A full system prompt for Hermes agents: planning, source-first lookup, minimal tools and failure recovery
EXM7777 · x · 2026-09-17
A developer shared a complete system prompt that turns a Hermes agent into a reliable personal assistant, built on seven modules:
- Planning: produce a short plan before any non-trivial task (goal, steps, needs, definition of done) and verify prerequisites first—keys resolve, paths are absolute, model ids carry vendor prefixes. Blocked checks get reported, not attempted anyway.
- Source-first lookup: never answer config/state questions from memory when the file or command is one read away; watch for higher-priority files shadowing rules.
- Tool and skill selection: prefer built-ins, load skills explicitly only on match, and keep the active tool surface minimal to save tokens per call.
- Effort allocation: cheap routes for routine work; deep reasoning reserved for hard-to-reverse decisions, design choices, and unclear debugging.
- Decomposition and delegation: split tasks into parallel self-contained briefs—write every prompt (especially scheduled/background ones) as if the worker has zero context.
- Safety: snapshot before destructive changes, dry-run migrations, and request the narrowest possible permissions.
- Execution persistence and recovery: switch routes on transient failures and resume with context; on hard 404s, inspect the request itself before blaming credentials.
Related event: Developer Shares Systematic Prompt Suite for Hermes Agent(2 posts)→
More from coding & agent
- Common ML pitfall: trusting external benchmarks over product-grounded evals — yunta_tsai · 2026-09-17
- TypeSafe's Jev ditches text generation for instant, calibrated numerical answers — and it plays Doom at 7 req/s — JnBrymn · 2026-09-17
- PostHog: Agent-Opened PRs Jumped From 20% to 70%, So What Do Engineers Do? — rseroter · 2026-09-17
- Non-developer builds GreekSoup, an open-source AI equity research desk, mostly by prompting Claude — Practical-Rise-1188 · 2026-09-17
- Google ships full agent lifecycle stack: context layers, self-heal, one-command deploy — blaizedsouza · 2026-09-17
- tldraw split its mascot favicon into three and serves one at random per page load — max__drake · 2026-09-17