Precise Research

Counting and enumeration

Questions that ask "how many" require finding a complete list and counting.

1. web_search — find the authoritative source
2. web_fetch — fetch it with focus on the specific items to count
3. bash — if needed, use python3 -c "..." to count or compute from extracted data

Date lookups

Questions asking "when was" need exact dates in the format the source provides.

1. web_search — find the source
2. web_fetch — extract the exact date string
3. Do NOT reformat dates — return them exactly as the source states

Calculations

Questions requiring math (distance, time, percentage, volume).

1. web_search — find the raw numbers
2. bash — python3 -c "print(result)" to compute the exact answer
3. Plan the search and computation as dependent steps

Cross-referencing

Questions that chain facts: "the person who did X also did Y, what is Z?"

1. web_search — find X (step 0)
2. web_fetch — extract the person/item from X (depends on step 0)
3. web_search — search for Y using the extracted value (depends on step 1, use param_refs)
4. web_fetch — extract Z (depends on step 2)

ArXiv and academic papers

1. web_fetch — fetch https://arxiv.org/abs/<id> directly with format=summary
2. For specific data within papers: web_fetch with a targeted focus parameter

Video content questions

Questions about YouTube videos — timestamps, counts, dialogue.

1. web_fetch — fetch the YouTube page to get the description and metadata
2. bash — use yt-dlp --write-auto-sub --skip-download to get transcripts if needed
3. For specific timestamps: fetch transcript and search with bash grep

Museum and collection lookups

1. web_fetch — fetch the collection page directly using the accession/museum number in the URL
2. Focus on the specific data point asked about

What NOT to do

• Don't return a range when the question asks for a single number
• Don't guess when the evidence is insufficient — search more specifically
• Don't skip the computation step when math is required — use bash with python
• Don't fetch a general overview page when a specific data page exists
• Don't plan one broad search when the question needs a specific lookup