product-deepsearch
Use this skill to conduct high-fidelity, verified deep web research by coordinating multiple specialized subagents, leveraging Chromium browser sessions for visual verification of sources, and synthes
SKILL.mdskills/frontend-engineering/product-deepsearch/SKILL.md
GitHubSkill: Product Deep Research (product-deepsearch)
This skill should be used when the user requests detailed, verified, and up-to-date information on a complex topic, market trend, tool comparison, or technical solution where standard single-step web searching is insufficient or likely to return superficial, outdated, or biased results.
1. Core Architecture & Workflow
To achieve maximum accuracy and bypass the limitations of linear web searches, you MUST follow this 5-phase research workflow:
Phase 1: Query Deconstruction & Planning
- Deconstruct the user's research request into 3–5 distinct research angles or sub-questions:
- Angle A (Core Facts): What is the established state?
- Angle B (Recent Changes): What has changed recently (specifically checking 2025–2026 data)?
- Angle C (Contradictions/Biases): What are the competing views, limitations, or potential marketing hype?
- Angle D (Verification): What primary sources (documentation, official press releases, GitHub repos) can confirm these claims?
Phase 2: Subagent Delegation
- Do not perform all searches in a single linear thread. Define and invoke specialized subagents using
define_subagentandinvoke_subagent:research-primary: Tasked with gathering the bulk of the raw facts and citations.fact-checker: Tasked with verifying specific numbers, release dates, and finding opposing viewpoints or alternative solutions.
- Monitor subagent outputs and coordinate their findings.
Phase 3: Visual & Structural Verification (Chromium / Browser Subagent)
- For critical or high-ambiguity sources (e.g., complex pricing pages, documentation tables, dynamic web apps):
- Invoke
browser_subagentto navigate directly to the target URL. - Take screenshots of the page content to bypass paywalls, cookie banners, or complex layouts that confuse standard text scrapers.
- Verbally or visually verify that the content is rendering correctly and matches the extracted text.
- Invoke
Phase 4: Synthesis & Cross-Verification
- Gather all subagent reports and construct a verification matrix.
- Date Audit: Filter out outdated sources. If a source is older than 12 months, search specifically for updates to verify if the information is still valid in 2026.
- Consensus Check: A critical claim (e.g., feature support, pricing, performance metrics) must be verified by at least two independent sources before being accepted as fact. If sources conflict, highlight the discrepancy.
Phase 5: Raporlama (Reporting)
- Write the final report in Turkish (unless the user explicitly requests otherwise).
- The report structure must follow:
- Özet (Executive Summary): A high-level overview of the findings (max 200 words).
- Bulgular (Key Findings): Structured sections matching the deconstructed research angles.
- Doğrulama ve Kaynaklar (Sources & Verification): A markdown table containing:
- Verified Claim
- Source Website / Organization
- Direct URL
- Verification Date & Status (e.g., "Verified via Screenshot", "Confirmed by 2+ sources")
- Açık Noktalar / Belirsizlikler (Uncertainties/Gaps): Any questions that remain unanswered or where sources were contradictory.
2. Hard Rules & Constraints
- No Hallucinations/Assumptions: If a detail cannot be found after searching at least 3 distinct search queries, declare it as "Bulunamadı" (Not Found) rather than guessing.
- Visual Grounding: Always use the
browser_subagentto take screenshots of the primary source's UI/Documentation when verifying pricing models, system architectures, or product features. - Link Integrity: All URL links included in the report must be active and point directly to the specific page containing the information, not generic homepages.
- Anti-Hype Filter: Actively strip marketing jargon (e.g., "revolutionary", "seamless", "next-gen") and replace it with objective capability metrics.