Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
How to Give AI Agents Enough Context to Be Useful
TL;DR
Agents fail from missing context, not lack of obedience — Sachin from Skyvern says AI produces “slop” because it follows instructions without the business context a new employee would need, so the fix is better grounding plus self-critique.
Skyvern built real leverage from internal agents at a $2 million+ run rate — Sachin says he handled PM, marketing, sales, and support largely solo by using agents for PRDs, SEO, content marketing, customer support, and even small bug fixes.
The best context is your company’s exhaust: email, Slack, Notion, call recordings, and even databases — his point is that agents become useful when they can connect customer complaints, product docs, and operational data like failed runs into one working memory.
Remote companies have an ‘unfair advantage’ because their context is already written down — in-person firms lose key details in hallway conversations, while remote teams naturally leave behind searchable call transcripts and Slack threads that agents can actually use.
Skyvern’s PRD agent works because it does research and adversarial review before anyone sees it — it searches calls, Slack, Notion, and customer comms, drafts a spec, then runs sub-agents and a RICE prioritization pass to strip out junk requirements.
A simple content agent turned customer calls into pipeline — every morning Sachin gets five post ideas based on the last 20 customer conversations, and one post about insurance carrier portal automation directly led to a warm intro from his network.
The Breakdown
Context Is the Difference Between Helpful and Sloppy
Sachin opens with the core claim: “context is all you need,” and says the reason agents generate slop is simple — they’re eager to help but don’t know your business. He compares them to a brand-new hire who tries hard but lacks onboarding, and says the fix is giving good instructions, rich context, and space to critique their own work.
How Skyvern Uses Agents to Punch Above Its Weight
He gives the credibility upfront: Skyvern, an open-source company serving mostly healthcare customers, scaled past a $2 million run rate while he personally handled PM, marketing, sales, and support. That only worked, he says, because agents were already doing real work every day: writing PRDs, managing SEO, doing content marketing, supporting customers, and fixing small bugs.
Your Company’s Knowledge Base Is Bigger Than You Think
For Sachin, “good context” means everything: email, Slack, Notion docs, customer call recordings, and even the database. The important bit is correlation — an agent should be able to connect “this customer hates this product behavior” with “these runs failed in this exact way,” instead of answering in a vacuum.
Why Remote Teams Are Weirdly Better Positioned
He makes a sharp cultural point: remote companies have an unfair advantage because their context is recorded, while in-person companies let crucial information evaporate in spoken conversations. His framing is blunt and useful — anything you don’t record isn’t saved, so if you want agent leverage, you may need to restructure how your company communicates.
The PRD Agent That Stopped Looking Like Slop
His first concrete example is a PRD-writing agent that intentionally starts from a vague prompt, then does the legwork: searching relevant call recordings, Slack threads, Notion pages, and customer communications. After drafting, sub-agents do adversarial review and a RICE prioritization pass to cut weak requirements; he says the team initially reacted with “this is slop,” and only after repeated tuning did they start saying, “actually, it’s not bad.”
A CAPTCHA Example That Grounded the Work
He points to a February issue around CAPTCHA solver failures, where several customers had complained. The agent didn’t just summarize the problem — it surfaced specific recordings tied to those complaints, so the people building the fix could jump directly into the source material.
Content Marketing From the Last 20 Conversations
His second favorite workflow is a daily email with five post ideas generated from the last 20 internal and external customer calls. The system buckets recurring pain points and contrarian observations, drafts versions for Twitter and LinkedIn, runs a cleanup step with Pangram to remove obvious AI-sounding phrasing, tries to add a meme, and sends it to him for review — enough to help him publish five times a week when he otherwise wouldn’t.
The Operating-System Change: No DMs, Record Everything
He closes with the organizational shift this requires: at Skyvern, no one is allowed to send DMs, and everyone has to ask questions in office channels so future hires — and agents — can benefit. They record every call, even one-on-ones with his co-founder, and give agents broad access across the company; when two new sales hires had to set up call recording on day one, they hated it for an hour and then adapted.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
