Thanks, Claude.

The Problem

I couldn't keep track of what I'd promised people.

Not the work itself, but the overhead of managing it. Commitments buried in Slack threads. Follow-ups I promised in emails and then forgot. Context for meetings scattered across four different apps. Every Monday morning I spent an hour doing archaeology just to figure out what I'd said I would do.

I tried the usual fixes: better notes, task managers, calendar blocking. They helped. But they all required me to be the integration point, manually shuttling information between systems, remembering to log things, hoping I didn't miss something.

So I tried something different: what if I built an application that actually understood my work context? Yes, that dreaded vibe coding for some, dopamine god-send for others.

What I Built

I call it Second Chris (super original, right?). It ingests my Slack messages, Gmail, Google Calendar, Zoom transcripts, and Wrike tasks, then uses Claude to make sense of it all.

It's not a chatbot. It's my data, searchable and analyzed.

Here's what it does:

• When I tell someone "I'll send that over by Friday" in a Slack thread, Claude extracts that as a structured commitment with an owner, action, and deadline, and reminds me before I forget.
• It flags conversations where something seems stuck before it becomes a problem.
• Before meetings, it pulls relevant context from past emails and Slack threads with those attendees.
• It DMs me when something needs attention instead of waiting for me to ask.
• And I can just ask questions like "what did we decide about shipping those samples from NYC to Dallas?" and it searches across everything with semantic understanding.

The shift is from me hunting for information to information finding me.

How It Works For 35-75 Cents A Day or Less (If you’re into that)

The system has four layers. I wrote zero code, but I'm technical enough to ask just the right questions or be the human-in-the-loop where needed which I think gives me at least a leg up in this endeavor. I re-learned how to deploy, fight environment variables, and to be incredibly careful to make sure I didn't have runaway costs on services I was still learning how to orchestrate. To be clear, I was an expert in none of this, but my helper was patient, and taught me. I learned to ask better questions.

1. Ingestion. Separate connectors for Slack (Bolt SDK), Gmail and Calendar (Google APIs), Zoom, and Wrike pull data incrementally. Each source normalizes into a common format.
2. Storage. PostgreSQL holds structured data: messages, commitments, sync state. Pinecone stores vector embeddings for semantic search. I use OpenAI's text-embedding-3-small (1536 dimensions) for embeddings. It's fast, cheap, and accurate enough.
3. Intelligence. Different Claude models handle different jobs. Sonnet powers the main conversational responses with full RAG context. Haiku handles query expansion. When I search for "Samples from NYC to Dallas" it rewrites the query to include related terms before hitting the vector store. Commitment extraction uses structured prompting to parse natural language into items with deadlines.
4. Retrieval. This took more iteration than I expected. Pure vector search missed obvious keyword matches. Pure fulltext missed semantic connections. The solution: hybrid search combining both, re-ranked with time weighting. 60% semantic relevance, 40% recency with exponential decay over 30 days. Recent messages matter more than old ones.

Interface = Slack bot. I live in Slack, so that's where it meets me. Natural conversation for questions, slash commands for structured stuff like /priority or /prep. Runs on Railway, auto-deploys from main. About 10K lines of Python. I made a dashboard too in a day and realized, my problem wasn't the need to surface my projects, commitments, blockers, etc. I just needed information PUSHED to me, I didn't want to pull, observe, filter, etc. (But I can if I want).

The takeaway: this isn't one AI doing everything. It's Claude applied to specific tasks with good context. The work is in the retrieval, not the generation.

What I Learned

• Retrieval is harder than generation. Getting Claude to write good responses is easy. Making sure it has the right context is hard. Vector search parameters, ranking weights, thread awareness. These details determine whether you get a useful answer or a confident hallucination.
• Query expansion helps more than I expected. Using Haiku to rewrite queries before searching made a real difference. "What did Sarah say about the timeline?" becomes a richer query that catches messages where she said "schedule" or "deadline" instead.
• Proactive beats reactive. The biggest change wasn't asking questions faster. It was not having to ask. When the system DMs me "You committed to sending the proposal to Sarah by EOD and haven't yet," that's different from remembering to check.
• 80% accuracy beats 30% memory. The commitment extraction isn't perfect. But imperfect automation is better than my memory. I iterated in production instead of waiting.
• Build for your actual problem. I didn't start with "I should build something with AI." I started with "I keep forgetting what I promised people." That kept me focused.

Where I Think This Goes

I built this for myself, but I don't think it stays personal.

Every knowledge worker has the same fragmentation problem. Slack, email, calendars, project management: they're silos. We're the integration layer.

LLMs change things. Building a system that reads unstructured text, understands context, and takes action is now accessible to one engineer on weekends. That wasn't true two years ago.

I expect personal AI to become normal for knowledge workers. I expect companies to build internal AI that understands their specific context. I expect new workflows where humans focus on judgment and relationships while AI handles information synthesis.

The organizations that figure this out will have an advantage. I owe all of this to the magic of Opus 4.5 and Claude Code (and now, Claude Cowork too).

If You're Curious

I'm still building. Still learning. If you're working on similar problems, I'd like to hear about it.

I use this system, and others that I’m experimenting with, as fast as I have the ability to give them attention, every day. It changed how I work forever. My recommendation is to start playing NOW. Model subscriptions in the $200/mnth range for the tokens and compute available won't be subsidized forever. The price will go up, the access will tighten. Play and learn as fast as you can.