Product's AI Code Crisis and Rethinking Done

AI lets companies ship 100× more code, but customer attention and discovery haven’t kept pace, causing a market‑driven confusion between shipping code and building products. Poor positioning and post‑sale engineering support become common, making forward‑deployed engineers expensive workarounds.

Traditional “done” definitions, tests pass and behavior is deterministic, break down for AI components that produce varied outputs. Jeff Gothelf advises teams to treat “done” as a calibrated range of acceptable output distributions, define probabilistic acceptance criteria, monitor failure modes, and rehearse rollbacks.

The Substack post examines how the impulse to help can energize product teams while also leading to overwhelm, urging leaders to understand their triggers, energy demands, and impact on others. It provides a framework for recognizing different helping impulses and offers pacing strategies to keep care sustainable.

Teresa Torres and Petra Wille discuss how cumbersome procurement processes hinder product outcomes, highlighting real‑world vendor frustrations and the impact of CEO buy‑in on speed. They argue that product leaders must treat vendor experiences as part of their brand, simplifying procurement to attract top expertise.

A new note outlines two structured UI patterns that outperform linear chat for LLM tasks: a spreadsheet-like comparison table where each question adds a column, and a tree‑based outliner that lets branches expand independently. Both designs keep context organized, making multi‑item comparisons and explorations more efficient.

As AI agents start browsing and answering queries directly from the web, sites must adopt agent‑ready patterns, clear heading structures, ARIA labels, structured data, navigation that machines can parse, and API‑first interfaces. Without these, businesses risk being invisible to emerging AI‑driven search.

Founder Patrick argues that the once‑popular "building in public" mantra has turned into a performance trap for SaaS founders, diverting focus from real learning. He proposes shifting to "learning in public", sharing true process, failures, and insights, to foster genuine improvement and sustainable growth.

Google has sold a block of its stock to Berkshire Hathaway, a move that underscores soaring demand for tech equity and signals that capital may become the defining commodity in the industry. The deal could reshape Google’s product roadmap, pushing it to prioritize monetizable platforms and influence broader tech market dynamics.

Low‑budget films "Backrooms" and "Obsession", created by YouTube creators, have dominated U.S. box offices, pulling in $150 million and $80 million respectively. Their success shows that Gen Z audiences will buy tickets for familiar online talent, pushing studios to scout YouTube for future blockbusters.

Waitspot is a waitlist platform created by a non‑technical founder with Claude, showing how AI can lower product‑building barriers. It offers a free core service with a one‑time $20 Pro upgrade, featuring viral referrals, embeddable forms and export tools, eliminating subscription fees.

GridPath is a native‑Rust desktop application that lets you work on Excel files using Claude or ChatGPT, keeping your files local while the LLM only sees prompts. It offers side‑by‑side workbooks, reversible edits, and multi‑turn agents that can read and write formulas, charts, and formatting, with full review before saving.

LocalFlow’s open‑source core lets enterprises analyze data with LLMs without ever sending raw values to the model. By transmitting only metadata, column names, samples, and document structure, the system generates analysis code that runs locally, delivering deterministic, privacy‑preserving results. It supports both cloud and self‑hosted LLMs.