CHAPTERS
Why agents produce “slop”: helpfulness without business context
Suchintan frames the core failure mode of AI agents: they execute instructions eagerly but lack the nuanced context that makes output useful. He compares this to onboarding a new employee who needs time and information to become effective.
What “good context” means in practice: your company’s recorded knowledge
He defines context broadly as everything your company knows and has documented—communications, documentation, customer feedback, and operational data. The more the agent can reference real artifacts, the more grounded and actionable its work becomes.
Remote companies’ advantage: context is written down instead of spoken
Suchintan argues remote-first teams naturally generate better AI-ready context because conversations happen in recorded or written channels. In-person context often disappears in hallway conversations unless explicitly captured.
Two agent “skills” that speed up execution: PRDs and marketing
He previews two concrete workflows where agents create leverage for a lean team: drafting PRDs and generating content ideas. The unifying theme is retrieval from internal sources plus iterative review.
PRD agent workflow: retrieve evidence from calls, Slack, and Notion
The PRD agent starts with a topic and then searches across call recordings and internal tools to assemble a grounded first draft. The prompt is intentionally a bit vague to give the system flexibility in how it gathers and structures information.
Improving PRD quality with sub-agents: adversarial review and critique
To avoid shallow specs, Skyvern runs additional agents that review and comment on the initial PRD. This creates an internal “pushback” loop that mimics how strong teams refine requirements.
Cutting ‘junk requirements’ with prioritization (RICE framework)
After drafting and critique, the PRD is filtered through a prioritization framework (RICE in their case). This step helps remove low-value ideas the agent might include by default.
From “this is slop” to “actually usable”: tuning via team feedback
Suchintan shares that early PRD drafts were rejected as low-quality, but improved through repeated tuning and real user (team) feedback. The goal became producing drafts engineers would actually start from.
Concrete PRD win: diagnosing CAPTCHA solver issues with linked evidence
He gives an example where the PRD agent helped address customer complaints about CAPTCHA solving. The agent surfaced and linked specific recordings so builders could quickly verify context and understand the problem firsthand.
Daily content marketing agent: post ideas from recent customer conversations
A second workflow emails him daily with social post ideas based on the last ~20 customer conversations. The agent identifies recurring topics and formats them into draft posts that he can quickly review and publish.
Polishing and packaging: de-“AI-ifying” wording and adding memes
Before sending drafts, the workflow runs a style pass to remove obviously AI-sounding phrasing and tries to add humor via memes. While the tooling isn’t perfect, it improves publishability and engagement potential.
Business impact example: hyper-specific post generates a warm lead
He describes a real inbound opportunity that came from an auto-generated post about a niche use case (carrier portal automation). The specificity—sourced from customer requirements—made the content more likely to connect with someone who needed it.
Operational changes to capture context: no DMs, office channels, record everything
To maximize usable context, Skyvern changed company norms: questions happen in public channels, and all calls are recorded. This creates searchable institutional memory that benefits both humans and agents.
Democratizing agent access: everyone gets an agent, even new hires
He closes by emphasizing broad access: agents should be available to the whole company, not just engineers. New sales hires were required to set up tooling (e.g., cloud code) and quickly adapted, reinforcing that agent workflows can become standard operating practice.
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome