This Week on How I AI: GPT 5.5, Claude Design, and GPT Images 2.0 — Plus the Memelord Story

Source: Lenny's Newsletter (How I AI) · Author: Lenny Rachitsky · Date: Apr 27, 2026 · Original link

This issue rounds up three episodes of How I AI with hands-on reviews of OpenAI's GPT 5.5, Anthropic's Claude Design, and OpenAI's GPT Images 2.0 — and a builder profile of Jason Levin, the non-coder who turned Memelord into a $3M API business. The connecting thread: AI tools are now strong enough to do days of expert work autonomously, but choosing which tool for which job (and how hard to push it) still matters a lot.

---

Episode 1 — "GPT 5.5 just did what no other model could"

Claire Vo (ChatPRD) put GPT 5.5 through real, messy problems. The verdict: it's the first model that genuinely behaves as a long-running autonomous engineer — but only worth its cost when your problems are actually hard.

The smartness/use-case mismatch

On Claire's personal ChatGPT account, GPT 5.5 felt like overkill. The clearest example: she asked it to build a simple subtraction app for her first-grader, and the model spent 17 minutes thinking about it. Impressive reasoning, but the task didn't need it. The lesson: super-intelligent models pay off only when the underlying problem is genuinely complex; for everyday queries you're paying for capability you don't use.

The "I trust you, figure it out" prompt

The biggest unlock wasn't a clever prompt template — it was delegating outcome instead of steps. Claire had a nasty data-migration job: 2 million rows of unstructured data full of edge cases. Her instruction was essentially:

"I trust you to make a call, figure out how to spawn a subagent to do this, test it, identify issues, repair them, and get this ready for production."

GPT 5.5 then ran almost six hours autonomously with zero follow-up prompts, zero steering, and only a single approval check-in along the way. That's a qualitative jump — previous models would stall, ask clarifying questions, or drift. This is the first time she's seen real long-horizon agent behavior in the wild.

The hardware-hacking test

The strongest signal of raw intelligence: Claire had spent months trying to reverse-engineer a cheap Chinese Bluetooth speaker that used a proprietary encoding. She'd tried Claude Code, GPT-4, every tool — none could crack it. So she went full detective: downloaded Bluetooth profiling tools, ran packet sniffers, and crawled Chinese documentation repos. When she dumped that pile of context on GPT 5.5, it figured out both the bitmap encoding and the Bluetooth transport layer. She can now message the speaker from her terminal and has wired it up as a notification display for her Codex agents.

Cost vs. value

Sticker shock is real: GPT 5.5 Pro is $30 per million input tokens, $180 per million output tokens. But the framing matters. In one session it produced six hours of autonomous work, validated 2M rows, and erased six months of accumulated tech debt — all of which would have cost more in human engineering hours, coordination meetings, and focus time. So the "expensive" model is cheaper than the alternative if you point it at problems that justify it.

The "baked potato personality" fix

Out of the box, GPT 5.5 inside Codex has what Claire calls a "baked potato personality" — dull and robotic. Type /personality in Codex and you can swap it for something friendlier. Some testers said the new vibe was "too Gen Z," but Claire prefers it over the bland default — small thing, big quality-of-life win on long sessions.

🎧 Listen: YouTube · Spotify · Apple

---

Episode 2 — "I spent $200 on Claude Design so you don't have to"

Claire stress-tested Claude Design and GPT Images 2.0 against real marketing assets — landing pages, slides, brand kits. Headline: AI design tools are now first-draft monsters but still lose the refinement game to Figma.

Design systems are now first-class

Claude Design's workflow starts with importing your design system — fonts, colors, components, brand assets — and structuring them so the AI uses them consistently. That's a fundamental shift from earlier prototyping tools that ignored brand. As a tell that this is the industry direction: Google just released Design MD, a proposed standard for describing design systems to AI agents.

Where it shines vs. where it struggles

• Marketing assets (good): landing pages, marketing sites, decks that need to match brand — Claude Design respects design systems impressively here.
• Product UX (weak): for app components and complex flows, it doesn't reason as well within the design-system constraints. Pick the tool to match what you're actually building.

Iteration speed is still Figma's moat

A subtle but big point: Claude Design takes 5–10 minutes per generation and every tweak is another LLM call. Figma lets you drag, swap fonts, adjust colors instantly with no model in the loop. That immediate feedback loop is more valuable than people admit when you're refining a design. AI tools = great first draft; traditional tools = still own the polish phase.

The "Claude Design slop tell"

Just like Claude-written prose has tells ("in summary…"), Claude Design has a visual signature: italicized serif fonts everywhere, especially on landing pages. Once you notice it, you can't unsee it — useful both for spotting AI-generated work and for explicitly overriding it in your prompts.

GPT Images 2.0 cracked typography

Previous image models botched text and layout. GPT Images 2.0 can generate multi-page brand kits with proper text rendering, consistent layouts, and sophisticated typography. For marketers who need assets that combine images + text + layout, this is the breakthrough — output looks expensive, not obviously AI-generated.

Let the AI run wild (sometimes)

Counterintuitive finding: when Claire asked Claude Design to make a '90s GeoCities version of Lenny's Newsletter without a design system, it produced "Lenny's Product Zone" — Comic Sans, brick backgrounds, and surprisingly great copy like "Your OKRs are cringe (and seven ways to fix them before Q3)." Reference styles and creative direction outperform rigid constraints when you want something genuinely unexpected.

The killer practical workflow: content → slides

Take an article, attach your design system, and Claude Design will spit out a beautiful, on-brand deck — including code-driven elements like animated terminals with blinking cursors. For product marketing, sales enablement, or any customer-facing deck work, this loop just works today.

🎧 Listen: YouTube · Spotify · Apple

---

Episode 3 — "How a non-coder built Memelord: From a $6.90 newsletter to a $3M API"

Jason Levin grew Memelord to $100K ARR without writing code, then rebuilt it as an API-first product designed for AI agents. His worldview is provocative on purpose, but the operational ideas underneath are sharp.

"Let your marketers cook — or watch them leave"

At Memelord every marketer has to vibe code. It's not a CEO slogan; it's survival. The free tools section — built entirely by non-technical marketers using Cursor — has produced things like a "bust down filter" that went viral in Turkey and pulled in hundreds of thousands of emails. The argument: when you stop routing marketer ideas through engineering prioritization, you ship weirder, faster, better products. And if you don't let them ship, they'll quit and raise $3M to compete with you (yes, that's how Jason got started).

"No UX is the best UX"

Jason spent months perfecting onboarding while knowing agents were coming and would skip onboarding entirely. His lead investor told him bluntly: "I don't want to use your software anymore — I just don't want to use anybody's software." The takeaway: build a beautiful human experience, but make sure there's an API key waiting at the end — because the winners will be the products agents can sign up for and consume frictionlessly.

Free tools are the new PDF lead magnets

Two years ago Jason wrote about this strategy for HubSpot; now he's living it. Building a free tool with Cursor takes less time than writing an e-book, drives more engagement, and naturally introduces users to the bigger product. Stop emailing people PDFs — give them a Giga Chad meme maker or a Steve Jobs portrait generator. More fun, more viral, more effective.

Build hyper-personalized software for an audience of one

Jason built a Raspberry Pi keyboard that lives by his bed so he can capture ideas at night without waking his wife. He's working on an in-home camera that uses AI to track where he leaves his keys. These aren't products — they're personal tools. The point: when AI makes software cheap and disposable, you can build absurdly niche solutions that would never make sense as VC-backed companies. Build for yourself first.

Obsession beats expertise (up to a point)

Memelord hit $100K ARR on Bubble with 395 workflows — a no-code codebase Jason describes as "easier to figure out Atlantis" than to understand. He got rate-limited on day two because he didn't know what rate limiting was. The lesson isn't "use Bubble"; it's that obsessive willingness to learn beats technical credentials when you're solving a problem you understand deeply. Start scrappy, prove demand, then hire engineers.

"Be mean to your AI (but not too mean)"

Jason's controversial take: AI is a tool, not a friend — stop saying "thank you" to robots, push them harder, tell them to curse, make them uncomfortable. He claims AI performs better under pressure, and that for creative/humor work you have to push past the safety rails. His rule of thumb: be mean enough that you'd apologize if it grew a body. He argues Grok and Gemini are funnier than Claude and ChatGPT because they're less politically correct. (Take the thesis with whatever salt you like — it's a useful provocation about how RLHF defaults shape output.)

"The most entertaining outcome is the most likely"

Jason's organizing thesis (borrowed from Elon Musk): in 2026, attention belongs to whoever is most entertaining. Who controls the memes controls the universe. The internet is getting more chaotic and more extreme; brands that take being funny seriously will win, boring brands will disappear.

🎧 Listen: YouTube · Spotify · Apple

---

Cross-cutting takeaways

• Match model power to problem hardness. GPT 5.5 is wasted on simple tasks; it shines on multi-hour, ambiguous, expert work.
• Delegate outcomes, not steps. "I trust you, figure it out" produced six hours of autonomous engineering. Specifying every step caps the model at your imagination.
• AI design tools win the first draft, lose the polish. Use Claude Design / GPT Images 2.0 to get to a believable v1; finish in Figma.
• Build for agents as customers. API-first beats UX-first as agents become the dominant buyers of software.
• Let non-engineers ship. Vibe coding turns marketers into product builders, which produces the weird, viral surface area legacy orgs can't match.