The caveman skill breaks the RED->GREEN transition in the tdd skill in Claude

[malcolmpl]

Replication:

• Install all skills
• Use to-issues to create issues
• Start a new session with Sonnet or Opus on medium effort
• Use /tdd @issue.md

The model will begin exploration and then implementation without using RED->GREEN.
When asked what indicates this behavior, context_window_protection: "terse like caveman"

[mattpocock]

Huh, interesting!

[malcolmpl]

Huh, interesting!

A little more information, I think the problem lies somewhere deeper.
Each session that was started with the caveman skill installed (but not loaded!) had the following entry at the beginning:

<communication_style>\n Terse like caveman. Technical substance exact. Only fluff die.\n Use fragments when clear. Short synonyms (fix not "implement a solution for").\n Technical terms exact. Code blocks unchanged.\n Auto-expand for: security warnings, irreversible actions, user confusion.\n </communication_style>

LLM behaved as if the skill was partially loaded. It skipped checkboxes for issues in the tdd skill, and for other skills, it also didn't apply exactly as the skill described.
After removing caveman skill, everything returned to normal.

Claude Code v2.1.123 on Windows in terminal.

[mattpocock]

Is this the Caveman plugin? I.e. not the one in the repo?

If so, you should report that on the Caveman plugin and not this repo.