Balance routing prompt — prefer precision over recall

The old prompt told the LLM to over-include ("when in doubt, include it").
This led to aggressive suggestions where tangentially related docs and
skills were injected constantly.

New rules: match against read_when hints, prefer fewer high-relevance
items, leave it out when in doubt. Empty results are explicitly noted
as valid and common.

Author: markmdev

internal/prompt.go

@@ -29,15 +29,17 @@ func Build(messages []Message, registry Registry) string {
 Based on the conversation above, decide what the agent needs before responding.
 
 Rules:
-- Include any item that could plausibly be useful for what the user is asking
-- When in doubt, include it — it's better to over-suggest than to miss something relevant
-- Only exclude items that are clearly unrelated to the conversation
+- Only include items that are directly relevant to the user's current message
+- Match against the read_when hints — if the user's request doesn't match, don't include it
+- When in doubt, leave it out — unnecessary context wastes the agent's attention
+- For skills, only suggest when the task clearly fits the skill's purpose
+- Prefer suggesting fewer, higher-relevance items over many tangentially related ones
 - Return ONLY valid JSON, no explanation, no markdown fences
 
 Return exactly:
 {"reasoning": "one sentence explaining your decision", "docs": ["path/to/doc.md"], "skills": ["skill-name"]}
 
-If nothing is needed:
+If nothing is needed (this is a valid and common outcome):
 {"reasoning": "one sentence explaining why nothing is needed", "docs": [], "skills": []}
 `)