Ray Weekly Update, 2026-04-25

Week of 2026-04-19 → 2026-04-25

Hey Ray. Most of the client work this week was with Mark on Orbiter. Most of the system work was on the head-screen, the operator UI for snappy-os.

Metrics

Orbiter, Mark, the Henry Pack push

Mark and I had four working sessions across Monday, Wednesday, and Thursday. The arc: load 2,000 LSI conference attendees into the knowledge graph, run targeted Cypher queries against that subset, generate a polished prospect report and pitch deck for Henry Pack (LSI Star 51 fund, ex-medical engineer, anchored by Mayo Clinic, actively raising LPs in medtech).

Mon, LSI Conference Data and Strategy. Mark walked me through the queuing system that turns a 500K-person fundable dataset into graph nodes plus a CRON-prioritized enrichment queue. We scoped the dog food test to the 2,000 LSI attendees, defined five use cases, and committed to a first report for Henry by Wednesday.

Wed afternoon, Dog Food Test. Reviewed the first pass. Pulled out work: replace company tiles with executive avatars, swap placeholders for real LinkedIn profile images, A/B Kimi 2.6 against Opus on two report entries for tone, add a hero image, package the Gemini API setup as a scope file Mark can drop in with his own key.

Wed evening, Henry Pack Proposal and Data. Reviewed the proposal document. Structural feedback on the "why" sections. Decision to keep the working draft over-explanatory and produce a clean client-facing version separately. A cross-referencing gap surfaced between the LSI table and the Fundable People table, queued for me to chase. Confirmed the new GCP/AlloyDB vector index work starts the Monday after, with Xano staying as the R&D playground.

Thu, Deck Review. Tone pass on the V2 Star 51 deck. AI-cliché phrasing pulled out in bulk with a Python filter. Replaced "Claude Opus 4.7" references with "Orbiter" so the output reads as proprietary. Split the deck into two PDFs (Luke and Caitlin Morse intros, Star 51 outcomes). Final delivery to Mark Thursday evening.

Billable on this work: $1,500. Invoice not yet issued.

Total CRM, Jordan on vacation

No new Total work touched my desk this week. Jordan back next week, the Total MCP project files are queued from the 2026-04-17 sync.

snappy-os, head-screen progress

Most of my non-client time went here. What's in place this week:

Home shows desks, not lanes. A desk is a category that can be renamed. Each row inside a desk is an agent, which is a frozen prompt plus its loaders and an optional skill. Desks are user-created, so a new one can be added without touching the code.

Home view: SNAPPY OPERATOR with the desks Sweep, Content, LinkedIn, Meetings, Pods, plus a live dispatch panel and trust ledger.
Home view: SNAPPY OPERATOR with the desks Sweep, Content, LinkedIn, Meetings, Pods, plus a live dispatch panel and trust ledger.

The brain panel is a real link-out. When an agent hits a gap in its loader during a turn, it logs a row to a feedback queue, the row gets approved or discarded, the loader gets patched. The brain page shows the loader inventory plus the last twelve feedback rows with concrete fixes (linkedin timeout, folder-shape fallback, prose-only sweep).

Brain view: full loader inventory with a feedback panel of approve, not-yet, and wrong actions on recent loader patches.
Brain view: full loader inventory with a feedback panel of approve, not-yet, and wrong actions on recent loader patches.

Clients view shows the active client modules (Orbiter, Pulse, Ray, Scott, Total) as cards with last-run and a recent score. Your row is the only one showing activity this week because the rest of the roster is on a slower cadence.

Clients view: cards for Orbiter, Pulse, Ray, Scott, Total. Ray's card shows recent activity, the others show idle.
Clients view: cards for Orbiter, Pulse, Ray, Scott, Total. Ray's card shows recent activity, the others show idle.

Trust ledger rolls recent skill runs into a green/yellow/red strip. Click a row, jump to the skill page. The strip shows recent runs, their scores, and which skill they belong to. It is the audit surface that proves the system is doing what it says it is doing.

Trust ledger: recent skill runs as a pass/weak/fail strip. Each row shows skill, score, verb, and time, with click-through to the skill page.
Trust ledger: recent skill runs as a pass/weak/fail strip. Each row shows skill, score, verb, and time, with click-through to the skill page.

Skills index is the real catalog read off disk. Every row carries category, loader presence, sidecar presence, 7-day run count, and 7-day mean. Filter by category, click into a skill, see its anatomy, see its rubric. The catalog is now sourced from the filesystem instead of being a list I kept in my head.

Skills index: full catalog read off disk, with loader and sidecar columns, 7-day run count and mean per row. Filterable by category, top rows are meeting-triage, contract-test, polish-digest.
Skills index: full catalog read off disk, with loader and sidecar columns, 7-day run count and mean per row. Filterable by category, top rows are meeting-triage, contract-test, polish-digest.

Ambient mode is the screensaver state. When the operator goes idle the screen flips to a quiet "still up?" face with the ready-for-you queue. Wake word is "snappy". Running locally on the dev machine this week.

Ambient view: large hand-lettered "still up?" face, "ready for you - 2 items" panel, mode listening, wake word "snappy".
Ambient view: large hand-lettered "still up?" face, "ready for you - 2 items" panel, mode listening, wake word "snappy".

The structural work that made all of this possible:

skills.snappy.ai landing: "Treat markdown like code" hero, a skill is one .md file with YAML frontmatter, optional siblings api.ts, state/bin/, .agents.md, plus an install/pull/doctor command card.
skills.snappy.ai landing: "Treat markdown like code" hero, a skill is one .md file with YAML frontmatter, optional siblings api.ts, state/bin/, .agents.md, plus an install/pull/doctor command card.

Action item status from last session

  1. FreshBooks daily pull into the business-tracking Sheet: still not shipped. Cron slot ready. Lands by Wednesday.
  2. Older-client outreach (Sylvan, Benoit, Sophie, Jim, Mallay): drafts staged from 2026-04-18, still not sent. Several recipients are on travel this week. Sending Monday morning lands replies on the right end of the week.
  3. Referral asks for Jordan, Mark, Scott: drafts staged. Holding Jordan's until he is back from vacation. Mark and Scott go out Monday.
  4. Send invoices: Mark/Orbiter $1,500 from this week's Henry Pack push goes out Monday.

Closer

One question for the call: should the desks page be the first thing a new client sees, or should that be the brain panel where loader feedback lives. I have arguments for both and I would rather pick on a call than guess.

Looking forward to our call.