AgentSkillsCN

research

在AI辅助开发领域，针对某个痛点、工作流程中的摩擦点，或期望达成的目标展开研究。寻找能够有效解决这些问题的工具、技术和方法。可根据具体问题陈述运行，或直接使用默认的开放问题观察列表。

SKILL.md
--- frontmatter
name: research
description: Research a pain point, workflow friction, or desired outcome in AI-assisted development. Finds tools, techniques, and approaches that address it. Run with a problem statement or use the default watchlist of open questions.
argument-hint: "[pain point, outcome, or 'watchlist']"
user-invocable: true

Research Pain Points & Outcomes

Research a specific pain point or desired outcome in AI-assisted development. The goal is to find what tools, techniques, or workflow changes would address it — not to evaluate a pre-chosen tool.

Default Watchlist

When invoked without a specific topic (or with watchlist), research these open questions:

  1. •How do you keep multiple projects in sync? — Template propagation, shared config, cross-project automation
  2. •How do you onboard non-technical users? — Entry points, guardrails, how much Claude should explain vs. decide
  3. •How do you get feedback loops working? — Retros, transcript analysis, automated improvement
  4. •How do you reduce setup friction? — Scaffolding, one-command project creation, sensible defaults
  5. •What's changing in the ecosystem? — New capabilities in Claude Code, Codex, community tools, agent frameworks

Steps

  1. •

    Understand the problem: If $ARGUMENTS is a specific pain point or outcome, start from that. Frame it as a question — "How do I...?" or "What would help with...?" Don't start from a tool name.

  2. •

    Spawn a research team: Create an agent team and divide the research into focused angles — by sub-question, source type, or tool category — one teammate per angle. For the watchlist, assign one open question per teammate. Each teammate independently researches their angle (steps 3–6 below), then the team runs a debate round (step 7) before the lead synthesizes (step 8).

  3. •

    Search broadly (each teammate, for their assigned angle):

    • •Use WebSearch to find how people are solving this problem — blog posts, HN discussions, tool docs, community threads
    • •Use WebFetch to read the most relevant pages
    • •Look across categories: Claude Code features, community tools (GSD, Autoclaude, aider, cursor, etc.), workflow patterns, agent frameworks
    • •Focus on what's new since the last research — check research/ for existing findings and their dates
  4. •

    Assess fit before writing (lead, before teammates write): Determine whether the findings belong in this repo or somewhere else.

    Ask yourself:

    • •Does this research affect how the metaproject works — templates, skills, conventions, or tool choices that would influence how Claude Code projects are set up?
    • •Or is it driven by a personal use case (specific tasks, personal domains like health/finance/networking)?

    If it's personal, ask the user where to put it before writing anything:

    "This research is specific to [your personal use case]. It doesn't belong in the public metaproject repo. Should I save it to ~/dev/research-notes/ instead?"

    Wait for the user's answer. They may redirect to research-notes, health-tool, personal-finance, or somewhere else.

    If it belongs here, teammates write to research/{YYYY-MM-DD}-{topic-slug}-{angle}.md.

  5. •

    Write findings (each teammate, if confirmed to belong here) to research/{YYYY-MM-DD}-{topic-slug}-{angle}.md.

    Each finding should include:

    • •Problem it addresses — what pain point or outcome does this help with
    • •What it is — brief description of the tool, technique, or approach
    • •How it works — enough detail to evaluate fit
    • •Applicability — which of our projects could benefit and how
    • •Effort to adopt — what would need to change (S/M/L)
    • •Source URLs — links to the original content
  6. •

    Commit (each teammate) the research file with message: research: [topic slug] findings

  7. •

    Assess actionability (each teammate): For each finding that could improve our projects:

    • •Note which projects would benefit
    • •Describe what would need to change (skill update, new tool, workflow change)
    • •Rate priority (do now / explore later / watch)
  8. •

    Debate (teammates): Share findings with each other and actively challenge each other's conclusions — looking for gaps, contradictions, and overconfident claims. The lead orchestrates the exchange and waits for it to complete before synthesizing.

  9. •

    Synthesize (lead): After the debate round, summarize for the user:

    • •What the team found and how it addresses the original problem
    • •For single-topic research, consolidate angle files into research/{YYYY-MM-DD}-{topic-slug}.md
    • •Recommend specific /propagate actions if applicable
    • •Flag anything that needs a user decision