[MODEL] Keyword choice when using [Web Search] tool restricts discovery
Preflight Checklist
- [x] I have searched existing issues for similar behavior reports
- [x] This report does NOT contain sensitive information (API keys, passwords, etc.)
Type of Behavior Issue
Other unexpected behavior
What You Asked Claude to Do
The model, when asked to do tasks that involve websearch, always uses keywords that will almost never yield the best results for that task, such as always using 2024 at the end of its queries (Even for Opus 4.5), or suggesting keywords for the search that are unrelated or will clutter the results, also on the results page for each, if its mentioned that the library is deprecated or have a new version, it will never follow through to read the subsequent page, resulting in outdated/subpar results.
Example when searching for a markdown library:
This is gold. The key insight: "connect, don't collect" and "read now or never".
Let me research local LLMs that excel at structured extraction for this pipeline.
● Web Search("best local LLM structured JSON extraction 2025 Ollama llama qwen mistral comparison")
⎿ Did 1 search in 27s
● Web Search("Qwen 2.5 vs Llama 3.2 vs Mistral local structured output markdown tagging 2025")
⎿ Did 1 search in 23s
● Web Search("best local embedding model 2025 bge nomic sentence transformers Ollama comparison")
⎿ Did 1 search in 10s
❯ don't hardcode model names into your search,
● Web Search("best local embedding models for semantic search 2025 benchmark comparison")
⎿ Did 0 searches in 5s
● Web Search("MTEB leaderboard 2025 open source embedding models ranking")
⎿ Did 1 search in 5s
❯ why are you searching for things that are not related to doc -> markdown?
● You're right, I'm overcomplicating this. The core question is:
URL/doc → structured markdown with frontmatter + tags
● Web Search("convert web article to markdown frontmatter YAML automated pipeline CLI 2025")
⎿ Did 1 search in 19s
● Web Search("LLM extract content generate markdown obsidian frontmatter local 2025")
⎿ Did 1 search in 20s
● Web Search("Jina reader markdownify readability web scraper to markdown 2025")
⎿ Did 1 search in 18s
● Now I have the full picture. Here's the pipeline:
[Redacted]
❯ literally on their page https://huggingface.co/jinaai/ReaderLM-v2, anyways, we have git to protect us,
so feel free to delete code, i like to use bun, with a modern tui library, if need any scripting
● Web Search("bun modern TUI CLI library 2025 ink clack terminal UI")
⎿ Did 1 search in 7s
❯ STOP HARDCODING LIBRARY NAMES INTO YOUR SEARCH WHEN YOU'RE "SEARCHING"
● Web Search("best terminal UI library TypeScript 2025")
⎿ Found 10 results for "best terminal UI library TypeScript 2025"
Thank you
What Claude Actually Did
Read log above, it always searches for unrelated keywords, constricting the search results.
Expected Behavior
It should look from the perspective of a user who would be looking for that topic and suggest keywords that don't narrow the search unless it's a followup query
Files Affected
Permission Mode
Accept Edits was ON (auto-accepting changes)
Can You Reproduce This?
Yes, every time with the same prompt
Steps to Reproduce
_No response_
Claude Model
Sonnet
Relevant Conversation
Impact
Critical - Data loss or corrupted project
Claude Code Version
2.1.19
Platform
Other
Additional Context
_No response_
This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗