AI Maturity Check · Marina Havlis

← marinahavlis.pages.dev

Custom Skills

+3 vs June 1

Recurring Agents

consolidated from 8

On-demand Routines

stable

Monitored Jobs

new, off-machine

Strong Level 9, holding. This month the setup got leaner. Under pay-per-use, the recurring fleet was pruned to essentials. The four polling digests folded into one fixed-tick run; the 1-minute chatbot poller and four scheduled reports were switched off the same day. What stays on cron is the morning digest and the Monday business-review draft, now watched by an off-machine dead-man's switch that emails if any job goes quiet, even with the laptop off. Nothing runs that I'm not using, and if a job goes quiet I hear about it.

What changed since June 1

Under 3 weeks of work, shipped, with evidence

Pruned to essentials under pay-per-use. Four polling digests (inbox, chat, fallback, voiceover) folded into one fixed-tick morning digest. The 1-minute chatbot poller plus four scheduled reports (a daily category dashboard, a margin scoreboard, a page content watch, a monthly category funnel) were switched off the same day. The morning digest and the Monday business-review draft stay on cron; the coordinator poller stays always-on and logged a run on June 19. The margin-tracking and audit work stays available as on-demand routines.
Off-machine monitoring shipped and reconciled. A dead-man's switch on healthchecks.io emails the moment a job misses its check-in past deadline, from the cloud, so it speaks even when the laptop is off. The check list was reconciled to the post-consolidation fleet: 6 live jobs watched, 5 retired ones turned off (expected), none false-alarming. It catches the case an in-process check misses entirely: a job that never started. Closes the silent-failure gap the June 1 snapshot named.
Portability hardened. A migration runbook plus monitoring and publishing moved onto personal accounts make the whole setup vendor-independent, with no single corporate login in the critical path.
Public site reorganized. The landing page is now a personal index with a real 404, and the weekly refresh dropped a build step it no longer needs.
Three new skills. bye (clean session close with lessons capture), marina-voice (first-person drafting), and pixel-perfect-doc (document fidelity) bring the library to 28.

10-Level Maturity Map

Prompt & Hope

Type prompts and hope for the best.

Done

CLAUDE.md & Basics

Structured CLAUDE.md with preferences, shorthand, safety rules. Workspace plus per-project hierarchy.

CLAUDE.md hierarchy with skill-enforcement gates and 6 workspace rule files

Done

MCP Connectors

Connected external tools: Asana, Gmail, GDrive, BigQuery, Keboola, Tableau, Jira, Confluence, Figma, OpenMetadata, Google Chat.

11 live connectors, with governed data access

Done

Custom Skills

Repeatable one-command workflows: weekly status report, business-review commentary, meeting notes to Asana, analytics pulls, audits, unit-economics guides.

28 local skills, each encoding institutional knowledge

Done

Memory & Context Engineering

Persistent memory with glossary, people profiles, data-model context, voice rules, P&L framing, and lessons learned.

Memory consolidated into user, workspace, and project layers, mirrored to git for laptop-loss durability

Done

Multi-Phase & Chained Skills

Skills that gather from 4+ sources and synthesize: the weekly report pulls Asana, Calendar, Drive, and chat history. The daily dashboard chains the warehouse to Python to HTML to chat.

Weekly report, business-review fill-in, daily dashboard pipeline

Done

Headless & Scripted

Python scripts run independently and pipe structured output to external services. Dashboards and scoreboards generate without a human in the loop.

Production scripts piping JSON to chat

Done

Browser Control & Generation

Browser automation. PDF, PPTX, DOCX generation skills. HTML dashboard generation from live data, deployed to the cloud.

Browser agent plus document generation producing daily reports and content watches

Done

Parallel & Specialist Agents

Orchestration that fans out to specialist sub-agents on a fixed pattern, with review gates. Internal review and second-opinion passes run routinely.

A swarm skill (phased fan-out with review checkpoints) plus the Asana coordinator poller hold the orchestrator layer.

Done

Cron Jobs & Background Agents

A lean recurring core: the morning digest (daily) and the business-review draft (Monday) on cron, plus the always-on Asana coordinator poller, backed by 6 on-demand routines. Pruned from 8 agents once the platform went pay-per-use: the four digests merged into one and the chatbot plus four scheduled reports were switched off.

The morning digest and the coordinator both logged runs on June 19. An off-machine dead-man's switch (healthchecks.io) watches the 6 live jobs (5 retired checks turned off) and emails on any missed check-in, even with the machine off.

↑ FEWER AGENTS, WATCHED HARDER. Silent-failure detection, the June 1 gap, now shipped.

Strong

Autonomous Loops

Agents that observe and improve the system. Self-optimizing setup. Anomaly detection that triggers investigation on its own.

Advancing: the dead-man's switch is real anomaly detection on job health, and the monthly self-audit still takes actions (model switch, 13.9% context trim). Missing: skill authoring tested against evals, and a business-metric anomaly that auto-spawns an audit.

↑ FUTURE to NEXT. Job-health anomalies now detected off-machine; two loops remain: skill authoring tested against evals, and metric-anomaly auto-investigation.

What's still missing

To keep Level 9 solid and reach Level 10

Monitor the self-improvement routines, not just the daily ones. The monthly self-audit and weekly strategic-alignment runs are paused on the dead-man's switch and have not checked in since June 4, so nothing would alert if they silently stopped. Wire them to ping, then re-arm their checks. The self-audit is the engine behind Level 10, so it's the last job that should ever go unwatched.
Skill authoring with evals (Level 10). The monthly self-audit observes and acts. The next step is a routine that proposes or improves a skill and tests it against examples before adopting it.
Anomaly-triggered investigation (Level 10). The monitor now catches job silence. The remaining loop is a business-metric anomaly, a margin or refund move on the daily dashboard, that auto-spawns an audit instead of waiting to be asked.