Agent-maintained "living document" skills (theorist/THEORY.md) are anti-patterns in branching git workflows

The theorist skill (blader community skill, v1.3.0) maintains a single THEORY.md at repo root — a "living essay" rewritten holistically every session to capture the operating theory behind the current work. In practice this causes three problems in any repo with branches:

1. Merge conflicts on every branch. THEORY.md is rewritten end-to-end per session. Two branches touching different subsystems both rewrite it → guaranteed content conflict that's impossible to auto-resolve (every line changed on both sides).

2. Stale on branch switch. The theory narrative is about the current work, but git checkout other-branch carries THEORY.md's content from the previous branch's last session. The agent reads it at session start, orients to the wrong context, and burns tokens updating it before real work begins.

3. Token cost with no durable value. The skill mandates rewrites every ~10 minutes of active work and after every investigate/implement/verify loop. These rewrites consume significant context window and output tokens for a document that (a) is overwritten next session, (b) conflicts across branches, and (c) duplicates information better captured in commit messages and PR descriptions.

The fix: Delete the skill and file. Commit messages (subject: what; body: why, alternatives considered, what broke the old theory) are the natural, branch-friendly vehicle. They're immutable, travel with the branch, and git log --first-parent reconstructs the theory narrative without maintenance cost. PR descriptions aggregate the branch narrative into one place GitHub preserves after merge.

When the skill might work: Solo developer, single long-lived branch, multi-day investigation where you want the next session's agent to cold-start with full context. Even then, a pinned issue or local:// file is more robust than a tracked file at repo root.