I built a custom Slack inbox. It was easier than you think. | Yash Tekriwal (Clay)
Yash Tekriwal, Head of Education at Clay, demonstrates how he built a custom Slack inbox management system using Perplexity Computer to transform overwhelming notification overload (100–150 daily messages) into a prioritized, actionable dashboard. The episode explores a practical framework for using AI agents to prototype deterministic tools (code + APIs) versus using AI for dynamic categorization tasks, and argues that the falling cost of software creation will spawn a "Cambrian explosion" of small, niche productivity tools built on top of existing platforms like Slack.
Key takeaways
- • Categorize notifications by urgency and type, not by treating all mentions equally: Tekriwal uses AI to bucket messages into "action required," "need to read," and "FYI" categories, reducing his 100–150 daily notifications to roughly 30–40 that actually demand attention.
- • Use AI to build deterministic tools with APIs (like fetching unread Slack messages via structured code) rather than relying on AI to repeatedly make the same decision: The Slack digest uses code to pull messages and timestamps, with AI only categorizing urgency—not regenerating the entire system each time.
- • Perplexity Computer's key advantages over competitors are multi-model orchestration (using Sonnet, Gemini, and Opus in parallel for different subtasks), cloud-native connectors (no manual OAuth setup per use), and instant web deployment—removing friction that persists in Claude or Codeex.
- • Build custom UIs and dashboards for existing tools using Kanban-style interfaces to make information digestible: Tekriwal's Slack digest, news/email consolidator, and Clay University prototypes all use visual boards with filtering and deep-linking back to source messages.
- • The future of productivity SaaS isn't replacement—it's composable, vertical-specific extensions: Instead of building "Slack 2.0," founders can charge $15/month for specialized tools that sit on top of Slack, unlocking a long tail of useful software that was previously too expensive to build.
- • When AI repeatedly fails at a task (e.g., date handling), refine the system prompt or skill definition incrementally over 10–20 iterations rather than starting over—and make the stakes explicit (frame requests with urgency, use caps, describe negative consequences).
Recommendations (10)
"What I sort of started doing with Perplexity Computer when it came out about a month ago is thinking if I could truly just design any software or like paradigm myself, what would I do, how, and why?"
Yash Tekriwal · ▶ 4:05
"I have a Google Calendar skill that I'm constantly refining and iterating"
Yash Tekriwal · ▶ 42:22
"Google Drive, Gmail with Calendar, Notion, Asana, Slack, Forms, Tasks"
Yash Tekriwal · ▶ 17:44
Mentioned (10)
More from these creators
Claude Code + 15 repos: how a non-engineer answers every customer question | Al Chen
Claude Code for personal productivity: a step-by-step tutorial | Hilary Gridley
How Stripe deploys 1,300 AI-written PRs per week
How Microsoft's AI VP automates everything with Warp | Marco Casalaina
From journalist to app developer using Claude Code
How Figma engineers sync designs with Claude Code and Codex