How I AII built a custom Slack inbox. It was easier than you think. | Yash Tekriwal (Clay)
CHAPTERS
- 0:00 – 4:05
150 Slack pings a day: separating real work from FYIs
Yash describes waking up to 100–150 Slack notifications daily and why that volume creates anxiety. The key insight: most notifications are informational, and the real goal is to surface the ~30–40 that truly need attention.
- •Gets 100–150 tagged/DM notifications daily, not just unread channels
- •60–80% are FYI; only a minority require action
- •Slack treats all notifications as equally urgent
- •Goal: reduce anxiety by triaging what actually matters
- 4:05 – 6:06
Designing the “better world” workflow before prompting AI
Instead of asking an AI to “make Slack easier,” Yash starts by defining what an ideal notification system would look like. He proposes a two-level categorization: first by Slack message type, then by actionability.
- •Start with the target workflow, not a vague AI request
- •Primary buckets: DMs, group DMs, threads, @mentions
- •Secondary buckets: action required, need to read, FYI
- •Outcome: a system aligned to how he personally processes messages
- 6:06 – 6:36
AI vs deterministic code: what to automate with models (and what not to)
Yash and Claire discuss a practical framework: use AI where judgment is needed (e.g., classifying intent), but rely on deterministic code when APIs and structured rules exist. This framing guides how he builds his Slack system.
- •Deterministic code is ideal when APIs expose reliable signals (timestamps, read states, etc.)
- •AI is best reserved for fuzzy tasks like categorization/summarization
- •Avoid over-using models when rules can solve the problem faster/cheaper
- •A “systems thinking” approach beats endless reprompting
- 6:36 – 11:51
Prototyping a Slack digest bot with OpenClaw (and why Discord helps)
Yash shows his first implementation: a Slack digest created via OpenClaw with a long back-and-forth to reverse-engineer Slack’s notification logic. He explains why Discord’s threading/search makes it a better environment for iterating on agent conversations.
- •Built a digest by reasoning through Slack’s timestamps/unread logic
- •Pulled only relevant context (e.g., latest messages in a thread) to avoid overwhelm
- •Discord chosen for threading and quick search to manage long build conversations
- •End result: digest grouped into the four main Slack buckets
- 11:51 – 12:21
Why the digest still felt draining: text summaries vs usable interfaces
Even after improving prioritization, Yash finds scrolling through a giant digest message exhausting. That motivates the next step: turning the digest into an interactive, “Superhuman-style” interface.
- •Digest required multiple screens of scrolling even after summarization
- •Better than raw Slack, but still cognitively heavy
- •Desire for navigability, sorting, and a clean UI
- •Shift from “summaries” to an actionable command center
- 12:21 – 14:29
Building the Kanban Slack inbox in Perplexity Computer
Yash demonstrates how Perplexity Computer quickly turns his digest into a dashboard: a three-column Kanban for action required, need-to-read, and FYI. The killer feature is bulk archiving FYIs that also clears Slack notifications.
- •Kanban columns: Action Required (red), Need to Read (yellow), FYI (green)
- •Sidebar controls to view by DM/group mention/thread/@mention
- •“Archive All” on FYIs removes them from the dashboard and Slack notifications
- •Transforms Slack from a flood into a triaged workflow
- 14:29 – 16:29
Why Perplexity Computer beats Claude Code/Codex for this use case
Yash outlines three core advantages: parallel task execution, cloud-native operation, and strong connectors. He also highlights Perplexity’s multi-model orchestration that reduces the painful reprompt/debug loop.
- •Runs multiple long tasks concurrently instead of single chat threads
- •Cloud-based agent is more fluid with web apps and context
- •Multi-model “ensemble” orchestration selects models per subtask
- •Fewer ‘try again’ cycles; more autonomous troubleshooting
- 16:29 – 19:53
Connectors in practice: meeting follow-ups across Notion and Asana
Using Perplexity Computer’s connectors, Yash describes automating meeting follow-ups: pulling transcripts from Notion, extracting action items, filing longer tasks into Asana, and drafting messages/emails. Claire highlights how broad his connector setup is.
- •Pull meeting transcripts from Notion AI and extract action items
- •Categorize importance; send long-term items to Asana
- •Draft responses for items that are “just a message/email”
- •Connector stack includes Gmail/Calendar, Notion, Asana, Slack, Drive, Airtable, more
- 19:53 – 24:08
The “micro-software” future: SaaS isn’t dead, it’s getting personalized
Claire and Yash zoom out: rather than replacing Slack, people will build lightweight layers on top of existing SaaS to match individual workflows. They predict a long-tail boom of small, paid tools that don’t need venture-scale outcomes.
- •Slack is great for sending, not always for reading/prioritizing
- •Personal layers (“Slack custom”) can close the last-mile workflow gap
- •Many niche requests were historically “never on the roadmap”
- •Low build cost enables profitable micro-businesses and acquisitions
- 24:08 – 29:49
Anti-to-do lists and building personal command centers (news + email + Slack)
Claire introduces the “anti-to-do list” as a way to identify high-leverage automations you never want to do manually again. Yash shows another Perplexity Computer app: a consolidated digest UI combining AI news, email, and Slack, refined through iteration.
- •Anti-to-do list: automate tasks you never want to do again (spam cleanup, meeting follow-ups, etc.)
- •Yash consolidates AI news, email, and Slack into a single UI
- •Iterative approach: get data right → build UI → improve usability (deep links)
- •Idea: present daily briefings in more readable “newspaper” formats
- 29:49 – 32:25
Auth + deployment: why Computer makes sharing and maintenance easier
Yash explains that Perplexity Computer apps are already deployed and shareable, removing the need to push repos and deploy to Vercel for simple internal tools. Connector authentication is reused automatically, and the agent can detect and fix auth issues.
- •Apps are instantly live on the web with a share button
- •No separate deployment pipeline for quick internal tools
- •Connector auth is reused across the app and the agent session
- •System warns about auth failures and can guide reauthentication
- 32:25 – 35:57
Team use case: prototyping persona-based Clay University journeys
A teammate uses Perplexity Computer to prototype a redesigned Clay University experience with persona-based learning paths. The key value is bridging the communication gap between stakeholders and designers by producing a visual, context-aware prototype fast.
- •Challenge: existing content isn’t tailored to personas/roles
- •Computer can “see” the current site in-browser and prototype against it
- •Prototype includes persona journeys (e.g., SDR/BDR/RevOps) and a logged-in view
- •Creates a faster, clearer handoff artifact for the design team
- 35:57 – 44:36
Lightning round: fun AI uses, debugging tactics, and where to find Yash
Yash shares his main “fun” AI use: brainstorming games and logistics for friend Olympics and board-game nights. He also reveals his strategies for when models fail—better context, strict prompting, and refining skills over time—then closes with where viewers can connect.
- •Uses AI for brainstorming party/game ideas and operational planning for events
- •Admits he doesn’t use AI as a “therapist,” mostly as a work tool
- •Debugging: recognize when AI shouldn’t do a task; add missing context (timestamps)
- •Prompting tricks: be strict/explicit; build/refine reusable skills iteratively
