How I AII built a custom Slack inbox. It was easier than you think. | Yash Tekriwal (Clay)
CHAPTERS
Meet Yash Tekriwal: hyper-optimizing work with Perplexity Computer
Claire Vo introduces Yash Tekriwal (Head of Education at Clay) and frames the episode around building custom tools with AI—specifically a better way to process Slack overload. The promise: a full tutorial plus a broader debate about AI-built micro-software and the future of SaaS.
- •Who Yash is and why he cares about workflow optimization
- •Episode goal: build a custom Slack inbox/digest using Perplexity Computer
- •Theme: AI as a leverage tool for personal productivity and prototyping
- •Tease of SaaS/micro-software discussion
The real problem: 100–150 Slack notifications a day (and most are FYIs)
Yash describes waking up to 100–150 tagged notifications daily, which creates anxiety and makes everything feel equally urgent. His key insight is that 60–80% are FYI noise; only ~30–40 truly require action, so the goal is triage and prioritization rather than “read everything.”
- •Slack treats a critical DM and a fun channel ping as equally urgent
- •Yash wants a system that reduces anxiety by surfacing only what matters
- •Target state: separate action-required vs. read vs. FYI
- •Need to bucket by message type (DMs, group DMs, threads, @mentions)
A practical framework: when to use AI vs. deterministic code
They distinguish between using AI to directly do tasks (summarize/categorize) versus using AI to help build deterministic systems with APIs. Yash argues Slack is a good candidate for deterministic logic because timestamps and message states are structured, while AI should be reserved for the ambiguous classification layer.
- •Two AI modes: do-the-work vs. build-the-tool
- •Slack APIs + structured timestamps enable deterministic filtering
- •Systems thinking: what changed since last view, which context to pull
- •Use AI mainly where judgment is needed (action/read/FYI classification)
Prototype v1: building a Slack digest using OpenClaw (in Discord)
Yash shows an early implementation built with OpenClaw, iterating through a long back-and-forth to reverse-engineer Slack notification logic and produce a daily digest. It works—but outputs a long scroll of text that still feels draining to process.
- •OpenClaw used as a coding/automation agent to build the digest pipeline
- •Slack notification nuances: unread logic, threads, last-seen timestamps
- •Digest groups messages into four buckets (mentions, DMs, group, threads)
- •Only repeated AI usage is the triage labels (action/read/FYI)
Why Discord for agent work: threads, search, and context management
Claire notices Yash is running OpenClaw in Discord and asks why. Yash explains that Discord’s threading and quick search (Command+K) make it easier to manage many parallel “projects” and keep context clean compared to other chat surfaces.
- •Threaded organization for long-running builds and debugging
- •Fast retrieval of past context via search/Command+K
- •Ability to archive/close threads to reduce clutter
- •Separating agent-building workflows from daily Slack noise
The UX problem: a digest that works… but is still overwhelming
After using the digest for a week, Yash realizes that even a well-structured text feed requires excessive scrolling and context switching. He wants a ‘Superhuman for Slack’—a clean interface that matches his mental model and supports fast triage.
- •Text digests reduce noise but can still be cognitively heavy
- •Scrolling through multi-screen summaries is fatiguing
- •Desire for navigability, filtering, and one-click batch actions
- •Motivation to move from prototype script to UI-first workflow
Building the visual dashboard in Perplexity Computer (multi-model agent workflow)
Yash switches to Perplexity Computer to turn the digest into an interactive dashboard. He highlights Perplexity’s orchestration advantage: it can use different models for different steps, run troubleshooting loops, and progress from prompt to working app with fewer manual reprompts.
- •Perplexity Computer can ingest context via browser + connectors
- •Ensemble approach: different models for planning/coding/reasoning
- •Less ‘it doesn’t work, try again’ reprompting due to better orchestration
- •Rapid progress—large portion built in first few messages
Three reasons Perplexity Computer beats Claude Code for this use case
Yash outlines why Perplexity Computer feels superior for his workflow: parallel task execution, cloud-native agent behavior, and smoother integration with everyday tools. The net effect is faster iteration, less setup friction, and more “speed of thought” building.
- •Concurrency: run multiple tasks in parallel without juggling terminals
- •Cloud-native: easier access to web tools and long-running tasks
- •Connectors feel more fluid for non-coding workflows, not just app generation
- •Recorded/trackable work and easier reuse across tasks
Connectors in action: automating meeting follow-ups (Notion → Asana + drafts)
They explore how connectors enable cross-app automations beyond Slack. Yash describes pulling meeting transcripts from Notion, extracting action items, routing longer-term tasks to Asana, and drafting responses for messages/emails—turning meetings into structured follow-through.
- •Perplexity Computer connected to tools like Notion, Asana, Slack, Gmail, Drive
- •Workflow: scan transcripts, extract action items, prioritize, and route tasks
- •Draft outbound responses when an action item is communication-based
- •Focus on practical value vs. “shiny object” connector collecting
The Kanban-style Slack dashboard: action/read/FYI + one-click ‘Archive All’
Yash demonstrates the core deliverable: a Kanban-style board with three columns—Action Required (red), Need to Read (yellow), and FYI (green). The standout feature is batch archiving: clearing FYIs from the dashboard also clears those notifications in Slack.
- •Three-column triage board aligned to Yash’s workflow
- •Additional grouping controls (DMs vs group mentions vs threads, etc.)
- •Deep links let him jump into the source message quickly
- •Batch ‘Archive All’ solves Slack’s all-or-nothing read state problem
The long tail of customer requests: micro-software, SaaS, and “Slack custom”
Claire and Yash discuss how AI lowers the cost of building niche workflow extensions that SaaS products won’t prioritize. They predict a “Cambrian explosion” of small, useful apps—often non-venture-scale—where individuals prototype solutions and creators package them cheaply ($15/month) or get acquired.
- •SaaS isn’t dead; customization layers on top of strong cores will grow
- •AI makes niche features economically feasible for small builders
- •Many valuable requests were historically “never on the roadmap”
- •Emergence of maintainable micro-products with small TAMs but real revenue
The anti-to-do list: systematically eliminating recurring pain
Claire introduces an “anti-to-do list” framework: write down tasks you never want to do again and invest time in automating them. Yash agrees, positioning Perplexity Computer as a practical way to burn down these recurring annoyances (Slack triage, email spam, meeting action items).
- •Identify high-friction recurring tasks to automate
- •Spend consistent time turning pains into systems
- •Examples: unprioritized Slack triage, manual spam deletion, manual task entry
- •AI tools as leverage to reduce life/work admin overhead
Building a consolidated digest: news + email + Slack as a command center
Yash shows a second dashboard that combines AI news, email, and Slack into a single prioritized view. He explains an iterative build loop: first get the right data, then add UI, then improve usability with deep links and interaction patterns.
- •Start with text output to validate what’s included and why
- •Iterate on categorization rules by interrogating model decisions
- •Add UI (often Kanban by default) once content is right
- •Improve usability: deep links back to source messages and emails
Authentication, deployment, and sharing: why cloud agents remove friction
They dig into the operational advantage of Perplexity Computer: apps are already deployed and shareable, and connectors reuse existing authentication rather than requiring repeated token setup. The agent can also detect auth issues and prompt reauthentication or use in-browser proof-of-concepts.
- •No separate hosting/deployment pipeline (e.g., GitHub/Vercel) for prototypes
- •Connector auth is reused across apps built in Computer
- •Agent warns on broken auth and can help reauthenticate
- •Shareable apps via a single link reduce collaboration overhead
Team use case: prototyping persona-based learning journeys for Clay University
Yash shares a teammate’s project: using Perplexity Computer to redesign Clay University into persona-based learning paths (SDR/BDR, RevOps, Marketing Ops, etc.). The key value is “visual bridging”—a functional prototype that helps cross-functional teams align without heavy design tooling overhead.
- •Computer can “see” the existing site in-browser and use visual context
- •Rapid prototype of logged-out persona journeys and logged-in dashboards
- •Helps communicate requirements to design teams with less translation loss
- •Demonstrates that Computer isn’t just personal productivity—it's team prototyping
Lightning round: fun AI uses + how to recover when models fail
In closing, Yash shares personal uses (brainstorming games and event ops) and candid strategies for when AI performs poorly. His tactics include adding missing context deterministically (timestamps), being extremely explicit/strict in prompts, and iteratively refining reusable skills for recurring failure modes (e.g., calendars/dates).
- •Fun use: brainstorming party/Olympics games and optimizing team rotations
- •AI as a starting point—humans still do final taste/edits
- •Debugging strategy: decide what should be deterministic vs. model-based
- •Iterate skills over many turns; ask the model to explain failures and propose fixes
