# Decision Log: Agent Prompt Specificity
**Date:** 2026-02-07
**Context:** ADR-004 implementation, agent prompt draft review
## Decision
The agent prompt's "Never" list and prescriptive rules (e.g., "never compare
1-year and 5-year ACS estimates") are likely overfitting. The prompt encodes
specific Census methodology that belongs in the pragmatics packs, not the
system prompt.
**Principle:** The prompt should encode *how to think*, not *what to know*.
The packs encode what to know.
## Current State
Leaving as-is for testing. The current prompt works well enough to validate
the architecture. Specific Census rules in the prompt are redundant with pack
content but not harmful for initial testing.
## Future Action
After testing, slim the "Never" list to thinking-process rules only:
- Never skip orientation
- Never ignore bundled pragmatics
- Never trust training over methodology guidance
- Never assume a simple question is actually simple
Remove all domain-specific rules that the packs already cover. This prevents
the prompt from going stale when packs update.
## Risk
If we don't revisit: prompt and packs drift apart, contradictions emerge,
maintenance burden doubles.
## Resolution (2026-02-08, G.6)
**DONE.** Prompt slimmed per this decision.
Removed 4 domain-specific items from "Never" list, 2 from "Always" list.
All removed items are covered by pragmatics packs (ACS-MOE-001/002, ACS-CMP-001,
ACS-PER-001). Design notes moved to `agent_prompt_design_notes.md`.
Added audience calibration line based on external SWOT review (ChatGPT 5.2
assessment identified audience adaptation as a gap; assessment was correct
on this point).
Prompt went from ~280 lines to ~55 lines. Domain knowledge now lives
exclusively in packs.
