Capability Scan — First Run

What you HAVE vs. the pain you keep HITTING — and the wirings that close the gap
Generated: Tuesday, May 26 2026 · 2:15 PM EDT
This is: the first run of the self-learning routine you asked for (answers SCOPE_BACKLOG 38-K7 "no real self-learning loop")
Inputs: live MCP inventory · skill catalog · local scripts · SCOPE_BACKLOG friction (sessions 36-39) · this session's friction

The one-liner

You have a lot of capability that isn't wired to pains you keep hitting. The biggest unlock isn't a new tool — it's the scheduler: you keep asking for things to "fire on their own" (morning brief, this scan, runway), and the schedule skill + your existing cloud-routine pattern already do that — they're just under-used. Second biggest: three connectors are dead weight (monday/supabase need auth, circleback is failing) — wire or drop them so the inventory stays honest.

Top move: stop hand-running build_index.py + hand-firing briefings — schedule them. Details in §4.

1What you have

CapabilityStateUsed for what today
Gmail — business (claude.ai connector)OnEden/HOA/Mildred threads, drafts, sends
Gmail — personal (gmail-personal, local MCP)On (new today)Just wired — ztreitel@ now readable directly (retires the ChatGPT relay)
Google Calendar (connector + bot CalendarApp)OnNow in the morning brief (added this session). Desktop sessions can read it too.
Google Drive (connector)OnReading MIS/obligations sheets; the future Drive-as-memory migration (39-7) rides on this
Cloudflare + Netlify (connectors)Onops portal deploy, LevSMS site
monday.com (connector)Needs authNothing — never authenticated. Are you using monday? If no, remove it.
Supabase (connector)Needs authNothing yet — but it's the backend for the 39-5 native/multi-user app idea
Circleback (connector)FailingMeeting-notes tool — failing to connect. Fix (useful for Eden mediation calls) or drop.
schedule skill (cloud cron routines)Under-used2 routines live (Weekly Review Sun, Mon-Eve plan). Could fire the morning brief, this scan, runway.
loop skill (interval re-run)UnusedPolling/iterating tasks in-session
hookify skill (behavior rules)UnusedCould codify recurring traps so they never bite twice (see §5)
claude-md-improver / revise-claude-mdUnusedYour CLAUDE.md + CONTEXT are stale (38-K2/K3) — these fix exactly that
Local scriptsOnnow.ps1 (date), notify.ps1 (SMS), inbox.ps1 (bot read/write), setup-gmail-personal-mcp.ps1
hookstreet-skills (your own)Onsession-operator, command-inbox-engine, mis-daily-report, underwriting-method, +10 more

2The friction you keep hitting

Recurring pains pulled from SCOPE_BACKLOG (sessions 36-39) + this session:

Recurring painEvidence
Things don't fire on their own — briefings, scans, reviews need you (or me) to hand-triggerMorning spine "paper-only" gap; you've said "I want it to self-learn and offer things"; 38-K7
Portal index lags — new outputs deploy but don't show until build_index.py is hand-run38-K6; I ran it manually ~6× this week
Self-capture broken — .remember/now.md writes empty38-K5; 38-K7
Same setup traps bite twice — .json.json filename, em-dash breaking PowerShell, clasp --forcethis session (all 3 cost minutes)
Docs go stale — CLAUDE.md / CONTEXT lag sessions behind38-K2, 38-K3
Money picture isn't live — cash-IN vs OUT not wired into the brief39-6 runway layer
Two inboxes, one assistantSOLVED today (gmail-personal)

3Dead weight — wire or drop

Honest inventory means cutting noise. Three connectors do nothing:

4Recommended wirings — ranked

1Schedule the things you keep hand-firing

You own the schedule skill + an existing cloud-routine pattern (Weekly Review, Mon-Eve plan already fire). The morning brief now has rich content but still depends on a 6:48 AM Apps Script trigger; this scan, the runway check, and the KB-health-check all run only when triggered by hand. Wire them as cloud routines.

Solves: "things don't fire on their own" · 38-K7 self-learning loop · the morning-spine gap. Effort: low (the pattern exists). Your call: which cadence per item.
2Auto-generate the portal index on push

Stop hand-running build_index.py. Make it a git pre-push hook (or a Cloudflare build step) so outputs/index.html regenerates from the folder automatically every push. The portal never lags again.

Solves: 38-K6. Effort: ~20 min (pre-push hook). I can build this now — it's local + reversible.
3hookify the setup traps

Three traps cost minutes this session (.json.json, em-dash-breaks-PowerShell, clasp push needing --force). The hookify skill turns "mistakes I made" into guardrails that auto-warn next time. Codify these three + the "write dedup-date AFTER send" MIS rule.

Solves: "same traps bite twice." Effort: ~15 min. This is literally self-learning in action.
4Fix the .remember writer

.remember/now.md writes empty (38-K5) — the auto session-state capture is broken, which is half of why there's "no self-learning loop." Fix the writer so each session leaves a real breadcrumb.

Solves: 38-K5 + 38-K7. Effort: needs a look at the SessionStart hook. Diagnose first.
5Run claude-md-improver on the stale docs

CLAUDE.md (Session 37 stale) + CONTEXT.md (Session 36 + self-noted gaps) lag reality. You have a skill built exactly for this. Run it as part of the monthly cadence.

Solves: 38-K2, 38-K3. Effort: low. Good monthly-routine candidate.

5Gaps — pains with NO capability yet

6How this routine recurs (so it self-learns)

Saved as docs/CAPABILITY_SCAN.md — a repeatable procedure any Claude session can run: inventory MCPs/skills/scripts → mine SCOPE_BACKLOG + .remember + recent outputs for friction → match capability to pain → rank → report. Cadence options (your pick):