Philip Behrends

VC, startup advisor, pitch designer, photographer,

Need a pitch deck update?
photos?

[Billy Dinh @billydeee_]

Summer in New York City (2025)

[Soleio @soleio]

Being wrong is forgivable. Being slow is not.

Years later, the primacy of speed is still the best lesson from FB.

[Philip Behrends @pbehrends]

@nicochristie Would love to use this, if you have another invite code? 😇

[Anthropic @AnthropicAI]

Some of those failures were very weird indeed. At one point, Claude hallucinated that it was a real, physical person, and claimed that it was coming in to work in the shop. We’re still not sure why this happened.

[GREG ISENBERG @gregisenberg]

friendly reminder: productizing AI agents is the biggest zero-to-one play on the internet right now.

step 1

start with the workflow. open someone’s “how-to” loom, watch for any section where the mouse repeats itself for more than 30 seconds. that is your extraction point. i call this step looking for drudge.

step 2

add the context. scrape the internal FAQ, the tribal acronyms, the edge-case screenshots that a general LLM will never see. that proprietary mess formatted into plain text or json is your first moat. sometimes you need an audience for this. it helps. usually i create an audience/community to start.

step 3

wire up an agent on an LLM or agent platform: claude, gpt-4o, gemini, lindy, string dot com. keep the prompt brutal-simple: “given this context and that input, return this output.” test until the agent’s average response time beats the human baseline by 10×. keep refining prompt, it will get longer, longer and more valuable. never share that prompt with anyone.

step 4

wrap the agent in the thinnest UI possible. one drop zone, one action button, one log window that says “done.” hide the settings; hide the model; hide the magic. outcome is the only feature. make a brand people fall in love with.

step 5

price the saved hours. take the median hourly cost of the person who used to run the workflow, multiply by the time you shave off each week, and anchor the subscription somewhere south of pain but north of hobby.

step 6

ship video proof before a landing page. record old way vs. new way in a split-screen clip; caption it; post it where the operators hang out. every “how did you do that?” reply is a warm lead. you can use platforms like lindy ai etc to do cold dm/email automation.

step 7

loop usage back into the context file. new ticket? new edge case? append it. your agent improves without code, and the context corpus becomes asset #2.

step 8

layer distribution inside the product. add a quiet “powered by” footer on every output, or trigger a shareable report by default. users spread the gospel while you sleep.

step 9

when churn stabilizes and support tickets hit near-zero, bolt on adjacent micro-agents each with its own promise, its own subscription tier, no extra complexity in the UI.

step 10

ignore vanity metrics. screenshots of ARR can be faked (and are constantly). My favorite was watching someone screenshot test payments. Literally fake money lol.

Anyways, time returned to a paying customer cannot be faked. compound that saved time across niches until you own a stack of invisible employees.

you own a portfolio of little products that save people time and you get paid for it. the market for this is in the billions.

do this on repeat and you won’t need funding, permission, or glossy charts.

you will be free and be creating impact.

this is probably one of the biggest opportunities out there.

[Philip Behrends @pbehrends]

@fabianarbor Exactly what I’ve focused on for the last 2 years. 30+ happy startup / scale up clients. VC background & strong narrative-first approach. DM or https://t.co/Ihm3srOhjv

[Philip Behrends @pbehrends]

@felixleezd @ADPList Design

[Philip Behrends @pbehrends]

@87xysr @harshaog @browsercompany 🫶 yes please!

[Philip Behrends @pbehrends]

@harshaog @browsercompany 🫶

[Philip Behrends @pbehrends]

@jeffreympark Yes

[Philip Behrends @pbehrends]

@NordicPatrik Landing

[Philip Behrends @pbehrends]

@laurentiu_ol Design

[Google @Google]

Let's go hands-on with #GeminiAI.

Our newest AI model can reason across different types of inputs and outputs — like images and text. See Gemini's multimodal reasoning capabilities in action ↓

[Philip Behrends @pbehrends]

@nicktrimmer @erihong Style looks very much like @Dannmace

[Philip Behrends @pbehrends]

@n335y Thank you thank youuu ❤️ also for officially declaring it a public holiday 😂

[Philip Behrends @pbehrends]

@heyeaslo @NotionHQ 👋

[Philip Behrends @pbehrends]

@heymattia 👋

[Philip Behrends @pbehrends]

@chddaniel 👋

[Philip Behrends @pbehrends]

@yungconnasse Definitely not as enjoyable. Ideal would be mid-morning coffee & siesta after lunch. Not the most work-friendly schedule though 😅

[Philip Behrends @pbehrends]

@yungconnasse 6am run 🤝 11am nap 😂