Judge atom answers with a different model family
Nikolay Petrov recommends using a judge or a panel of judges — ideally from a different family of models — to evaluate the model's answers to the atom questions in skill atomization.
42,607 results · Most relevant
Nikolay Petrov recommends using a judge or a panel of judges — ideally from a different family of models — to evaluate the model's answers to the atom questions in skill atomization.
Nikolay Petrov recommends checking the strictness of model knowledge in two ways: a lenient check for the general idea and a strict check for exact details such as exact keys, parameters, values, or steps — since the gap between them is large, especially in frontier models.
Sparsh Jain recommends routing surface-level research queries to only the active memory nodes, and routing deeper research queries to both active and inactive memory nodes.
Sparsh Jain recommends building looping and eval infrastructure that cleans things which are no longer important, updates things which should be prioritized, and rearranges things so the agent performs better as context accumulates — mimicking sleep: cleaning unimportant items, taking prioritized items to the front, and rearranging everything so the agent thinks clearly next session.
After building the memory layer (formation, updation, retrieval), they built an observability layer around it. With observability in place, build a looping layer combined with an eval layer that, after every session and every day, extracts the learnings, improves the systems, tests it independently, and then makes the changes in the real system.
The host recommends that designers use the design system to build out ideas with mock data templates and refine them further in a Lean UI code playground.
A design system site with embedded React Storybook provides guidance on when to use components and describes their behaviors and states.
The host recommends that designers use a design prompt containing all established context to spin up an automatically generated prototype built on the design system.
Set up Figma Code Connect to map components, their variants, and properties between design and code. This is also very useful for new Figma features.
hedged
Set up mock data templates to easily build out ideas, and use a Lean UI code playground to refine them further.
Set up an MCP server to provide design system context to AI agents, similar to how a design system site provides context for humans.
Frank Long describes a pipeline approach: first deduplicate information, then filter it, then pull external figures as an alternative configuration.
Frank recommends using humans for tasks requiring judgment and real-world navigation, like meeting a company's management team. He discourages sending an LLM, AI agent, or an if statement into a management meeting — judgment and the ability to navigate the real world make a person the right choice. He also notes that LLMs are really bad at translating information to portfolio impact — that step should be a human using a deterministic system to help them search and pull information.
hedged
Jeff Cavaliere advises that after reaching positive failure on floor flies, you bring your elbows into your sides and press up to reposition for the next eccentric fly, going through failure for better gains. He also cautions not to always look at the fly as an exercise that has to be done light — do it with heavier weights for better gains.
Jeff Cavaliere advises that after reaching failure on dumbbell presses, you drop to the floor and perform a set of push-ups to failure using body weight as resistance. He calls this push-up drop set one of his favorite techniques for building the chest bigger, stronger, and better.
The creator recommends running the team through a two-to-three hour OKR planning process once all information and data is gathered, and running the same process across every organization so everyone is aligned; the process includes a SWOT analysis and is supported by an OKR agent built on the Measure What Matters framework.
The creator recommends that before setting team goals at the QBR, you plot all planned initiatives against the capacity budget and decide what to give and take — e.g., when at 65 hours with everyone thinking everything has to happen, or when a huge AI training and a product release are both proposed.
The Vercel COO asks his GTM engineering team every quarter whether another company has built something better, faster, or cheaper than their in-house agent — and if the answer is yes, the agent goes.
The Vercel COO advises applying the new build-versus-buy test: buy from any vendor who can sell you the same outcome; if they can't, build it yourself by carving out a few hours a week to start experimenting today.
Cohen advises testing new pricing and positioning on sales calls first, where you control the message and get immediate feedback, rather than running website split tests. On a call, say a new higher price out loud and watch the prospect's reaction — you'll often be surprised how easily they accept a price you were scared to quote (e.g., jumping from $99/month to $1,000/month).
Choose sources → to get a library of your own and a daily brief.