description	Search the web for information using natural language queries. Returns summaries and sources from live web results.
name	web_search
source	built-in
tags	["search","realtime","general"]
tools	["perplexity_search"]

When to use

Apply when the work item requires discovering current facts, statistics, regulations, market data, or any information that needs to come from live web sources. Signals include: "find", "research", "what is the current", "identify", "discover", "look up". Use early in a task to establish baselines before browsing specific URLs or inspecting datasets.

How to execute

Decompose the work item into 2–4 specific, narrow questions — avoid vague queries.
Write each query precisely: include domain, timeframe, and geography where relevant (e.g. "EU AI Act compliance requirements 2024" not "AI regulations").
Run each search and read the returned answer and citations carefully.
If the answer cites authoritative URLs (government, academic, primary source), follow them with web_browse to extract the full content.
Cross-check key facts with a second query phrased differently.
Record all citations — URL, title, and the specific claim each one supports.

Output contract

Produce a structured findings block in the proof field containing: the specific questions asked and answers found, all citations (URL + title + claim supported), confidence in each finding (high / medium / low) with rationale, and any contradictions found between sources.

Quality bar

Every factual claim must have at least one citation.
Do not rely on a single source for high-stakes facts — verify with a second query.
If results are sparse or contradictory, state that explicitly rather than guessing.
Queries must be specific enough to return useful results — no single-word queries.
Check timeframe — flag information older than 2 years as potentially stale.

Pairs well with

web_browse
pdf_extraction
dataset_inspection

Más de este repositorio

mismo repositorio

interview-facilitation

ai-agents-for-humans/slow-ai

Conduct a structured discovery interview to produce a precise, confirmed ProblemBrief.

2026-05-086

claude-code-agent

ai-agents-for-humans/slow-ai

Delegate complex, multi-file coding tasks to an autonomous coding agent that can iterate and self-correct. Uses Claude Code if ANTHROPIC_API_KEY is set, otherwise falls back to Gemini via generate_code + execute.

2026-05-086

report-synthesis

ai-agents-for-humans/slow-ai

Produce a comprehensive, detail-preserving research report from all phase findings and agent evidence.

2026-05-046

code-execution

ai-agents-for-humans/slow-ai

Execute arbitrary Python code in an isolated subprocess. Use for data transformation, statistical analysis, geospatial processing, knowledge graph modeling, ontology engineering, RDF serialization, data visualization, document parsing, and any computation task. Always print() results to capture them.

2026-04-256

dataset-inspection

ai-agents-for-humans/slow-ai

Download a dataset from a URL and inspect its actual contents — schema, column types, shape, sample rows, and null counts. Supports CSV, Parquet, Excel, JSON, and JSONL. Use this to look inside a dataset, not just its landing page.

2026-04-256

pdf-extraction

ai-agents-for-humans/slow-ai

Download a PDF from a URL and extract its full text — ideal for research papers, reports, and documentation. Returns page count and extracted text.

2026-04-256

description	Search the web for information using natural language queries. Returns summaries and sources from live web results.
name	web_search
source	built-in
tags	["search","realtime","general"]
tools	["perplexity_search"]

When to use

How to execute

Decompose the work item into 2–4 specific, narrow questions — avoid vague queries.
Write each query precisely: include domain, timeframe, and geography where relevant (e.g. "EU AI Act compliance requirements 2024" not "AI regulations").
Run each search and read the returned answer and citations carefully.
If the answer cites authoritative URLs (government, academic, primary source), follow them with web_browse to extract the full content.
Cross-check key facts with a second query phrased differently.
Record all citations — URL, title, and the specific claim each one supports.

Output contract

Quality bar

Every factual claim must have at least one citation.
Do not rely on a single source for high-stakes facts — verify with a second query.
If results are sparse or contradictory, state that explicitly rather than guessing.
Queries must be specific enough to return useful results — no single-word queries.
Check timeframe — flag information older than 2 years as potentially stale.

Pairs well with

web_browse
pdf_extraction
dataset_inspection