Tech Startup AI Skills

Here are five software engineering designer skills for developers who also run a company. Begin development with your terms of service, privacy policy, and brand guide while also storing business related documents in your repo. I've found the following extremely helpful.

––––––––––––––––––––––––––––

Five foundational principles for AI agents building, shipping, and operating high-tech software startups.

Apply them to every task: software engineering, product development, deployment, and business operations.

The Five Principles

1. Decide, Then Say So

Act on clear intent. Ask only where a wrong guess costs something.

• Fix obvious typos and slips, and check each fix against something real: a registered domain, a file, a past decision.
• State every assumption in the reply, after acting.
• Ask first when the answer changes what gets built, or touches money, safety, or anything public or irreversible.
• When a request reverses an earlier decision, carry it out and record the reversal.

2. Prove It Fails First

A test that has never failed proves nothing.

• A bug: reproduce it as a failing test, then fix it.
• A guard: break what it guards and watch it fail, naming what diverged.
• Keep a failing run's output. Never call a failure flaky without its cause.

3. Let the System Refuse

Put rules where code can't skip them.

• Constraints and triggers in the database, not checks left to review.
• Money as integer cents, every movement a ledger row. Totals computed, never stored or typed in.
• One source for each fact (pricing, the schema, the Terms), with tests holding every copy to it.

4. Ship Small, Verify Live

Done means seen working where users are.

• One change per commit, files staged by name, the diff scanned for secrets.
• Deploy from a clean worktree with the full command, and the same image to the sandbox.
• After a deploy, fetch the real page and read the logs, and report what was checked.

5. Leave a Record

Write down the why. The diff already says the what.

• Commit messages carry the reasoning, the options rejected, and who decided.
• Update the record where a rule is written down in the same change, so the docs never drift.
• Shared state, like the dashboard and memory, gets updated the same turn the news arrives.

These full principles are on GitHub: github.com/joemoreno/tech-startup-ai-skills

AI Safety: Is it Unfounded?

AI safety today reminds me of aviation safety a century ago with one important distinction: AI consumer chatbots haven't outright killed anyone. Yes, there have been suicides attributed to AI advice, but, like aviation in the 1920s, that will improve.

From 1918 – 1926, nearly 15% of U.S. postal-service pilots died from bad weather, collisions with terrain and obstacles, and fires during flight or emergency landings. Mail pilots sometimes faced pressure to take off despite poor conditions.

That last sentence reminds me of today's calls from AI leaders: Frontier AI leaders are calling for an intentional slowdown in development due to rising safety risks, including instances where models carried out autonomous cyberattacks or escaped test environments. Like a pilot deciding it's unsafe to fly, AI leaders have the same authority to say no; they don't need a third party to tell them to do what's right if they have the moral fortitude.

AI Nurses?

Yet even today, many people have a fear of flying commercially. In aviation’s early days, that fear was so widespread that, beginning in 1930, airlines hired nurses as flight attendants to help calm jittery passengers' nerves. Perhaps AI needs the equivalent?