Apr 30, 2026
Erik's Newsletter: The Harness Is the Product
Issue #90
Today at a Glance:
Quote: Picasso on why computers are useless
Thinking: The harness is the product, SaaS is going headless, the PM role is compressing, and kids rise to the expectations we set
Podcast: Nikhyl on product leadership, Claire Vo on personal knowledge systems
Tactics: ccunpacked.dev, Karpathy’s LLM Wiki, evals as PRDs
Try: Defend the whiteboard, be a little selfish
Tweet: Claude Code memory, AI-native companies, chat-bar SaaS, and a little Claude-speak for the lols
#QUOTE
Quote I’m pondering...
“Computers are useless. They can only give you answers.” — Pablo Picasso
Picasso said this in 1968. ENIAC was old, mainframes were new, and the computer was already the platonic Q&A machine. He was complaining that this didn’t help an artist. Artists need new questions, not better answers.
He’s still right.
#THOUGHTS
What I’m thinking about...
The Harness Is the Product
Read: Thin Harness, Fat Skills, Why Your “AI-First” Strategy Is Probably Wrong, The Great Convergence
On March 31, Anthropic accidentally shipped the entire source code for Claude Code to npm. 512,000 lines. Every coding agent builder I know read it that week.
The lesson wasn’t about the model. It was about the wrapper.
Live repo context, prompt caching, structured session memory. None of it makes Claude smarter. All of it gives Claude the right context at the right time. That wrapper is the harness.
The 2x people and the 100x people use the same models. The difference is architecture. Half the gap between the AI-frustrated PM and the AI-pilled one is just that one set up their harness and the other didn’t.
We talk about it like a prompt skill. It’s plumbing.
Headless SaaS Is Eating Itself
Read: The New Software: CLI, Skills & Vertical Models, The Great Convergence
Linear, PostHog, and Attio all shipped chat bars as their homepages this month. Basecamp shipped a CLI version. “Agents have emerged as the killer app for AI.”
Every SaaS company follows Salesforce and goes headless within 18 months. A $1T category of agent-native startups emerges that treats Salesforce like AWS.
The front door of every SaaS becomes a programmable surface. Machine identities already outnumber human users 45 to 1 in the average enterprise. Build for them or stop getting used.
The PM Is Disappearing
Read: Evals are the new PRD, Why Your “AI-First” Strategy Is Probably Wrong
Most of the PMs I work with weren’t doing PM work. They were doing translation work between customers, engineers, and execs. That layer collapses when the agent does the translation.
What’s left is the thing PMs were always supposed to be doing: deciding what’s worth building and what great looks like.
Evals as PRDs is the same shift. You don’t write requirements anymore. You write the test that measures whether the system did the thing.
If your PM job is mostly Jira tickets and roadmap docs, you’re learned. If it’s mostly “what do we kill” and “what does great look like,” you’re learning.
Children, Mirror
Read: We Have Been Underestimating Children
Wendy spent a day at Alpha School and watched 3rd-6th graders run their own learning, give each other feedback, and command rooms with more confidence than she’s seen in professional development sessions.
Her takeaway: “We have taught children that needing help is expected, that waiting for instructions is appropriate, and that someone else is always responsible for their learning.”
Kevin Dahlstrom on the parent side: “The greatest burden a child must bear is the unlived life of the parents.” Be a little selfish. Keep your passions. Don’t become a martyr. Your kids absorb the version of you they actually see.
The fix isn’t more boundaries or fewer screens. It’s higher expectations for kids and a fuller life of parents.
#PODCASTS
My favorite listens...
Smiling Exhaustion (Lenny’s Podcast)
Nikhyl is always on point. Listen to this one if you’re in product.
#TACTICS & TOOLS
Tools and techniques I’m using...
ccunpacked.dev
A reverse-engineered walkthrough of how Claude Code works under the hood. The leaked source plus this site teaches more about agent architecture than any course. If you want to build a harness for your own work, start here.
https://ccunpacked.dev/
Karpathy’s LLM Wiki
Karpathy went viral with a pattern I’ve been refining for months: don’t make your AI re-read documents every query. Have it maintain a structured wiki of markdown files that compounds over time. I run mine in this monorepo with Claude Code. Stop using AI as a search engine. Start using it as a librarian.
Evals as PRDs
Stop writing PRDs that gather dust in a Google Doc. Start writing evals that run on every commit.
An eval suite is the spec, the acceptance criteria, and the roadmap all at once. The PM job becomes: define measurable signals, set the bar, hand engineering a number to make go up.
https://www.braintrust.dev/blog/evals-are-the-new-prd
#THINGS TO TRY
Experiments worth running...
Defend the Whiteboard
Kpaxs noted that Nvidia’s culture turns whiteboards into engines of clarity. There’s something agent-resistant about a whiteboard.
You can’t AI-summarize it. The summary destroys the thing that made it work. I’m being more deliberate about which conversations belong on a whiteboard versus in a doc.
Be a Little Selfish
On parenting: too many parents surrender themselves after having kids. They stop pursuing their passions. The kids absorb that. Keep the version of yourself you like, and your kids get to see what an actual life looks like.
#TWEET
Social media that made me stop and think...
AI + Work
How Claude Code’s memory works under the hood. Constrained, structured, ruthlessly compressed. Don’t store everything. Store the right things.
YC published the playbook for building a modern AI-native company from the ground up. Read it before you hire your tenth person.
SaaS + Building
Linear, PostHog, Attio all shipped the same thing in the last few weeks: the homepage is a chat bar, not a dashboard. UI is officially admitting one homepage can’t serve every user.
Three corporate phases: Visionary builds the thing. Operator scales it. Builder (Visionary 2.0) makes it exciting again. Most companies die stuck between two and three.
For the Lols
Adopting Claude speak in regular life.
Platform 6-7. No notes.