AgentSkillsCN

researching

当需求模糊不清、存在多种技术方案,或变更可能影响架构、API、数据或安全性时使用。

SKILL.md
--- frontmatter
name: researching
description:
  Use when requirements are fuzzy, multiple technical approaches exist, or
  change affects architecture, API, data, or security

Researching Skill

Use when the how is unclear; output is a chosen approach with evidence.

Prerequisites: If what/why is unclear, use discovering skill first.

Quick Reference

TrackWhenOutput
Fast (default)Standard decisionsBrief: problem, decision, risks
FullAPI/security/multi-subsystemResearch brief + optional design doc

When to Use

  • Requirements are fuzzy or incomplete
  • Multiple plausible approaches exist (spike, proof of concept needed)
  • Change affects architecture, API, data, or security

Skip if: obvious bug, trivial change, pattern already exists.

Fast vs Full Track

code
Is it high-risk?
├── No → Fast Track (default)
└── Yes → Full Track
    ├── Changes public API
    ├── Security/data implications
    ├── Team disagreement
    └── Touches multiple subsystems

Core Rule

Evidence before opinions. Never recommend anything until facts are gathered: what already exists in code, which constraints are real, what decisions were made before. If evidence is missing — label it as assumption and lower confidence.

Quick Rules

  • One question at a time while clarifying requirements
  • If multiple interpretations exist, clarify before researching
  • State assumptions explicitly with confidence (H/M/L)

When to Ask vs Act

Ask (one question at a time; prefer multiple choice) if:

  • Multiple interpretations
  • Critical context missing
  • Answer would change direction

Act (state assumptions + confidence) if:

  • Default interpretation is clear
  • Request is specific
  • Assumptions are easy to verify

Workflow

Fast Track (default)

  1. Clarify & Frame — purpose, constraints, success criteria
  2. Collect Signals — code, docs, history
  3. Options + Decision — 2-3 options, pick one
  4. Brief — Problem (1-2 sentences), Decision, Key Risks (1-3 bullets)

Full Track

PhaseFocusOutput
0. FrameProblem, non-goals, success criteriaConstraints
1. SignalsCode, docs, historyCurrent state + assumptions
2. Options2-4 approaches with trade-offsComparison
3. EvaluateDecide using Phase 0 criteriaRecommendation + fallback plan
4. ArtifactsResearch brief (always)Optional: design doc

Research Brief template: Problem & Context → Constraints → Current State → Options Compared → Recommendation → Risks

Stop Conditions

Stop research when:

  • Decision criteria are satisfied
  • Remaining unknowns won't change the decision
  • Next step is prototype/measurement

Best Practices

  • Assumptions + Confidence — label assumptions (H/M/L), propose quick validation
  • Spike as Research — 30-90 min spike with clear question and success metric
  • Checkpoints (Full Track) — after each phase: "Constraints complete? Which risks matter?"

Common Mistakes

MistakeFix
Recommend without evidenceGather facts first, then opinions
Single option, no alternativesAlways 2-3 options with trade-offs
Drift into tangential topicsStay connected to problem statement
Hide uncertainty behind confidenceState unknowns explicitly
Finalize with disputed framingResolve framing first