Token optimization: how we cut system prompt costs 73%

[thegdsks]

We were burning 2,401 input tokens per message on system prompts. Cut it to 650. Here is what changed.

What was eating tokens

Component	Tokens	Needed?
Skills list (55 skills)	~800	No, not for regular chat
Model aliases table	~150	No, not for local models
AGENTS.md context	~400	No, dev docs not chat context
Recent activity stats	~50	Only when non-zero
Preset prompt + grounding	~350	Yes

What we did

1. Removed skills list from non-tool chat mode. Only inject top 15 when tools are active.
2. Skip model aliases for Ollama models. They don't need to know about cloud model names.
3. Stopped injecting AGENTS.md into every conversation. It's dev workflow docs, not useful for chat.
4. Condensed activity context to one line when data exists.
5. Added "do NOT repeat this" instruction so small models use context as background knowledge instead of dumping it back.

Results

Metric	Before	After
Input tokens per message	2,401	650
Reduction		73%
Monthly savings (100 msg/day, Sonnet)		~$5
Monthly savings (100 msg/day, GPT-4o)		~$8

Small models like Gemma 4 E4B benefit the most. Less context means faster responses and fewer hallucinations.

The changes are in src/chat/system-prompts.ts if you want to see the diff.