AGENT SKILL

research

Investigate properly: frame questions, build source ladders, triangulate, and report findings. Use when investigating complex questions with primary sources.

License 许可
MIT
最近更新
2026-10-02
Stars
★ 3
Forks
0

01 前言 Front Matter

name
research
group
Research
description
Investigate properly: frame questions, build source ladders, triangulate, and report findings. Use when investigating complex questions with primary sources.
last_reviewed
1788652800000

02 资源文件 2

03 SKILL.md 正文

research

Core Philosophy

Research is not browsing search engines, skimming blog posts, and summarizing the first three links. Anyone can query an LLM or Google and produce a surface-level synthesis of conventional wisdom. Rigorous, high-trust research is an investigative discipline: it frames sharp, refutable hypotheses, climbs the "Source Ladder" to primary provenance, triangulates claims across independent conflicting vectors, audits funding and bias, and delivers structured intelligence with auditable citations.


4-Step Investigative Research Framework

Step 1: Question Framing & Scope Bounding

  1. Refining the Research Perimeter:
    • Broad questions produce generic fluff. Narrow inquiries produce high-signal intelligence:
      • Fluffy: "How does WebAssembly perform?"
      • Incisive: "What is the memory and cold-start latency overhead of Wasmtime versus Node.js v20 when executing identical JSON-parsing microservices under 1,000 QPS on AWS Lambda?"
  2. Defining the Falsification Criteria:
    • State what empirical evidence would refute your initial operating hypothesis.

Step 2: The Source Ladder Architecture

  1. Climbing the Provenance Ladder:
    • Tier 1 (Primary Provenance - Highest Trust): Source code, benchmark raw telemetry, official RFC specifications, SEC 10-K filings, court transcripts, peer-reviewed clinical trial datasets.
    • Tier 2 (Authoritative Secondary): Official technical documentation from system authors, seminal textbooks, peer-reviewed review articles.
    • Tier 3 (Practitioner Analysis): Production post-mortems, engineering blogs from scaled tech companies (e.g. Netflix, Cloudflare).
    • Tier 4 (Unverified / Opinion - Zero Trust without Verification): Marketing whitepapers, Medium posts, corporate PR releases, Reddit comments.
  2. The Golden Rule: Never cite a Tier 4 claim without tracing it back to its Tier 1 origin.

Step 3: Triangulation & Cross-Vector Verification

  1. The 3-Vector Rule:
    • A non-obvious empirical claim cannot be accepted as fact until verified across three independent, non-affiliated sources:
      • Vector 1: Empirical benchmark or source code audit.
      • Vector 2: Third-party independent academic replication or audit.
      • Vector 3: Production post-mortem documenting real-world failure/success.
  2. Reconciling Conflicting Findings:
    • When Source A says X and Source B says Not-X, identify the hidden variable: Did they test different hardware? Different compiler flags? Different network conditions?

Step 4: Structuring the Research Dossier

  1. Standardized Dossier Structure:
    • Executive Summary (BLUF: Bottom Line Up Front).
    • Core Hypotheses & Findings Matrix.
    • Deep Evidence & Methodology Dissection.
    • Contradictions & Caveats.
    • Primary Source Provenance References.

Deliverable Format: High-Trust Research Dossier

# Research Dossier: Edge Compute Cold Starts — Wasmtime vs V8 Isolates

### 1. Executive Summary (BLUF)
Wasmtime delivers a 12x faster cold-start initialization compared to V8 isolates (0.45ms vs 5.6ms), but suffers a 1.8x throughput penalty on heavy numerical compute tasks due to the lack of dynamic JIT profiling and tier-2 optimization.

### 2. Empirical Triangulation Matrix

| Vector | Source / Provenance | Methodology / Scope | Verified Finding |
|---|---|---|---|
| **Vector 1: Primary Telemetry** | Bytecode Alliance Wasmtime Benchmarks (v18.0) | Memory footprint per instance | Baseline idle footprint is 180KB RAM vs 18MB for Node.js V8. |
| **Vector 2: Academic Study** | ACM SIGPLAN 2024 (Müller et al.) | 100k synthetic microservice requests | Cold start latency: 0.42ms (p50), 0.98ms (p99) on Linux x86_64. |
| **Vector 3: Production Audit** | Fastly Lucet/Wasm Architecture Report | Global edge POP telemetry | 99.8% of tenant requests execute in existing pre-warmed memory instances. |

### 3. Critical Contradictions & Edge Cases
- *Discrepancy*: Several commercial serverless benchmarks claimed Wasm cold starts were 50ms.
- *Root Cause Analysis*: Tracing the setup revealed those benchmarks included container cold start times (Docker daemon initialization), not raw WebAssembly runtime compilation.

### 4. Primary Source Citations
1. Bytecode Alliance. (2024). *Wasmtime Performance Suite*. GitHub: https://github.com/bytecodealliance/wasmtime
2. Müller, K., et al. (2024). "Sandboxing at the Edge." *ACM SIGPLAN*, 42(3), 112–126.

Worked Example: Investigating an Enterprise Vendor's "Zero Overhead" Database Claim

  • Context: A vendor claimed their distributed database layer introduced "zero performance overhead" over raw NVMe disk operations.
  • Investigation:
    1. Cloned the vendor's open-source benchmark repository and reviewed the configuration.
    2. Discovered the benchmark had disabled write-ahead logging (fsync=off) and kept the entire test database within the 64GB OS page cache.
    3. Reran the benchmark with persistent disk flushing (O_DIRECT / fsync=on) with working set sizes exceeding RAM.
  • Outcome: Uncovered a 42% write latency regression under durable transaction conditions, saving the enterprise from a flawed multi-million-dollar infrastructure migration.

Verification Checklist

  • [ ] Specific, falsifiable research question formulated before gathering data.
  • [ ] Source Ladder climbed to primary source code, raw data, or official specs.
  • [ ] Key claims triangulated across at least 3 independent, non-affiliated vectors.
  • [ ] Benchmark conditions audited for hidden configuration biases (e.g. caching, compute parity).
  • [ ] Conflicting data and outlier results acknowledged and reconciled.

Anti-Patterns

  • Single-Source Paraphrasing: Relying entirely on a single corporate whitepaper or sponsored blog post to make architectural decisions.
  • Citing Secondary Aggregators: Citing a news article that cites a blog post that misquotes a scientific paper. Always read the original paper.
  • Confirmation Bias Hunting: Searching only for queries that confirm your favorite technical bias (e.g. searching only "Why Rust is faster than Go").

04 同仓库更多 Skills