Skip to main content

​
Phantom Concept Elimination in Agent System Prompts

​
The Problem

Grace’s predecessor agent files (Zoid template) referenced an 8-agent council (Anvil, Bard, Vault, Cortex, Tempo, Compass, Sentinel, Atlas) that didn’t exist anywhere in the STRATT codebase. STRATT has 3 real councils with 16 agents using Martian designations. The template also used trigger words (FORGE, TOMMY, POSTIE, STORM) that didn’t map to any real CLI commands or workflows.

​
The Insight

Phantom concepts in agent system prompts cause compounding errors. The agent routes work to non-existent councils, references non-existent workflows, and generates responses grounded in fiction rather than reality. This is worse than having no agent context at all — it’s confident misdirection.

​
The Pattern

Every reference in an agent system prompt must pass a verification gate:
  1. Councils — verified against councils/{name}/council.yaml files
  2. Agent designations — 16 designations confirmed 1:1 against council configs
  3. CLI commands — 21 commands confirmed by listing packages/cli/src/commands/
  4. Failure modes — FM-01-08 confirmed against packages/graph/src/ci.ts
  5. Cross-repo paths — test -d <path> for every referenced directory

​
Business Impact

Grounded agent context means every routing decision, council consultation, and tool invocation maps to a real capability. This is the foundation for reliable autonomous operations across the STRATT + Choco HQ ecosystem.