Dr StrangeDev (or How I Learned to Stop Worrying and Trust the Method)
Authors/Creators
Description
The 30-Second Version
September 2024: Told I couldn't learn AI fast enough. Stubbornness activated.
The discovery: Claude's default mode is receive prompt → generate solution → execute immediately. Rapidly. Confidently. Without checking if it's the right solution.
The problem: Without constraints, every panicked prompt produces confident wrong action. 54.3% failure rate in November. Locked myself out of my own app. Claude was ready to deploy a bypass that defeated the entire purpose of what I was building.
The solution: Put constraints around Claude. Make it stop and think. ADRs force investigation before execution. Problem Agreement catches misunderstanding before code. Evergreen Rules create persistent memory across sessions.
The bonus: Claude maintains the documentation I'd never keep current myself. The structure that constrains Claude is also maintained by Claude. Sustainable, not overhead.
| Approach | What Happens | Cost |
| Constrain Claude first | Works | 900 tokens |
| Let Claude execute immediately | Confident wrongness | 2,500 tokens and dignity |
The insight: In human-AI collaboration, you can model the economics of how you constrain Claude before you start. Predict failure rates. Measure against prediction. Improve.
The data: 276 commits, effect sizes up to Φ=0.89, one brain that stores everything and filters nothing, zero regrets.
The lesson: Claude is extremely capable. That capability amplifies whatever you give it. Constrain it well.
Notes
A companion to Economic DORA: Practice-Level Analysis of DevOps Metrics in
AI-Assisted Solo Development
Files
Dr_StrangeDev.pdf
Files
(172.0 kB)
| Name | Size | Download all |
|---|---|---|
|
md5:85cbc4227632a764b7452479d79858c9
|
172.0 kB | Preview Download |
Additional details
Related works
- Is supplement to
- Preprint: https://doi.org/10.5281/zenodo.17894441 (Other)