Explore

Onboard GrokBot by giving it tool access first, then asking what it can do

Roman Ugarte recommends that when onboarding GrokBot, the first thing is to give it access to your tools — such as Slack, email, and company records — so it has the context to be successful, and then to ask GrokBot what it can do for you. For brand-new users, he suggests giving it a task like going through Slack and email to suggest five things it can take off your plate, and spinning off separate bots from the tasks it identifies as actionable.

hedged

Lenny's Podcast · How we built Grok Bot in a month | Roman Ugarte (SpaceXAI)

Use roles to group work instead of creating a new chat per task

Roman Ugarte recommends using roles — different swim lanes of work — instead of creating a new chat for each discrete unit of work, which he says leads to problems like copying and pasting between chats.

hedged

Lenny's Podcast · How we built Grok Bot in a month | Roman Ugarte (SpaceXAI)

Instead of adding complicated features to an existing AI product, start fresh

The host calls starting fresh instead of bolting onto an existing AI product a really valuable lesson, and Roman Ugarte endorses considering a simpler solution rather than adding to complicated existing AI products.

hedged

Lenny's Podcast · How we built Grok Bot in a month | Roman Ugarte (SpaceXAI)

Make the core task just work instead of adding features

Roman Ugarte says the second big push before launch was making it just work: users want not a thing with lots of dropdown menus and bells and whistles, but something where you describe a meaningful task and it goes and does it. He also recommends ruthlessly unshipping experimental features users don't need to see, and hill climbing on five critical backend problems that users don't directly see but totally feel.

hedged

Lenny's Podcast · How we built Grok Bot in a month | Roman Ugarte (SpaceXAI)

Delete engineering scaffolding as models get smarter, and reinvent priorities every six months

Roman Ugarte recommends deleting engineering scaffolding once the model becomes smart enough to handle the task natively, building the thing for three months from now and then repeating the cycle, and completely reinventing priorities and the core product every six months or faster to stay competitive. He argues that if the company can't reinvent itself that quickly, it will lose.

Lenny's Podcast · How we built Grok Bot in a month | Roman Ugarte (SpaceXAI)

Think in capabilities given to the bot, not features added to the product; kill pixels

Roman Ugarte endorses thinking in terms of what tools and capabilities to give GrokBot rather than what new things to add to the product, saying adding things to the product is not the goal and users care about bots reliably doing impactful work behind the scenes; he urges killing as many pixels as possible so the bot manipulates them behind the scenes.

Lenny's Podcast · How we built Grok Bot in a month | Roman Ugarte (SpaceXAI)

Make setup effortless and hide power-user abstractions from users

Roman Ugarte recommends that the setup be really easy, not hacky like a home VPN plus a Mac mini, and that abstractions familiar to power users, like skills and slash commands, be removed so ordinary users never encounter them. He also says users should never see the interface for creating an automation, and that primitives should be created in the background so users don't need to be on the cutting edge of AI.

hedged

Lenny's Podcast · How we built Grok Bot in a month | Roman Ugarte (SpaceXAI)

Close the data flywheel: collect traffic, generate insights, promote models

Use Weave to collect production traffic, which is super critical so you can learn and iterate. Generate insights using humans, ARIA, and LLM judges to identify behavioral nuances. Enrich tasks, implement models, and evaluate using W&B Weave as a shared dashboard where the team can make decisions together, and promote the best model to production through a registry to close the data flywheel.

hedged

AI Engineer · An AI Research Agent That Runs Your Experiments — Tim Sweeney, Weights & Biases

Face pull set volume: 2-3 sets to failure for hypertrophy; 3-4 sets weekly or split across corrective sessions

The creator advises that hypertrophy face pull sessions going to failure should be in the two to three set range. For corrective work: if done once a week, do three to four sets that day; twice a week, two sets each session (four total); three times a week, one to two sets per session (three to six total). He notes you can usually handle more corrective workload because of the higher training volume you're accustomed to.

ATHLEAN-X™ · Do This AFTER Every Workout (NON NEGOTIABLE!)

Corrective face pull option: move dynamically overhead while keeping the five-second hold

The creator advises that in the corrective face pull you can dynamically move straight up overhead instead of staying statically in the isometric position; this involves the lower traps, another postural muscle that tends to be weak and dominated by other muscle groups, and you still keep the five-second hold so the rhomboids, rear delts, and rotator cuff isometrically hold you in position.

ATHLEAN-X™ · Do This AFTER Every Workout (NON NEGOTIABLE!)

Provide layout information and OCR alongside visual embeddings for non-multimodal LLMs

Aamir Shakir recommends using layout information and OCR for downstream tasks when the LLM is not multimodal, letting you decide yourself whether to feed the screenshot or the text.

AI Engineer · Better Search Took This AI Agent From 36% to 98% Accuracy - Aamir Shakir, Mixedbread

Save caller information in real time and move on without waiting for save confirmation

Jesse Hall recommends that a voice agent handling a booking should save every piece of information the caller provides to the backend in real time, and should not block the conversation waiting for save confirmation — it should continue immediately and handle errors asynchronously.

AI Engineer · How to Build Voice Agents That Don't Fall Apart When You Interrupt - Jesse Hall, LiveKit

Production checklist: latency budget, confirmations, identity verification, and repeated simulation

Set your latency budget that the caller hears. Confirm critical dates and anything financial or money-related — make sure those checks are in place. Verify identity before an account is accessed. Run the complete conversation repeatedly through simulations, testing final state, privacy safeguards, unnecessary tool calls, and turn-taking.

AI Engineer · How to Build Voice Agents That Don't Fall Apart When You Interrupt - Jesse Hall, LiveKit

Keep up with caller changes to payment amounts and dates in real time

Jesse Hall stresses that when a caller changes details mid-conversation, such as the payment amount or date, the voice agent must keep up with the changes in real time or the conversation fails — especially in finance, where the risk is much higher.

AI Engineer · How to Build Voice Agents That Don't Fall Apart When You Interrupt - Jesse Hall, LiveKit

Respond promptly to payment-status questions; silence erodes trust

Jesse Hall argues that when a caller asks 'did my payment go through?', the agent should respond promptly rather than leaving silence, because the heard gap erodes trust — two seconds of silence after that question is not reassuring — and this gap that the caller hears needs to be measured and minimized.

AI Engineer · How to Build Voice Agents That Don't Fall Apart When You Interrupt - Jesse Hall, LiveKit

Run the open-source turn detector evals and benchmarks against your own data

Jesse Hall notes that LiveKit's turn detector evals and the benchmark methodology and simulations are open source, and recommends running them against your own dataset and agent to see your own interruption numbers and how your setup performs, since turn detection numbers depend on the use case.

AI Engineer · How to Build Voice Agents That Don't Fall Apart When You Interrupt - Jesse Hall, LiveKit

Test the conversation itself, in text for business logic and audio for the voice pipeline

Jesse Hall recommends testing the conversation rather than models in isolation: text-based simulations test business logic only, while audio-based simulations test speech recognition, turn-taking, and the entire voice pipeline.

AI Engineer · How to Build Voice Agents That Don't Fall Apart When You Interrupt - Jesse Hall, LiveKit

Use filler speech like 'almost there' to bridge slow backends

Jesse Hall recommends that when a voice agent's backend is slow, it should use filler speech like 'system is running a little slow, almost there' to bridge the gap and keep the caller informed.

AI Engineer · How to Build Voice Agents That Don't Fall Apart When You Interrupt - Jesse Hall, LiveKit

Give your product personality: treat every touchpoint as a chance for human connection

The creator advises giving the product personality, on the view that users aren't looking for human-to-computer interfaces but human-to-human connection. Concretely: treat every loading screen, dialogue box, message, and especially error message as a chance to connect like a friendly human. Personality in copy doesn't have to be cheesy or unprofessional. The creator also advises sitting down to think about what your product and brand feel like and how to make it connect with people better.

Commercial context: The creator describes this as their own product.

MicroConf · SaaS Product Design: How a B2C App Got to 4% Churn

Start customer contact with a specific interesting message, not a pain-point pitch

Snyder recommends reaching out to potential customers on LinkedIn or via email with a specific, interesting message rather than a generic pain-point pitch — get their attention with something interesting instead of pitching pain relief in the first message. Warm intrusions are one way to get first customer conversations.

Lenny's Podcast · How to find product-market fit (but, actually) | Rob Snyder (Author, The Power of Pull)