CHAPTERS
- 0:02 – 0:32
Skyvern’s agent-driven workflow to scale a lean team
Suchintan introduces Skyvern and explains how AI agents enabled him to handle product, marketing, sales, and support while scaling the company. He previews the kinds of agents they rely on day-to-day across multiple functions.
- •Skyvern automates manual browser workflows (primarily for healthcare)
- •Scaled past $2M run rate with a very lean operator setup
- •Agents help across PRDs, SEO, content marketing, support, and small bug fixes
- •Talk focus: favorite agent workflows and what makes them effective
- 0:32 – 1:02
Why agents generate “slop”: helpful intent without business context
He frames low-quality AI output as a context problem rather than a capability problem. Agents behave like new hires: eager to help, but ineffective until they’re properly onboarded to the company’s reality.
- •Agents follow instructions but often lack critical context
- •Analogy: onboarding a new engineer/employee
- •Quality improves with better instructions plus richer context
- •Goal is to reduce generic output by grounding in company specifics
- 1:02 – 1:17
Three levers for better agent output: instructions, context, self-critique
Suchintan outlines the practical recipe for making agents useful: give them clear direction, feed them comprehensive context, and make them evaluate their own work. This combination helps move from plausible-sounding to actually correct and relevant.
- •Provide good instructions (clear goals and constraints)
- •Supply broad, high-signal context about the business
- •Add critique loops so agents review and improve their drafts
- •Treat agent performance as iterative tuning, not one-and-done prompting
- 1:17 – 1:32
What “good context” looks like: connect the agent to your knowledge sources
He enumerates the concrete data and systems that form a company’s knowledge base. The more the agent can see communications, docs, and product/customer data, the more it can produce grounded work rather than guesses.
- •Email and Slack for decisions, discussions, and customer details
- •Notion/docs for canonical product and process knowledge
- •Customer call recordings for pain points and nuanced context
- •Databases/logs for diagnosing real usage, runs, and failures
- 1:32 – 2:02
Remote companies’ advantage: context is already written down
He argues remote-first workflows naturally create persistent, searchable context, unlike in-person conversations that disappear. To benefit from agents, companies need to treat recorded artifacts as the real operational memory.
- •In-person context is often spoken and lost
- •Remote context tends to be recorded in Slack, docs, and calls
- •Tools become the company’s knowledge base
- •Anything not recorded can’t be leveraged by agents later
- 2:02 – 2:32
Agent workflow #1: generating a PRD grounded in real evidence
Suchintan walks through a PRD-writing “skill” that searches across recordings, Slack, Notion, and customer communications to draft an initial spec. The emphasis is on evidence-backed requirements rather than generic templates.
- •Start with a topic prompt; keep it intentionally flexible
- •Search call recordings for relevant discussions
- •Pull supporting context from Slack, Notion, and customer comms
- •Produce a concise first draft with linked evidence
- 2:32 – 3:02
Adding rigor: adversarial review and prioritization to remove junk requirements
After the first PRD draft, sub-agents critique it and leave comments, then a prioritization framework (RICE) is applied to filter out non-essential requirements. This turns raw drafting into a higher-signal spec process.
- •Sub-agents perform adversarial review and annotate issues
- •Iterative feedback is embedded directly into the document
- •Use a prioritization framework (e.g., RICE) to focus scope
- •Framework step removes “made-up” or low-value requirements
- 3:02 – 3:33
From “slop” to usable: tuning until the team trusts the output
He describes initial skepticism from the team and how repeated tuning improved quality. A concrete example shows the agent finding specific customer recordings about CAPTCHA solver issues, accelerating investigation and implementation.
- •Early drafts were rejected as low-quality (“slop”)
- •Quality improved through repeated tuning and iteration
- •Example: CAPTCHA solver strategy spec from customer complaints
- •PRD included links to specific recordings for engineers to reference
- 3:33 – 4:03
Agent workflow #2: daily content ideas and drafts from customer conversations
He explains an automated pipeline that converts recent customer calls into post ideas and draft social content. This creates a consistent publishing cadence without requiring daily manual ideation.
- •Daily email: post ideas from the last ~20 customer conversations
- •Topics bucketed into pain points, contrarian takes, high-performing angles
- •Drafts multiple posts (e.g., Twitter + LinkedIn variants)
- •Enables publishing five times per week with minimal time investment
- 4:03 – 4:18
Polishing and distribution: de-AI-ifying text and adding engagement hooks
The workflow includes a “de-AI-ifier” step to reduce obvious AI phrasing and optionally inject humor/memes before delivery. The final human review step remains, but the heavy lifting is automated.
- •Run drafts through a de-AI-ifier (e.g., Pangram) to reduce AI-isms
- •Attempt to include memes/engagement elements
- •Send to email for human review and approval
- •Human remains the final gate before publishing
- 4:18 – 4:34
Business impact: niche, customer-derived posts that drive inbound opportunities
Suchintan shares an example where an auto-generated post about a specific automation use case led to a direct referral from his network. The key benefit is capturing highly specific needs he wouldn’t have written about manually.
- •Auto-generated posts surface niche customer requirements
- •Example: carrier portal automation post sparked a referral
- •Demonstrates how specificity increases relevance and response
- •Content becomes a growth loop powered by customer reality
- 4:34 – 5:21
Operational changes to maximize context: default-public channels and record everything
He closes with concrete policy changes Skyvern made to preserve institutional context for both humans and agents. By eliminating DMs, recording all calls, and broadening access, they create a searchable company memory that compounds over time.
- •No internal DMs; questions go to an “office channel” for visibility
- •Record every call—internal and external (even founder 1:1s)
- •Give agents access to everything; give everyone access to an agent
- •New hires benefit from accumulated, searchable context
