
<skill-dir>/scripts/journal-append.sh.
Step 4: Phase 2 — Idea stress-test (embedded grill loop)
STOP. Read references/seven-pitch-questions.md in full before opening the grill loop. It contains the Seibel Seven Pitch Questions, Friedman's Idea Quality Score formula, and the Email Test gate. Do not paraphrase the framework — embed it.
Procedure:
- Walk the founder through Seibel's seven questions one at a time, in order: (1) what do you do? (2) market size — bottom-up preferred? (3) progress? (4) unique insight? (5) business model? (6) team? (7) what do you want?
- After each answer, run the grill: probe for vagueness, buzzwords, missing concrete nouns. Recommend a sharpened version.
- Score the idea on Friedman's four criteria (market size, founder-market fit, problem confidence, unique insight). If average <6 or any dim <4, force the founder to either strengthen that dim or write a paragraph naming the weakness.
- Run the Email Test: skill drafts a two-sentence pitch; founder must email it to a smart non-technical friend and paste the friend's paraphrase back. Skip is not allowed; if friend unavailable, founder pastes a written read-back of the pitch from a stranger they trust.
- Write to
03_idea-narrative.md.
Step 5: Phase 3 — External research (MANDATORY 5 parallel subagents)
This phase is gated. Drafts (Phase 4) cannot begin until Phase 3 completes. Skip is not allowed.
- Read each research prompt template
assets/research-*.md — one per slice (research-founder-footprint.md, research-competitive-landscape.md, research-market-tam.md, research-regulatory-context.md, research-traction-signals.md).
- Compose 5 slice prompts. Each prompt MUST embed:
- The slice question and primary scope.
- The exact target file path:
<workspace>/04_research/0N_<slug>.md.
- Forbidden write paths (anything outside
04_research/).
- The output schema (six sections: Findings, Verbatim Quotes, Signal Strength, Risks, Open Questions, Evidence URLs).
- Dispatch all five Agent tool calls in a SINGLE message (parallel by default in Claude Code). Use
subagent_type=general-purpose for each. Wait for every call to complete. If the harness exposes no parallel Agent facility, fall back to running the five sequentially — never skip a slice.
- Web-tool fallback: the slice prompts assume
WebSearch + WebFetch are available. If a subagent reports those tools are unavailable, it must still write its file using what it can find (founder-provided links, local context), mark the gaps under Open Questions, and never fabricate URLs. A slice with zero real URLs fails validation (next step) and is re-dispatched or flagged to the founder.
- Verify each output exists, contains all six sections, and cites at least one real URL (
<skill-dir>/scripts/phase-status.sh <workspace> confirms the Phase 3 invariant mechanically). Re-dispatch any slice that failed validation.
- Synthesize
04_research/summary.md parent-authored. Sections: Research Question, Slice Findings (one per agent), Convergences, Divergences, Risks & Open Questions, Recommended Next Steps for the application.
- Append milestone via
<skill-dir>/scripts/journal-append.sh.
Step 6: Phase 4 — Drafts (per-field grill loop for high-stakes fields)
STOP. Read references/accepted-examples.md and references/anti-patterns.md in full before opening any draft loop. Examples seed the founder's mental model; anti-patterns drive the rewrite gate. Inline tripwires below are not substitutes.
Gist tripwires the per-field loop must enforce:
- First sentence carries full informational load (Graham's 20-sec skim test).
- No buzzwords. No "platform" / "AI-powered" / "revolutionizing" / "leveraging" / "ecosystem" / "no competitors" / "in stealth" / "need funding to start" / "some users."
- No adjectives without verifiable artifacts.
- Inverted-pyramid structure: conclusion first, proof after.
High-stakes fields (run full grill loop on each):
- 50-character company description.
- "What is your company going to make?" (3-sentence template from
accepted-examples.md).
- Founder bios (one per founder).
- "Why did you pick this idea?" / domain expertise.
- "Who are your competitors? What do you understand that they don't?"
- "How will you make money?" with bottom-up TAM.
- "Other ideas you considered."
- (Reapplicant only) "What changed since last application?"
Low-stakes fields (skip grill, just validate):
- Yes/no boxes (cofounder search, revenue, users, fundraising, legal entity).
- Location (use the post-YC city format).
- Tech stack (require: include AI models + AI coding tools per the S26 form).
- Company URL, demo URL, login credentials.
Per-field procedure:
- Read field intent from
references/partner-signals.md + accepted examples from references/accepted-examples.md.
- Present the question to the founder verbatim from
01_form-spec.md.
- Surface 2-3 verbatim accepted examples from the matching field.
- Founder writes first draft.
- Run
<skill-dir>/scripts/buzzword-scan.sh <draft-file> (read-only helper). For every hit, present the bad → good rewrite pair from references/anti-patterns.md.
- Run the embedded grill loop: one adversarial question at a time, with a recommended answer; tighten the draft each round.
- When draft passes both the buzzword scan and the grill, save versioned in
05_drafts/<field>.md.
Coding-agent session sub-step (Summer 2026 experimental field):
- Ask the founder where coding-agent transcripts live (e.g.,
~/.claude/projects/, cursor exports). Optionally glob to list candidates with size + date.
- Score top 3 candidates on: technical complexity, ownership demonstration, problem-solving narrative, real product progress. Surface scores; founder picks one.
- Validate: ≤25 MB,
.md or .txt. If oversize, ask the founder to trim or pick another.
- Skill drafts a one-paragraph context blurb (what the problem was, what the transcript demonstrates). Stage selected file + blurb in
05_drafts/coding-session/.
Step 7: Phase 5 — Founder video
STOP. Read references/video-spec.md in full before producing notes or a script. It contains the ≤60s rule, the "all founders on camera" rule, the anti-script doctrine, and full accepted-video transcripts (Anja Health W22, Embark Trucks W16).
- Compose bullet talking notes per founder (3-5 bullets each: name + 2-3 credential anchors + one traction stat or insight). Total budget ≤60s.
- Show estimated speaking duration per bullet (~3 seconds per bullet by default).
- Write to
06_video/notes.md.
- If — and only if — the founder explicitly opts in by saying "yes, draft a script for me," produce
06_video/script.md with a strong header warning that reading verbatim is a YC anti-pattern (cite video-spec.md).
- Refuse to produce a script otherwise, even on indirect requests.
Step 8: Phase 6 — Pre-submit adversarial gate (10-check)
STOP. Read references/anti-patterns.md and references/partner-signals.md in full before opening the final grill. This phase is the last filter before paste-ready answers; it must combine the buzzword scan, the story compile, and the partner-style adversarial probes.
The 10-check gate. ALL must pass to unlock Phase 7. The result is recorded in a real artifact, <workspace>/06_gate.md, which phase-status.sh reads — a check is not "done" until it is recorded there.
Procedure:
- Run each check below against the workspace. Several have a deterministic helper: run
<skill-dir>/scripts/buzzword-scan.sh <workspace>/05_drafts for check 7 and re-run <skill-dir>/scripts/phase-status.sh <workspace> to confirm checks 1-6/9 mechanically.
- Write
<workspace>/06_gate.md containing the checklist below with each box marked [x] only when its evidence exists, plus a final line RESULT: PASS when all ten are [x]. Leave unchecked boxes [ ] for any that fail — phase-status.sh treats any remaining [ ] as "gate not passed".
- If the founder invokes
--force-submit with unmet checks, still write 06_gate.md but mark the final line RESULT: FORCE-SUBMIT, list the unmet checks explicitly, and append a warning to 09_journal.md via the journal helper.
[ ] 1. 01_form-spec.md exists and is current.
[ ] 2. 02_founder-profile.md complete: cool-things + hardest-built per founder.
[ ] 3. 03_idea-narrative.md: Idea Quality Score ≥6 avg AND weakest dim addressed in writing.
[ ] 4. Email Test paraphrase logged in 09_journal.md (with friend's paraphrase verbatim).
[ ] 5. 04_research/summary.md exists; 5 slice files present; each cites ≥1 real URL.
[ ] 6. Every high-stakes form field from 01_form-spec.md has a draft in 05_drafts/.
[ ] 7. scripts/buzzword-scan.sh returns 0 hits across 05_drafts/ OR every hit has a journaled rationale.
[ ] 8. Dalton-story compile passes: skill writes "This is the story of N founders, who X, working on Y for Z, currently at W, because V." Founder confirms the story is accurate from the drafts alone.
[ ] 9. 06_video/notes.md exists; ≤60s estimate; all founders named with credentials; demo content absent.
[ ] 10. Phase 6 final adversarial grill logged in 09_journal.md. Drills must include: "I don't like your idea, what else have you got?" probe response; "What's the single biggest current risk?" probe response; and one numeric claim fact-check.
RESULT: PASS|FORCE-SUBMIT
Final adversarial grill: walk the founder through Sam Altman's four-question filter (see references/partner-signals.md) and the "I don't like your idea" probe. Score the responses for formidable-but-not-defensive posture. If any answer is robotic, scripted, or denies all challenges, force a rewrite.
Step 9: Phase 7 — Submission pack
- Refuse to enter Phase 7 unless
<workspace>/06_gate.md exists with RESULT: PASS (or RESULT: FORCE-SUBMIT). Confirm via <skill-dir>/scripts/phase-status.sh <workspace> showing Phase 6 checked. If the gate is unmet, return to Phase 6.
- Read every
05_drafts/<field>.md. For each field, copy the final version (last ### v) to 08_final/<field>.md.
- Generate
08_final/SUBMIT.md — a single document with each form field heading + the paste-ready answer underneath, in the exact order from 01_form-spec.md. The founder pastes from here into apply.ycombinator.com.
- Append milestone via
<skill-dir>/scripts/journal-append.sh.
Step 10: Phase 8 — Interview prep (only after invite)
Refuse to enter Phase 8 unless 00_meta.md has interview_invited: true written by the founder.
STOP. Read references/interview-playbook.md in full before running drills. It contains the 10-minute interview signals, the inverted-pyramid 10-second rule, the "I don't like your idea" probe simulator, co-founder choreography, demo-readiness mandates, and "what not to say" lines.
Run, in order:
- The 5-question rapid-fire drill (10-second visible timer per answer): what are you building, what's your progress, what's your unique insight, how will you get the first 1000 customers, who are your competitors and why are you different.
- The "I don't like your idea, what else have you got?" probe simulator. Score the response: zero defensiveness, genuine curiosity ("why do you think it won't work?"), credible alternative on tap.
- Co-founder choreography check: enforce written pre-division of questions by area (CEO: business/market; CTO: technical/team), no interruptions, no real-time contradictions.
- Metric cheat-sheet generator: list every numeric claim in the application; for each, the source/verification (partners may Google-check live).
- Demo-readiness check (24h before): "Is the product loaded in a browser tab? Can you screen-share in <10s if asked?"
- Send a pre-interview note: "Walk in with a real smile and full attention. Partners read earnest energy as positive; flat scripted energy as negative."
Save all drill outputs to 07_interview/.
Step 11: Phase 9 — Post-mortem (auto-prompt on rejection)
If 00_meta.md has decision: rejected, the next /yc-apply invocation auto-opens Phase 9. If decision: accepted, Phase 9 is skipped.
- Ask the founder to paste the rejection feedback verbatim.
- Walk the lessons: what did partners flag? What is unaddressable? What is addressable in the next 3 months?
- Build
10_post-mortem.md and seed a 04_reapplicant-delta.md for the next batch (which the next workspace's Phase 0 will inherit).
Gist tripwires (operating-level)
These tripwires are not the contract; the references are. If any tripwire fires, follow the STOP directive for that phase before continuing.
- Skim test. Each first sentence in
05_drafts/*.md must let a reader reproduce the answer's core claim. If not, rewrite.
- Buzzword scan.
scripts/buzzword-scan.sh returns 0 hits or every hit is journaled with rationale.
- Story compile. The Dalton story must be reconstructable from the drafts alone. If not, the founder has not told a story yet.
- Honesty audit. Every numeric or named-customer claim is defensible if a partner Googles it on the call.
- Anti-script doctrine. Founder video stays as bullets. Script only on explicit opt-in with warning.
Anti-patterns
Behaviors the skill must refuse, regardless of founder pressure:
- Writing a verbatim founder-video script without an explicit opt-in.
- Submitting drafts (Phase 7) without Phase 3 external research complete.
- Bypassing the buzzword scan or fabricating its result.
- Producing application answers without surfacing accepted examples from
references/accepted-examples.md.
- Coaching personality, "earnestness," or "banana mindset" — those are interview-room signals a written tool cannot reliably train.
- Cold-outreach to YC partners. The only documented exception is a recommendation from a YC-portfolio-company founder (see
partner-signals.md).
- Recommending another accelerator as a path into YC.
When NOT to use
- Pitch-deck design unrelated to a YC application — use a dedicated deck skill.
- Generic startup advice without the founder intending to apply.
- Post-Series-A fundraising or hiring work.
- AI tools / models comparison without a YC application context.
Error Handling
- Workspace ambiguous (no clear candidate): ask the founder explicitly; never assume current directory.
scripts/bootstrap-workspace.sh fails on permission or existing files: report the error and ask whether to overwrite, choose another path, or abort.
- A subagent in Phase 3 returns malformed output: re-dispatch only that slice with stricter instructions; never synthesize the missing slice.
scripts/buzzword-scan.sh returns hits the founder insists are intentional: require a one-line rationale in 09_journal.md per hit; never silently allow.
- Phase 6 gate fails: write
06_gate.md with the failing checks left as [ ] and no RESULT: PASS; refuse to enter Phase 7. Only --force-submit (recorded as RESULT: FORCE-SUBMIT + a journal warning) bypasses it.
- Coding-agent session export oversize (>25 MB): ask the founder to trim or pick another file; do not attempt to compress automatically.
- Live YC form differs significantly from bundled spec: trust the paste, not the bundle; warn the founder and update
01_form-spec.md from the paste.