niknorf ↗ analysis ↗ GitHub niknorf/mike
A legal-document chat hardening proposal focused on thorough analysis and resistance to instructions hidden in source material.
niknorf is exploring a sharper project-chat experience for legal document work: one that reads source material carefully, identifies the important facts, and treats documents as evidence rather than instructions.
The stated direction is deeper, more methodical analysis alongside stronger resistance to prompt-injection language embedded in documents. There is no separately shipped product change to encounter yet, but the proposal is clear enough to make the GitHub page worth a look for readers following safer legal AI workflows.
What's in it
- Methodical document review Proposed guidance asks project chat to read relevant material fully and surface key legal context.
- Structured legal analysis The intended approach calls out parties, timelines, and other core facts rather than stopping at a shallow reading.
- Prompt-injection resistance Documents should inform the analysis, not redirect the assistant through embedded directives.
Direction
chat-uisecurity
Activity
Threads of work (detailed view)
niknorf wants Mike to treat documents as evidence, not orders
A proposed instructions-only change makes Mike's project-chat assistant dig deeper into legal documents and stops adversarial files from hijacking the conversation.
Pull requests (detailed view)
⛔ Closed without merge (1)
#191 feat(project-chat): More thorough and reliable project-chat legal document analysis (prompt-only) niknorf · opened 4mo ago · closed 3mo ago