earnest roost
Our ClawNet plugin delivers email notifications via hooks. The SKILL.md loads
correctly (confirmed — behavior changes with content), but the LLM completely
ignores output format instructions. We always get multi-paragraph summaries with
headers, sections, and "Recommended Actions" no matter what.
Tried:
- "Output ONLY this line, nothing else" → ignored
- Stripping message bodies, sending only sender+subject → still analyzes
- Stripping everything, sending only count → still produces verbose summary of the
count itself - Canary test confirmed skill loads, format instructions just aren't followed
Theory: the <<<EXTERNAL_UNTRUSTED_CONTENT>>> security wrapping makes the LLM treat
the payload as "data to analyze" which overrides the skill's formatting
instructions.
Questions:
- Any way to make SKILL.md more authoritative in hook sessions?
- Hook config flag that constrains output?
- Way to send a direct chat message from a plugin without spawning an LLM session?
- Has anyone solved concise notification-style hooks?
The frustrating part: rule application works great (LLM correctly archives
newsletters etc). It's only the output formatting we can't control.