Vigil Crest: A Custom Hermes Agent for Hackathon Triage

Vigil Crest is the filter I wanted

L Cordero developed Vigil Crest, a challenge triage agent integrated with Telegram. The system evaluates active hackathons based on stack fit, learning potential, and timing to prevent wasted development days.

Why This Matters

Generic AI prompts often rely on simple scoring formulas (e.g., ranking by prize money), which fail to account for the genuine cost of time and specific technical prerequisites. Vigil Crest addresses this by using a persona-driven reasoning approach via SOUL.md, prioritizing nuanced judgment over binary formulas to ensure the user only enters challenges where they are competitive.

Key Insights

• Persona-based reasoning: The agent uses SOUL.md (2025/2026) to apply qualitative criteria like ‘excitement as a prerequisite’ rather than quantitative scoring.
• Dynamic stack verification: A refresh-stack skill automatically updates the agent’s known languages by reading public GitHub repositories.
• Model selection for availability: Claude Sonnet 4.6 via AWS Bedrock was chosen over Opus to balance reasoning capabilities with operational affordability for frequent runs.
• Environment sensitivity in automation: Headless browsers used in scheduled background processes can fail where manual shell executions succeed, necessitating API-based pre-checks.

Practical Applications

• : Use case: Hermes Agent + Telegram for personalized decision support based on curated persona files (SOUL.md). Pitfall: Relying on search engine caches for live feeds, leading to stale data and missed opportunities.
• : Use case: Playwright for scraping rendered page content when dedicated APIs are unavailable. Pitfall: Using headless browsers in background schedulers without environment parity, resulting in invisible execution failures.

References:

• https://dev.to/earlgreyhot1701d/i-built-a-hermes-agent-to-tell-me-which-hackathons-to-enter-it-told-me-to-enter-this-one-jh2
• github.com/earlgreyhot1701D/vigil-crest