Renfei Design Design Jarvis Design Jarvis

Online Research — Speed-Optimized

Ground content in current facts with minimum fetch calls. The golden rule: draft from training knowledge, fetch only to verify and update.

When to Use

• Content involves pricing, specs, rankings, dates, or statistics that change over time
• User explicitly asks to "research," "look up," "compare," or "find the latest"
• Creating a doc or brief where outdated info would hurt credibility

Skip fetching when: the topic is conceptual, historical, or opinion-based — training knowledge is sufficient.

Speed Rules

1. Max 2 fetch_webpage calls per research task — hard cap
2. Batch all URLs into one call — urls accepts an array (3-6 URLs per call)
3. Draft first, fetch second — write full content from training knowledge, mark gaps, then fetch only for gaps
4. Use sharp query strings — focused query extracts faster, less noise
5. Official sources only — skip aggregators, comparison sites, and forums on first pass

Workflow

Step 1 — Triage (no fetching)

Split the question into two buckets:

If everything is stable → skip to output, no fetches needed.

Step 2 — Draft from training knowledge

Write the full content structure before fetching:

• Headings, sections, table layout
• Fill in all stable facts
• Mark volatile facts with [VERIFY] placeholders

Step 3 — Fetch (one call, multiple URLs)

fetch_webpage(
  urls: ["https://product-a.com/pricing", "https://product-b.com/pricing"],
  query: "pricing tiers, per-seat cost, free tier limits"
)

URL selection:

Step 4 — Patch the draft

Replace [VERIFY] placeholders with fetched data:

1. Update numbers, dates, facts
2. Add source URLs as inline citations
3. If source contradicts training knowledge, use fetched data (newer)
4. If fetch returned nothing useful — use training knowledge with caveat

Step 5 — Second fetch (only if critical gaps remain)

Only make a second call if:

• A primary URL was gated or empty
• Two sources conflict and a third can resolve it
• A high-stakes data point couldn't be verified

Research Patterns

Price Comparison

1 fetch call. Batch all vendor pricing URLs. Draft table from training knowledge. Patch prices.

Feature Comparison

1 fetch call. Batch official feature pages. Use training knowledge for framework. Fetch for specifics.

Market Analysis

1 fetch call. Training knowledge for narrative. Fetch one high-quality article for current numbers.

Technical Deep-Dive

Usually 0 fetches. Training knowledge covers architecture well. Fetch only for current API details.

Quality Checklist

• Volatile data verified from fetched sources
• Source URLs cited inline
• Used ≤2 fetch_webpage calls
• Pricing/dates include "as of" timestamp
• Gaps disclosed where applicable

Anti-Patterns

• ✗ Fetching 4+ pages one at a time — batch them
• ✗ Fetching to confirm what training knowledge already covers
• ✗ Fetching aggregator sites when official pages are available
• ✗ Reading entire pages for one data point — use sharp query
• ✗ Exploratory fetches before knowing what you need — draft first

Source: renfei-design/design-jarvis — distributed by TomeVault.