Evals as Product Guidance

★MediumAct: MediumFull content

Non-Determinism at Temperature=0: A Hidden Eval-Design Trap

ByteByteGo·Aug 31, 2026

Deep technical walkthrough of everything between pressing Enter and the first token (context assembly, safety classification, tokenization, batching, prefill/decode). Notes that identical prompts can …

View details →
★HighAct: HighFull content

A 'Design Critic' Subagent Pattern for Turning AI Output From Bland to Distinctive

Lennys Newsletter·Sep 1, 2026

Ex-Apple AI design lead Anshu Chimala lays out a Discover/Define/Deliver framework for getting genuinely distinctive design output from AI agents, centered on a 'design critic' subagent — shown only a…

View details →
★HighAct: MediumFull content

LLMs as a Judge: The Full Evaluation Stack Reference

ByteByteGo·Sep 14, 2026

Comprehensive explainer on why deterministic tests fail for LLMs and lays out the full evaluation stack: golden datasets, automated metrics, LLM-as-judge (point/pass-fail/pairwise/error-ID), human cal…

View details →
★Very HighAct: HighFull content

Advanced Evals Sequel: 'Criteria Drift' and a Free Open-Source Error-Discovery Skill

Lennys Newsletter·Sep 22, 2026

Hamel Husain and Shreya Shankar publish a sequel focused on the step most teams skip before writing eval metrics: systematically discovering failure modes from real traces. They give a concrete 3-step…

View details →
★Very HighAct: HighFull content

Anthropic's Head of Product for Research/Labs: Evals Are the Primary Artifact of Product Work, Not the PRD

Lennys Newsletter·Jul 26, 2026

Dianne Penn, Anthropic's first technical PM, describes how her team generates 30-40 representative examples encoded as prompts with golden answers for every major feature — that eval set, not the PRD,…

View details →
Very HighAct: HighFull content

Braintrust CEO: Evals Are the Modern PRD — Build the Feedback Loop First

Lenny's Newsletter / How I AI·Jun 15, 2026

Ankur Goyal (Braintrust CEO) lays out a concrete eval-driven product process: treat evals as the modern PRD, build a feedback loop that turns real production data into new evals continuously, and run …

View details →
HighAct: HighFull content

How DoorDash Built a Simulation + Eval Flywheel for LLM-Based Support

ByteByteGo·May 30, 2026

DoorDash built an offline simulator that generates multi-turn support conversations, then an eval framework grades them against quality criteria. The flywheel: simulation generates eval cases, eval fa…

View details →