A two-and-a-half-hour course rewrites how the team plans every project — and a couple of dashboards get fixed along the way.
Immediate, user-facing work. API endpoints, request/response, sub-second.
Background processing. Job tracking, status, anything that can wait.
Constant-time lookups. Caching, sessions, fast reads.
Large blobs with metadata kept separately.
Ordered messages, event-driven dispatch.
Structured data with relationships. ACID guarantees.
Similarity search. Embeddings. The AI substrate.
Now or later. Service or Worker. That's the entire surface area of how work gets done.
Because "where should this live?" has many right answers, each shaped by access pattern, scale, and consistency needs.
Design the architecture. Specify the blocks. Write the PRD. Strategic planning, system shape, the call on what goes where and why.
Implements what the PRD says. Translates blocks into code. Fast, literal, scales with the quality of the spec it's given.
"The planning phase becomes the hardest, most critical part. Implementation falls out clean."
Is this about doing work, or storing data?
What are the specific access patterns and performance requirements?
Answer those well, and the right blocks fall out. Hand the blocks to Claude and the implementation is a routine translation. Skip them, and you're vibe-coding in a more expensive chair.
Same systems we already shipped — now seen through the framework.
Blocks: Relational Key-Value
Sprints and tasks are structured data with relationships — Postgres territory. Hot reads for the dashboard view get a Redis cache so we don't hammer the relational store every render.
Blocks: Vector
Already running similarity search on embeddings. The pattern was always there — we just didn't have the name for it. Now we do.
The HelpFlow org repo is the source of truth for the team, but the account that pays Vercel can't write to it. helpflow-dev (personal) is the bridge — it's what Vercel actually pulls from. Until permissions get reshuffled, this is the path.
The B Premium client gave a 30-day shutdown notice, and Veneers itself is winding down — not switching providers, not pausing, fully closing. Their bot was basic relative to what we build now, so there's nothing to migrate or salvage. We let it run out the clock.
PST · client website launch
A dev gets dropped into a temporary Slack channel for the launch window. The client's own dev team owns the website. We're there for one thing.
GTM-related issues only. Website bugs go back to their dev team. Hold the line.
The individual completion-rate filter no longer requires a Sprint selection before Assignee or Month can be applied. Each filter is now standalone. Requested by Shay · implemented by Micah
Reveals once at least one filter is active. One click, fresh slate. Small change, big quality-of-life. Implemented by Micah
— end of edition —