Lantana: With a Little Help of My (Claude) Friend

From beta to v1 in 2 months

2026-07-15

Agenda

TODAY

1. Why Lantana — the honeypot-as-code motivation
2. 84 years ago, in 2025 — the pre-modern AI workflow
3. With Claude — what actually changed
4. Caveats & takeaways — what to try Monday

Why Lantana

MOTIVATION

I’d deployed two SIEMs. I wanted to reason about the log pipeline end-to-end — without building fake log sources.

So I built a Honeypot-as-Code platform.

Cowrie and Dionaea, wrapped in Terraform, Ansible, and systemd. Fewer decoys, stronger narratives, open pipeline.

The point of the “why”: this wasn’t a honeypot project pretending to be an AI project. It was a real thing I needed for pipeline work, and AI happened to change how I built it. Lantana doesn’t reimplement honeypots — Cowrie (SSH/Telnet) and Dionaea (HTTP, MySQL, SIP…) already exist and are solid. Lantana is the IaC + pipeline that wraps them: Terraform provisions, Ansible configures, Vector streams, Polars processes. T-Pot exists in the same space but locks you to ELK and 20+ decoys with weak narrative — I wanted the opposite. Skip the T-Pot detail unless Q&A asks.

84 years ago, in 2025

THE PRE-MODERN WAY

Open a chat. Paste context. Argue.

Google Gemini Pro. Every session a blank slate. Context saturation → prompt rage → new thread → paste context again.

December 2025 through March 2026 — Alpha built this way. The design decisions held up under later review by smarter models, so the thinking was fine. The workflow was prehistoric. Set up the contrast for the numbers slide.

The numbers

BEFORE / AFTER CLAUDE

9× more lines written

5× more commits

3.5× more files touched

Same project. Same person. Not full-time. 2.5 months before → 2 months after.

Exact figures if pressed: 46 → 234 commits (215 substantive + 19 dependency bumps), 65 → 226 files, +3,180 → +29,204 lines. Claude onboarded 2026-04-26 (commit f6f69f5). Last commit in window: 2026-06-24. Don’t read the numbers aloud — let the ratios land. Source screenshot: assets/claude-usage-statistics.webp.

The insight

WHERE I ADDED VALUE

me

Is it feasible and easy to create a check like this: if the last N enrichment tries for that source failed due to rate_limit, skip the rest? It’d save much much time.

claude

Feasible and easy — ~20 lines of code in one function, plus tests. Circuit breaker would have bailed in ~70 seconds instead of thrashing rate limits for 30 minutes. Implementing now.

This is the archetype of the collaboration for me. I saw the shape of the problem — a threat detection engineer notices API rate-limit thrash the way a mechanic hears a bad bearing. Claude implemented a per-provider consecutive-failure counter, faster than I could have looked up the tenacity docs. Screenshot source: assets/claude-enrichment-optimization.webp.

6h → 5min

THE RESULT

6h enrichment run

→

5 min after the circuit breaker

One insight from me. Ten minutes of Claude working. ~99% cut.

Alone, neither of us would have done this in an afternoon. Together, ten minutes. That’s the frame — not “AI wrote my code” but “AI let me spend my time where it mattered.” Land this line: together we’re stronger; alone we’re both weaker.

Caveats & takeaways

WHAT I LEARNED

Configure the AI
Keep CLAUDE.md lean. Bloat makes agents worse.

Plan first
Let Claude interview you before it writes code.

Manage context
/clear between tasks. Context rot is real.

Try again in 3 months
Models keep improving faster than intuition.

Configure — ETH Zurich showed LLM-generated context files made agents worse in 5/8 settings (via Chase H on Pragmatic Engineer). Plan first — Cherny’s “make it interview you” pattern; the interview is the plan. Manage context — /clear beats /compact when switching tasks; context rot is a real degradation, not a vibe. Three months — Cherny’s rule; try the same idea after a model bump before deciding it doesn’t work. If a “let it commit” question comes up: yes, let Claude commit; give it clear phasing and message conventions in your user CLAUDE.md — that’s what unlocked the second half of Lantana for me.

Thank you

FIND MORE

• Lantana — github.com/lopes/lantana · Lantana 1 · Lantana 2
• Pragmatic Engineer — Cherny · Building Claude Code · Yegge · From IDEs to AI Agents
• The 6 Levels of Claude Code — Chase-H-AI
• My CLAUDE.md — github.com/lopes/dotfiles/…/CLAUDE.md

lopes.id jlopesjr lopes

Land the frame one more time: neither Claude nor I would have shipped Lantana v1 in two months alone. The AI didn’t replace me — it made me faster where I was slow, and stayed out of the way where I was already fast. Then open the floor.

Q&A

QUESTIONS