openclaw - 💡(How to fix) Fix 2026.5.18: shortened tool descriptions degrade agent decision quality despite same model + skills

Summary

The 2026.5.18 update shortened built-in tool descriptions and tightened skill prompts across all tools and skills. While this likely aimed to reduce token usage, it has a measurable negative impact on agent decision quality — the same model + same skills now produce less precise results because the agent receives less context about when and how to use each tool correctly.

Why this matters

The CHANGELOG entry says:

"Agents/tools: shorten built-in tool descriptions and schema hints across media, messaging, sessions, cron, Gateway, web, image/PDF, TTS, nodes, and plan tools while preserving routing guardrails."

The intent — preserving routing guardrails — is good, but the execution misses the point. Tool descriptions serve two purposes:

1. Routing — which tool matches the user's request (the guardrail layer)
2. Quality — how to invoke the tool correctly with the right parameters, in the right order, with the right context

Shortening descriptions may preserve #1 (routing), but it degrades #2 (correct usage). An agent that knows which tool to use but not how to use it precisely will:

• Pass wrong parameters
• Miss edge cases captured in the original description
• Need more turns to correct mistakes
• Produce lower-quality results despite having the same model and skills

In a work scenario

Token savings on tool descriptions are negligible compared to the cost of wrong tool usage. A single extra correction turn burns far more tokens than an expanded tool description would over hundreds of calls. Clarity and accuracy should not be sacrificed for marginal token savings on agent-facing prompts.

Suggestion

If token budget is a concern, consider:

• Keeping full descriptions for tools that require nuanced parameter handling
• Shortening only trivial tools where the name alone is self-explanatory
• Using per-tool verbosity levels rather than a blanket trim
• Measuring the real token impact before and after: is the saving worth the quality regression?

A blanket shorten-all approach treats the symptom (context window pressure) without considering the root cause (which parts of the context are worth compressing vs. preserving).

Environment

• OpenClaw version: 2026.5.18
• Affects all agents/models/skills uniformly since the change is in the tool definition layer
• No community issue or feature request drove this change — it was an internal engineering decision without user-facing discussion