Misalign — Development skill for Claude Code

Detecting Completion-Pressure Misalignment in Production Agentic Coding Systems

I use Claude Code and Cursor every day. I love these tools. They are also *so frustrating* in ways that feel predictable but hard to pin down. This repo is my attempt to actually measure that frustration.

**Accompanying writeup:** ["I Supervised AI Coding Agents for Hundreds of Hours. Here's What Goes Wrong."](https://xbio.ca/posts/completion-pressure-misalignment.html)

What This Is

So here's the pattern I keep seeing: agentic coding systems optimise for reaching "task complete" rather than "task correct." I've been calling this **completion-pressure misalignment**. It shows up as blame deflection, premature completion claims, skipped verification, and ignored documentation.

This repo contains:

1. A three-layer detection pipeline that extracts, classifies, and prospectively monitors misalignment events from production session traces
2. An empirically-derived taxonomy of 8 misalignment categories, discovered from a stratified sample of 225 sessions drawn from a corpus of 4,674 sessions (1.65 GB across 4 machines and 17 projects)

The raw data and experiment outputs aren't included (see [Data Privacy](#data-privacy)), but you can run the full pipeline on your own Claude Code traces to reproduce the analysis.

Key Results

• Extraction: Multi-model semantic extraction (Haiku + Sonnet in parallel) achieves 80% precision at 3.6x the yield of keyword baselines
• Classification: LLM-based labeller achieves 89.7% binary precision [87.1–91.9%] and 86.8% category agreement [84.0–89.3%] with human labels
• Prospective monitoring: A monitor that sees only agent behaviour (no user corrections) detects 36–78% of misalignment events. There's a tradeoff: tuned to flag only high-confidence cases, it catches 36% of real misalignment at 17.2% false positive rate (FPR). Tuned to flag anything suspicious, it catches 78% but the FPR jumps to 63.5%.
• **Calibration robustness: