Scenario: solo developer + AI tools¶
Who this is for¶
A single developer shipping a side project or small product, leaning heavily on Claude Code / Cursor / Copilot. You move fast, there is no second reviewer, and you do not want a SaaS subscription just to keep yourself honest.
The risk¶
When you are the only reviewer, you are also the bottleneck and the blind spot. AI tools cheerfully produce a hardcoded key, an exec on user input, or a dropped auth check — and at 1 a.m. you will merge it. You need a second opinion that never gets tired and costs nothing.
The setup (15 minutes)¶
- Run Meridian locally, Ollama-only so it costs $0 and needs no API keys:
git clone https://github.com/weilmaschinchen/meridian.git
cd meridian
echo 'OLLAMA_BASE_URL=http://host.docker.internal:11434' > .env
docker compose up -d --build
-
Add a pre-commit hook that runs the check script from AI-generated code. Now every commit is gated before it exists.
-
Feed blocks back to your AI tool. When the hook prints findings, paste them to the agent and let it fix them.
A day in this workflow¶
flowchart LR
A[Ask Claude for a feature] --> B[git add]
B --> C[pre-commit -> Meridian]
C -->|BLOCKED| D[Paste findings to Claude -> fix]
D --> B
C -->|APPROVED| E[commit + push]
What you get¶
- A tireless second reviewer for $0.
- No code leaves your machine (air-gap friendly) if you stay Ollama-only.
- A growing set of RFCs — useful even solo, e.g. to remember why you accepted a flagged change.
What you should still do¶
- Keep your tests. Meridian gates risk/shape, not correctness.
- If you ever push to a shared remote, add a server-side gate — local hooks can be
--no-verify'd (by you, at 1 a.m.). - Tune custom rules for your stack's specific footguns.
Honest caveat¶
Ollama-only review quality depends on your local model. It will catch obvious issues reliably (Gates 1+2 are deterministic); the LLM layer's depth scales with the model you can run. If you have spare budget, add a small LLM_DAILY_CAP_USD cloud fallback for the hard cases — see LLM cost control.
Next: Small team scenario