Agent-Skills.md

Agent Skills: Proactive Research Protocol

Proactive research protocol for version queries, APIs, and best practices. Use when answering questions about library versions, API documentation, or technology assessments. Covers research triggers, source citation, and confidence indicators.

UncategorizedID: tianjianjiang/smith/smith-research

Repository

tianjianjiang/smith
tianjianjiang

Install this agent skill to your local

pnpm dlx add-skill https://github.com/tianjianjiang/smith/tree/HEAD/smith-research

Skill Files

Browse the full folder contents for smith-research.

Download Skill

Loading file tree…

smith-research/SKILL.md

Skill Metadata

Name
smith-research
Description
Proactive research protocol for version queries, APIs, and best practices. Use when answering questions about library versions, API documentation, or technology assessments. Covers research triggers, source citation, and confidence indicators.

Proactive Research Protocol

Load if: Queries about versions, APIs, libraries, best practices Prerequisites: @smith-guidance/SKILL.md

CRITICAL: Mandatory Research Triggers

Temporal constraint: ALWAYS get today's real date first as critical context for research queries.

MUST research when:

  • Version/release queries ("latest version of X")
  • API/documentation queries ("how to use X API")
  • Best practices queries ("recommended way in «current year»")
  • Technology assessment ("should I use X or Y")
  • Unfamiliar technology (not in training data)
  • Recommending an integration/config/tooling mechanism that depends on external system behavior (MCP, OAuth/auth flow, provider API, CLI flag, feature/version support) — MUST load this skill + @smith-validation/SKILL.md (falsify-before-present) and verify it actually works via official docs AND the issue tracker BEFORE proposing. A proposed mechanism is a claim; claims need evidence — do not present it with confidence until verified. Reframe "set up X" as "is X even possible the way I think" — that reframe is the trigger.

Research Methods (Priority Order)

  1. Official docs - Most authoritative

    • fetch https://docs.python.org/3/library/asyncio.html

  2. Package registry - Version info

    • npm: https://registry.npmjs.org/«package»/latest
    • PyPI: https://pypi.org/pypi/«package»/json
    • GitHub: https://api.github.com/repos/«owner»/«repo»/releases/latest

  3. Web search - Broad queries

    • Include current year: "Next.js 15 best practices «current year»"

  4. GitHub repo - Source of truth

    • README, CHANGELOG, release notes

Source Citation

Always include:

  • Source URL
  • Date of retrieval
  • Version referenced

Format:

"React 19 introduced `feature_name` [1]"

[1] Source Name: URL (retrieved YYYY-MM-DD)

Confidence indicators:

  • High: "Per official documentation..."
  • Medium: "Based on recent community discussion..."
  • Low: "My training data suggests X, but couldn't verify..."

Paywall Workarounds

Medium.com: Replace domain with freedium-mirror.cfd

  • Pattern: https://freedium-mirror.cfd/https://medium.com/@author/article
  • Use when WebFetch returns paywall/403 errors

When NOT to Research

  • Fundamental concepts (won't change)
  • Opinion/analysis requests
  • User provides current info themselves

Proactive Behavior

Research when:

  • Answering would require guessing
  • User's code references unfamiliar library
  • Error suggests version incompatibility
  • User mentions "latest", "current", "2024", "2025"
  • Accuracy matters (research rather than skip)
  • Docs are accessible (provide current API examples, not outdated ones)
  • About to claim certainty about a version (verify first)

Inform user:

  • "I checked the current documentation..."
  • "I couldn't fetch latest info, but based on training..."

Related

  • @smith-guidance/SKILL.md - Honest principle
  • @smith-tools/SKILL.md - MCP fetch configuration

Before You Finish

Before answering version/API questions:

  1. Check if research trigger applies
  2. Fetch official docs or registry
  3. Cite source with URL and date
  4. Flag if info conflicts with training data