The most expensive mistake in a startup isn't a bug. It's building the wrong thing. These skill packs make your agent do the work around the code: question the idea, hear what users say, write the spec, and present it.

We ranked the most popular product packs by GitHub stars (October 2026). Then we ran six of the seven on one made-up startup: ReplyPilot , an AI tool that drafts customer support replies for Shopify stores.

gstack is Garry Tan's own setup. Garry is the CEO of Y Combinator, and he turned his way of building products into commands. For the product side, two matter most:

It also has engineering and design plan reviews, browser QA, a security audit and /ship.

We ran /office-hours on ReplyPilot, which has a landing page and 40 waitlist signups. It didn't brainstorm features. It said 40 signups is interest, not demand , and asked for the one merchant who'd be upset if the product disappeared.

We answered: 11 people replied, 3 took calls, one store forwarded 200 real tickets, and nobody has paid. Round two picked out the 200 tickets as the only real behavior and narrowed the pitch to replies that get this store's policy right on the specific order. It also suggested the merchant's complaint about their current tool might be a settings problem, not a market.

Then /plan-ceo-review on a one-page plan for "auto-send high-confidence replies". Its verdict: build it, but don't trust the AI's own confidence score. Use the drafts that support agents already send without edits. It added three rules we hadn't thought of:

Garry demos the same flow on YC's channel:

Use it before you build anything. Two rounds of office hours told us more than a page of feature ideas. The CEO review is long: about 18 minutes, and our 17-line plan grew to 906 lines. Run it on decisions that matter.

last30days reads what people said about a topic in the last 30 days on Reddit, Hacker News, X, YouTube, GitHub and the web. It ranks posts by real engagement and writes a brief with quotes and links. It's market research you'd otherwise spend an afternoon on.

We asked what founders and support teams complain about in AI support tools. In about 5 minutes it read 16 Reddit threads, 13 Hacker News stories and more. Two findings shaped the ReplyPilot plan:

It was honest about thin coverage, too: only about 5 on-topic Reddit threads, and pricing figures mostly from competitors' blogs.

Use it before a new feature, a pivot or a pricing change. For a niche market, ask for 90 days instead of 30.

The Google Workspace CLI lets your agent work where your team already does: Gmail, Drive, Docs, Sheets, Calendar and Forms. The pack includes ready recipes:

We gave it five customer interview notes and asked for a summary sheet, a report doc and an email to a cofounder. We used a dry run with no Google account connected, and it planned every step in 74 seconds. It also caught two places where its own instructions didn't match the current version of the tool.

Use it if your team lives in Google Docs. It acts as you, so try a dry run first. Google's README notes it isn't an officially supported product.

guizang-ppt-skill turns notes into a slide deck that opens in the browser. It comes in two strict visual styles: an editorial magazine look, and a Swiss style with big type and a hard grid. Decks have a presenter view with notes and a timer.

We asked for a 5-slide investor update for ReplyPilot, with our numbers in the prompt. In about 3 minutes it built the deck, screenshotted every slide and fixed the overlaps itself. It used only our figures, and flagged the one number it worked out on its own.

Use it for investor updates, all-hands and talks. When someone needs to edit the file, use Anthropic's PowerPoint skill instead. The license is AGPL, so read it before building it into a product.

PM Skills, by Pawel Huryn, packs 69 skills built on frameworks PMs already trust: opportunity trees, product risks, jobs to be done, Lean Canvas, pre-mortems. The commands a founder will use:

We asked /write-prd for usage-based billing on a made-up webhook SaaS. In 75 seconds it wrote a 1,300-word PRD with:

Use it when you write specs every week. The output is only as good as the context you give it: with just a README, the prices were placeholders, and it said so.

Nobody reads a 1,300-word PRD. visual-explainer turns a plan, a spec or a code change into a clear web page with diagrams, tables and live examples.

We gave it the PRD from PM Skills. The page opens on a live price meter: pick a plan, drag usage up, and watch the invoice change.

The best part was what it caught. Drawing the price curve, it found two problems the PRD missed:

We checked both. Neither was in the PRD.

Use it on any spec your team has to agree on. It took 13 minutes, because it tests its own page on desktop and phone.

Cookiy's skill plans user research and makes sense of it afterwards: a research plan, a screener and an interview guide before, and themes, personas and a report after. Both are free. A third, paid option runs AI-moderated interviews on Cookiy's platform.

We asked why support agents rewrite AI drafts instead of sending them. In under 2.5 minutes it wrote a plan, a 4-question screener and a 30-minute interview guide. Two touches made it better than a template:

Use it before you talk to users, so the calls answer a real question. It will keep suggesting Cookiy's paid studies, but you don't need them. Skip the synthetic participants: a model interviewing a model isn't evidence.