I used YC's GStack for 3 projects – no-nonsense review

I spent 3 weeks using Garry Tan’s GStack to plan 4 startup ideas (2 SaaS, 2 robotics). My setup was Claude Opus 4.7/4.8 in VS Code.

What worked

- /office-hours - sharp, 6 forcing questions catch defensiveness, pattern-blindness. Forced me to think harder, I realized one of my product ideas was untestable.

- /plan-ceo-review - sketched out a technical architecture (~ Claude /plan), features, schedule. Injected discipline.

- /plan-eng-review - reviewed architecture, features, schedule. Caught race conditions, idempotency, prompt-injection, magic-link prefetch attacks. Locked decisions. Useful to anyone who hasn’t shipped SaaS at scale.

- /plan-design-review - neat, generated 3 landing page variant mockups (via the OpenAI API).

- The “outside critique” pattern - ran fresh/empty-context Claude instances against my work. Caught warming up Anthropic API to up its rate limits for launch, new domain email delivery warmup.

What’s missing

- No marketing skill. Biggest gap, especially for me/tech founder. GStack recommended “distribution: Show HN.” as the entire marketing strategy. Single post, no audience, no plan. None of subsequent skills flagged it.

- CEO review is 95% CTO review. Architecture, security, error handling, performance. Almost no real business strategy.

- Office-hours skip competitor search. I discovered my closest competitor during CEO review on day 3, not office-hours on day 1. A 30-second web search on day 1 would have changed my whole positioning.

- Time estimates are wildly wrong. Day-0 spike planned for 1 hour. Took a full day. Other days estimated at hours, took 10 minutes.

Where it falls apart

- /plan-ceo-review, /plan-eng-review planned everything up-front - locked features, schedule, and pass/fail tests early. Correct when building to a known customer spec. Wrong when you’re at the brainstorming stage, need to stress-test, iterate the idea.

- The multi-agent chain caused context poisoning. GStack splits work across agents: office-hours, ceo-plan, eng-review, implementation. Each agent reads a document written by the previous one, trusts upstream conclusions. Office-hours Claude trusts my prompts.

But I was doing product research — which is, by definition, what you do when you don’t know what you’re doing. By the end of the chain, things get weird. Implementation Claude insisted on launching a paid email digest service on HN with a single post — even though HN readers are tech-savvy, have Claude chat, can write their own digest in 30 seconds, and would never pay. Wrong audience entirely.

- GStack’s own skills don’t ship with regression tests. Shipping skills without tests is like shipping software without tests, but worse — broken software crashes, while broken skills produce confident, polished, wrong output. Read more here https://news.ycombinator.com/item?id=48405221

- GStack is gameable, and gamed scores feed YC’s application pipeline. SKILL.md is open source, so the rubric is visible. Claude is allowed to grade Claude. I scored 8/10 on office hours, the skill invited me to apply to YC (with ?ref=gstack). Gameable rubric → easy 8/10 → tracked referral to YC application → YC partners read polished output that masks weak fundamentals.

- Conflicting interests? Office-hours complimented me on psychological traits, like I passed a psychological screening test. That’s fine, but is office-hours key goal recruiting or empowering founders or both? I think the skill did a good job recruiting, not a good job empowering (see above about marketing missing).

Bottom line

1. Try /office-hours, /plan-design-review once.

2. Don’t trust/use GStack output for marketing.

3. Avoid workflows/multi-agent chain for early-stage exploration.

GStack is strong scaffolding for a self-aware founder who treats its output as a draft to react to. It’s actively confusing for a founder who treats its output as truth.

iliaov.substack.com