Sarah Ponn Foreman

Idea girl who writes the code too. Obsessed with UI.
Music tech, agents, consumer apps.

[Sarah Ponn Foreman @SPF_Labs]

@dashboardlim Meta Claude

[Suhas @zuess05]

Every single person coding with AI right now:

> Get an idea at 11 PM.

> Build the entire app with Claude by 3 AM.

> Realize you actually have to talk to
humans to get sales.

> Panic and start a new project to avoid marketing.

Can we coin a term for whatever mental illness this is? 😭

[Sarah Ponn Foreman @SPF_Labs]

$60 billion for a code editor and honestly? Underpriced. Cursor is where the actual work happens. Models are commoditizing, the interface isn't

[Sarah Ponn Foreman @SPF_Labs]

@Jason I'm in

[Jon Lai @Tocelot]

a16z @speedrun request for startups: GUIs for Agents

we’re still in the MS-DOS era of agents today - CLI, terminal sessions, file directories deleted by openclaw etc. while a small slice of silicon valley are power users, we're SO early for the rest of the world

at Speedrun, we’re looking for bold founders excited to bring the power of agents to normies everywhere. there's a whole slew of products to be built here - from agent builders to marketplaces to managed infrastructure

one broad idea we’re excited about are visual abstraction layers for agents. if you don't know exactly what you want, a command line / chat interface is paralyzing - you need to see options

1 example - think of a GUI or visual command center inspired by strategy games (ex. Factorio) where agents and workflows are represented graphically. skills, tools, MCP connections, background processes, etc could all be configured and shown visually in a workspace

on UX, strategy games have long perfected agent management. zoom to get a birds-eye view of your agents, batch and queue orders via shortcuts, assign agents in multiplayer etc. a well-designed agent command center would make multi-agent orchestration for normies feel easy & intuitive

most folks today still haven't moved beyond ChatGPT. the potential is enormous - just as Windows unlocked mass-market use of personal computers, the right visual abstraction layer could unlock agentic work for everyone - from individuals to enterprise teams

if you share our vision, we'd love to chat!

[Sarah Ponn Foreman @SPF_Labs]

The news that Tom Steyer is now the heavy favorite to be CA’s next governor (@Polymarket ~69%) is the final “thanks, but no thanks.”

More taxes on corps & property, single-payer healthcare for a system already drowning in fraud, Green New Deal 2.0, strict AI regs, and “free” everything paid for by… us.

This isn’t fixing affordability - it’s accelerating the entrepreneur exodus. California used to build things. Now it just builds barriers.

Who else is done?

[Sarah Ponn Foreman @SPF_Labs]

Stuff you might have missed this week:

• Anthropic's api is already returning references to Opus 4.7.
• Opus 4.6 performance appears to be getting dialed back - same pattern they've done before new releases in the past.
• Screenshots leaked of a full-stack app builder inside Claude - not Claude Code, a completely separate no-code tool. Live preview, integrated database, one-click deploy.

Lovable's entire $6.6B thesis is prompt-to-app, and Anthropic might just... include it for free inside Claude.

[Sarah Ponn Foreman @SPF_Labs]

Everyone's building AI products. Almost nobody's building products that use AI well.

The AI wrapper trap: you take an existing workflow, stick a chatbot on it, and call it AI-powered. The user now has to explain what they want in English instead of just clicking a button.

You made it harder and called it innovation.

❌ AI chatbot that helps you write emails
✅ Tool that reads the email you received and has a draft reply waiting before you open it

❌ AI that generates social media posts when you ask
✅ Tool that watches your blog/podcast/video and auto-generates post across platforms without you touching it

❌ AI dashboard that answers questions about your data
✅ Tool that texts you at 8 AM when something in your data looks wrong - you never asked

❌ AI meeting note-taker you have to set up and invite, and then read the notes
✅ Tool that's already in every meeting, already summarized it, already updated your CRM and sent follow-ups to the people you said you'd follow up with

The pattern: the best AI products don't have a prompt box. The user never types anything. The AI just watches, decides, and acts - and the human only shows up to approve or override.

A prompt box is a confession that you didn't know what the user needed. The best AI features feel like the product just got smarter, not like a chatbot moved in.

Build the thing that makes people say "wait how did it know to do that" - not "what do I type."

[a16z @a16z]

Jesse Genet on Agentic Parenting

Jesse Genet joins a16z's Sarah Wang and Katherine Boyle to discuss her journey from founder to parent, how she's using agents in her household, and how AI could transform parenting for the better.

00:00 YC founder turned homeschool mom
03:00 Discovering Claude Code and agentic building
06:00 Building while homeschooling 4 kids under 5
11:00 How AI generates personalized lesson plans and logs progress
18:00 Jesse's 11-agents
27:05 Agent tech stack deep dive
33:56 How agents improve daily life
40:04 Letting kids interact with AI: values, risks, and the future of parenting

@jessegenet @KTmBoyle @sarahdingwang

[Sarah Ponn Foreman @SPF_Labs]

There's a writing skill emerging that has no name yet. I'll call it architecture prompting.

It's not prompt engineering - that's about getting a good answer to a question. This is about writing something that an AI agent will use as its operating manual for hours of autonomous work.

The difference is subtle but it changes everything: human-readable specs are narrative. They explain WHY. They assume shared context. They leave things implicit because a human reader will fill in the gaps.

AI-readable specs are declarative. They state WHAT. They assume zero context. They make every decision explicit because an agent will take the path of least resistance through any ambiguity.

Human spec: "Auth should be simple and frictionless"
AI spec: "Supabase Auth phone OTP. Primary flow: enter phone → 6-digit SMS code → session cookie. No email required at signup. Optional email add later via settings. Cookie expiry: 30 days."

The first version sounds better in a meeting. The second version produces working code on the first try. The people shipping the fastest right now aren't the best prompters or the best coders. They're the best spec writers.

They've figured out how to think in a way that translates perfectly into an AI's execution layer - precise, exhaustive, zero ambiguity.

It's technical writing for a non-human audience. And it's quietly becoming the most valuable skill in software.

[Sarah Ponn Foreman @SPF_Labs]

The EU is deciding whether ChatGPT is too popular and needs more regulation. Meanwhile the US just had the VP, Treasury Secretary, and Fed Chair war-gaming AI cyber threats with bank CEOs in real time.

One side is managing risk. The other is managing paperwork.

[Sarah Ponn Foreman @SPF_Labs]

SpaceX merged with xAI, going public at $1.75T.

One man is building rockets, satellites, AI, and data centers in orbit while regulators in Europe are still debating what a chatbot is.

The gap between builders and bureaucrats has never been wider.

[Sarah Ponn Foreman @SPF_Labs]

AI psychosis is being more excited about the tools than anything you're building with them.

[Sarah Ponn Foreman @SPF_Labs]

Tribal knowledge is the real moat now. Every company has access to the same AI models, the ones that win will be the ones who actually documented how they do what they do.

The rest are sitting on institutional knowledge that walks out the door every time someone quits.

[Chamath Palihapitiya @chamath]

AI is stripping companies down to three key elements:

1) Expert Knowledge - this is the vertical specific understanding of what to do

2) Tribal Knowledge - this is knowledge of the fuzzy grey details that are poorly document and is what creates a dispersion in performance in companies doing roughly the same things

3) Hardware and Software - this is the glue that humans use to stitch Expert and Tribal Knowledge in order to perform

If 1 and 2 are documented well, it allows you to rebuild how you do work. This isn't about replacing humans - in fact I think it allows you to grow faster and hire more people because with a great understanding of what to do, it requires more and more people to express judgement.

This is why we built Software Factory and its being used by many established and large enterprises.

[Jon Lai @Tocelot]

a lot of talk on how 1000 startups just died due to Claude managed agents. I think that’s overblown - the truth is the moat for agentic products has been shifting from infra engineering to domain expertise + data for a while, managed agents is GOOD news if you’re a domain expert turned founder

Previously if you were say a CPA building an agent - you had to wrestle with a lot of infra complexity just to get things working. Sandboxes for execution, state / session management, error handling etc

Now with Claude handling the plumbing, you can focus on domain specific value creation, for example
- tax logic / insights from the 1000s of returns you’ve processed as proprietary data for your agents to pattern match against
- tribal knowledge on tax optimization strategies, state-specific quirks, weird behaviors that increase audit risk etc - things that aren’t found in generalized LLMs today
- domain specific integrations into quickbooks for accounting, plaid for banking, avalara for tax reconciliation etc

At end of day, for startups building vertical AI agents for “X” - what you really want to be selling is not the agent scaffolding itself but the codified expertise of a top practitioner of “X” - the judgement and outcomes of an expert doctor, accountant, lawyer, etc, encoded into software

[Sarah Ponn Foreman @SPF_Labs]

Anthropic just casually shipped what took some startups 18 months to build. This will keep happening.

The only safe place to be is building something where the AI is the engine, not the product. If Claude can replace you with a feature update, you were never a company.

[Sarah Ponn Foreman @SPF_Labs]

Every AI company is racing to ship agents that can write code, browse the web, and use your computer autonomously.

The cybersecurity industry is about to have its best year ever.

[Sarah Ponn Foreman @SPF_Labs]

Sam Altman's own chief scientist tried to remove him over safety concerns.

His co-founder wrote "the problem with OpenAI is Sam himself." His board member described him as someone with "a sociopathic lack of concern for consequences."

And he's still in charge. Still racing toward AGI. Still making decisions that will reshape every industry, every economy, every classroom on earth.

At what point do we admit that the AI future we're sleepwalking into wasn't chosen by us, it was chosen for us by people who can't even keep their own house in order?

[Sarah Ponn Foreman @SPF_Labs]

The most interesting thing about Karpathy's LLM knowledge base is that he's spending more tokens on knowledge than code.

We went from "AI writes my code" to "AI manages what I know." The IDE era for thought is starting.

[Andrej Karpathy @karpathy]

LLM Knowledge Bases

Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So:

Data ingest:
I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them.

IDE:
I use Obsidian as the IDE "frontend" where I can view the raw data, the the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides).

Q&A:
Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale.

Output:
Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base.

Linting:
I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into.

Extra tools:
I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries.

Further explorations:
As the repo grows, the natural desire is to also think about synthetic data generation + finetuning to have your LLM "know" the data in its weights instead of just context windows.

TLDR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts.