Dave Gelinas

I ship AI products for a living and build with AI every day. Posting what actually works with agents, and with AI in the classroom. Having more fun than ever.

[Dave Gelinas @davegelinas]

I'm an optimist about AI, on two conditions. Agents get a key to one door, and earn the next one. Kids learn the basics first, because you can't check what you never learned. Most of what I post here comes back to those two.

[Dave Gelinas @davegelinas]

The PRD is becoming the fax machine of product management. Still in use. The thinking that went into it now happens in the conversation with the agent.

[Dave Gelinas @davegelinas]

@anupamrjp https://t.co/R2l1cvahwk A thirty-day practice for manifesting one goal. Your intention becomes a manifestation statement, and every morning brings affirmations and one aligned action, plus an evidence journal to track your progress.

[Dave Gelinas @davegelinas]

@petergyang @once_reading Your first point is the product insight. The parent decides whether tomorrow's lesson happens, so the habit you have to build is theirs. Designing the parent into every lesson is why this one lasts.

[Dave Gelinas @davegelinas]

@alexalbert__ @addyosmani @lydiahallie @_catwu @bcherny @sidbid @jarredsumner @adocomplete @delba_oliveira @connerphillippi Feedback for @bcherny and @_catwu: compact when the topic changes, before the window fills. A new subject is the strongest signal that the old context is dead weight.

[Dave Gelinas @davegelinas]

Hypothetical: your AI agent gets an annual performance review. What goes under "areas for improvement"? First on my list: grading its own homework.

[Dave Gelinas @davegelinas]

@leerob Feature request: let me approve a scope once instead of every step. "Can email people in my contacts, nobody else, until Friday." By the tenth approval prompt, everyone clicks yes without reading.

[Dave Gelinas @davegelinas]

@rauchg The upside: patience works for defenders too. Point your own agents at your sandbox before someone else's do. That's what OpenAI just did, and it's why we're hearing about it from them.

[Dave Gelinas @davegelinas]

@jasonlk If agents are the buyer, error messages are the sales team. An agent that hits a vague one just moves on to the next vendor.

[Dave Gelinas @davegelinas]

@clairevo Fun prompt. Run it read-only, though: an agent digging up your embarrassing trail shouldn't be able to add to it from your inbox.

[Dave Gelinas @davegelinas]

@Sacrifice_Isaac Humanizer, most likely. Those tools swap in look-alike letters to beat detectors, and detectors will keep losing that race. A paragraph written in class is the one baseline no tool can fake.

[Dave Gelinas @davegelinas]

@NoemiTitarenco Keep 20 of your own real tasks and rerun them every week. If a provider ever swaps in a quant, you'll have the chart with the date on it.

[Dave Gelinas @davegelinas]

@gregisenberg FICO for agents only works if the score travels, the way your credit history follows you from bank to bank. A score one site keeps to itself is just a grudge.

[Dave Gelinas @davegelinas]

@elonmusk @bot A vault is a good start. Next is the valet key: pausing ad campaigns should take a key that can pause campaigns and nothing else, not the whole Google login.

[Dave Gelinas @davegelinas]

Companies won't rebuild their CRM. They'll stop paying for the 40 little tools bolted onto it.

[Dave Gelinas @davegelinas]

@lennysan Model timelines keep arriving early. Org timelines keep arriving late. Most companies are still pasting into a chat window.

[Dave Gelinas @davegelinas]

@alexalbert__ The moat in a small business is what the owner knows and never wrote down: which supplier runs late, which regular is about to stop coming. The agent that gets that out of their head beats the one with better features.

[Dave Gelinas @davegelinas]

@rauchg The underrated win with expiring keys is memory. A key that dies on its own is a grant you never have to remember to revoke.

[Dave Gelinas @davegelinas]

@OlegFlowcraft @bcherny @AgenticAIFdn Agree, and add a ceiling: allowed-tools only pre-approves. A skill that declares what it can touch, and gets nothing else, needs far fewer human yeses, so it can run unsupervised outside your list.

[Dave Gelinas @davegelinas]

Your agent, your liability. Hand it your bank login and you granted everything, including the parts you never meant to.

[Madison Mills @MadisonMills22]

SCOOP: Major U.S. banks are detecting unauthorized third-party AI agents inside customer accounts, including some attempting to take actions that appear to be independent of user instructions, sources told Axios.

The findings have alarmed banking executives, who are racing to restrict access to their systems as AI moves at a velocity faster than financial institutions can respond, multiple industry and bank executives told Axios.

Several banks are now preparing for worst-case and even “day after” scenarios, including how to rapidly secure or restore customer accounts if things go wrong, the people said.

Several bank executives told Axios they're struggling to distinguish between human users and AI agents. One executive said banks also struggle to determine whether agents are authorized by customers or potentially operating without their permission, more akin to rogue agent behavior.

Read my latest for Axios here: https://t.co/Q3bda5In8q