What 2 Months of AI-Assisted Work Actually Produced

€214 total. 400 hours. One AI coding assistant.

Here's everything that was built — with the numbers verified from our analytics and project systems.

The Deliverables

What	Numbers
Automation workflows (n8n)	45 active
LinkedIn posts published	91 total (across 2 accounts)
Newsletter subscribers	644
Cold outreach leads discovered	638
Emails sent (outreach)	616
GitHub repositories managed	23
Production skills/commands	20
Operational rules	6
Session summaries documented	41 (in first 6 days of February alone)
Lessons learned documented	19 incidents
Documentation health score	8.5 → 9.5/10 (21 stale docs fixed)

The Systems

Beyond individual deliverables, several complete systems were built:

JUVENAL — A knowledge management system with 20 production skills, 6 operational rules, persistent memory across sessions, and automated daily standups. It reads the entire project context and picks up where the previous session left off.

Analytics Hub — 5 database tables, 2 views, 3 automated collection workflows. Pulls data daily from newsletter, LinkedIn, cold outreach, websites, and automations into a single dashboard.

Content Hub — A planning system where each content theme adapts across all channels (newsletter, LinkedIn, Instagram, Reddit). One source of truth, five weeks of content planned at once.

Cold Outreach Pipeline — Lead discovery, email composition, sending, follow-ups, bounce detection. 638 leads discovered, 616 emails sent, dual A/B testing across two industry pipelines.

The Strategy Decisions

The AI wasn't just an implementer. It helped analyse data and inform strategy decisions:

Decision	What Changed
ListingForge demotion	Zero-usage product removed from portfolio
4-tier pricing	€750 to €15,000 service tiers finalised
Dual pipeline A/B test	Web design vs AI/automation cold outreach
Content Hub	Scattered planning → unified themes × channels
Git branching policy	Absolute rule: never commit to main

The Refactoring

One standout session: a monolithic Python file with 4,824 lines and 40 API routes was decomposed into 24 files — 11 functional blueprints, 10 shared components. All routes preserved. Code review found 3 issues, fixed in the same session.

The Failures

19 documented incidents. The highlights:

• Double newsletter send — two workflows with the same trigger, old one not deactivated
• Tracking broken for a month — privacy setting changed, container not restarted
• 14 workflows offline 12 hours — API key rotated, 14 hardcoded references not updated
• SQL injection via single quotes — apostrophes in post text broke webhook SQL
• Zombie process killed OAuth — stale process on port 8000 blocked new sessions

Each incident was documented with root cause, fix, and prevention in our lessons-learned system.

The Real Assessment

The AI multiplied capacity. Including the capacity to break things at scale.

What would have taken 2-3 months alone took 1 month with Claude Code. The cost — €107/month — is trivial compared to the output. But the first two weeks were frustrating: learning to communicate intent clearly, learning when to trust and when to verify.

The skill isn't using AI. The skill is knowing when not to trust it.

---

All numbers from verified ImparLabs sources: Analytics Hub, Listmonk, session logs, and the JUVENAL knowledge system.