A plugin for Claude Code agent team workflows — session viewing, analysis, and best practice evaluation

GitHub stars
128

Install

/plugin install agent-teams@ai-launchpad

Run it inside Claude Code after adding the marketplace with /plugin marketplace add kenneth-liao/ai-launchpad-marketplace.

Contents

Discovered from source

Skills 2

Analyze Team Session

Analyze an agent team session export against the official Claude Code agent teams best practices. Produce a structured report with a suitability verdict, scorecard, actionable recommendations, and an improved prompt rewrite.

Core Principle: Every recommendation must cite a specific best practice from the official docs and quote specific evidence from the session. No vague assessments.

When to Use

  • User provides a team session export and wants feedback
  • User asks how to improve their agent team usage
  • User wants to know if a task was a good fit for agent teams
  • After running an agent team, to identify what to do differently next time

Prerequisite: This skill requires a markdown session export. Use the agent-teams:view-team-session skill to generate one from a session ID.

Do NOT use for:

  • Analyzing solo (non-team) Claude Code sessions
  • General Claude Code usage advice unrelated to agent teams

Workflow

Step 1: Read the session export

Read the full markdown session export file. If too large, read in chunks.

Extract these elements:

  • Team name and session ID
  • Duration and agent list
  • Original user prompt — the first user message that kicked off the team
  • Team structure — lead + teammates, their roles, models used
  • Task list — tasks created, assigned, completed, dependencies between tasks (look for TaskCreate/TaskUpdate calls)
  • Communication flow — DMs between teammates (SendMessage calls), broadcasts
  • Tool calls — what each agent did (reads, edits, web fetches, etc.)
  • Quality gates — any plan approval workflows, TaskCompleted hooks, TeammateIdle hooks
  • Final output — what the lead produced as the end result

Step 2: Fetch the official documentation

Search for the latest agent teams best practices. Try these approaches in order:

  1. WebSearch: Search for "Claude Code agent teams best practices site:code.claude.com" to discover the current URL
  2. WebFetch: Fetch the discovered URL, or fall back to https://code.claude.com/docs/en/agent-teams
  3. Costs section: Also fetch https://code.claude.com/docs/en/costs for the "Agent team token costs" and "Manage agent team costs" sections

These are the authoritative source for all rubric evaluations. If both WebSearch and WebFetch fail, inform the user that analysis cannot proceed without the official docs.

Step 3: Evaluate against the rubric

Read the evaluation rubric from references/rubric.md (10 categories). For each category:

  • Rate it: Gap, Partial, or Strong
  • Quote evidence: Cite specific passages from the session export
  • Explain: Reference the specific best practice from the official docs

Step 4: Write the analysis report

Save the report to .claude/output/<team-name>-analysis.md. Actionable content comes first, detailed evidence last.

Use this structure:

# Agent Team Session Analysis: <team-name>

**Session:** <session-id> | **Team:** <team-name>
**Duration:** <duration> | **Agents:** <comma-separated list>
**Analysis date:** <today's date>

---

## Suitability Verdict

[1-2 paragraphs assessing whether this task was a good fit for agent teams.
If not, explain what approach would have been better and why. Be direct.]

**Verdict:** Good fit / Marginal fit / Poor fit

---

## Summary

| Category | Rating |
|----------|--------|
| Suitability | [Strong/Partial/Gap] |
| Context Sharing | [Strong/Partial/Gap] |
| Task Sizing | [Strong/Partial/Gap] |
| Task Dependencies | [Strong/Partial/Gap/N/A] |
| Communication Quality | [Strong/Partial/Gap] |
| File Conflict Avoidance | [Strong/Partial/Gap/N/A] |
| Lead Orchestration | [Strong/Partial/Gap] |
| Model Selection & Cost Efficiency | [Strong/Partial/Gap] |
| Quality Gates | [Strong/Partial/Gap/N/A] |
| Cleanup | [Strong/Partial/Gap] |

---

## Top Recommendations

[3-5 recommendations ranked by impact. Each must be specific, actionable,
and reference what happened in the session.]

1. **[Category]** — [Recommendation with concrete example of what to change]
2. ...

---

## Improved Prompt

If the original prompt could be rewritten following all best practices,
here's what it would look like:

[Rewrite the user's original prompt incorporating all recommendations.
This should be ready to copy-paste for their next run.]

---

## Detailed Rubric Scorecard

### 1. Suitability — [Rating]

**Evidence:**
> [Quoted passages from the session export]

**Assessment:** [Explanation referencing specific best practice from the docs]

[...continue for all 10 categories...]

After writing the report, tell the user where it was saved and give a 2-3 sentence summary of the key findings.

Quality Checklist

  • All 10 rubric categories evaluated with evidence and doc references
  • Suitability verdict is direct and honest (not hedging)
  • Recommendations are specific and actionable (not "improve communication")
  • Improved prompt is a complete, ready-to-use rewrite
  • Every rating cites a specific passage from the session export
  • Every assessment references a specific best practice from the official docs
  • Report saved to .claude/output/<team-name>-analysis.md

Common Pitfalls

  1. Vague ratings: Saying "Communication was Partial" without quoting the actual messages (or lack thereof). Always cite evidence.
  2. Hedging the verdict: If it was a poor fit for agent teams, say so. The point of the analysis is honest feedback, not diplomacy.
  3. Generic recommendations: "Improve task sizing" is useless. "Split the single review task into 3 focused sub-tasks (voice, engagement, accuracy) with 2-3 checkpoints each" is actionable.
  4. Ignoring newer features: If the session didn't use task dependencies, quality gates, or plan approval, note whether these would have helped — they may not have been available when the session ran, but they're available now.

View Team Session

Generate a self-contained HTML viewer from Claude Code session JSONL logs. Works for both solo sessions and agent team sessions, showing the full conversation timeline with filtering, search, and collapsible tool calls.

How It Works

Claude Code stores conversation logs as JSONL files at ~/.claude/projects/<project>/<session-id>.jsonl. When agent teams are used, each agent gets its own JSONL file in the same project directory, linked by a shared teamName field. Team config is stored at ~/.claude/teams/{team-name}/config.json.

Browsers can't read local JSONL files directly, so the skill uses a Python script to extract the data and embed it into a self-contained HTML playground.

Usage

Run the generate script with a session ID using uv run:

uv run ${CLAUDE_SKILL_DIR}/scripts/generate.py SESSION_ID [--output path] [--no-open]
  • SESSION_ID — the UUID from the JSONL filename (e.g., 605eef0e-40be-4e0f-8a34-0a977c60399d)
  • --output path — custom output path (default: .claude/output/<session-id>.html)
  • --no-open — skip auto-opening in browser

The script will:

  1. Scan ~/.claude/projects/*/ to find the session's JSONL file
  2. If it's a team session, discover all teammate JSONL files via the shared teamName field and enrich with team config from ~/.claude/teams/
  3. Parse all files into structured events (speech, DMs, tool calls, task management, thinking, etc.)
  4. Deduplicate inter-agent DMs (each DM exists in both sender and receiver logs — only the sender's copy is kept since it has richer metadata)
  5. Embed the data into the HTML template and write the output file

Finding Session IDs

If the user doesn't provide a session ID, help them find one:

# List recent sessions for the current project
ls -lt ~/.claude/projects/<project>/*.jsonl | head -10

# Search for sessions containing specific content
grep -l "search term" ~/.claude/projects/<project>/*.jsonl

The session ID is the filename without the .jsonl extension.

Output

The generated HTML is a single self-contained file with:

  • Header showing team name, session ID, duration, event count
  • Sidebar with agent toggles, event type filters, and content search
  • Timeline of event cards with agent-colored borders
  • Task management events highlighted (task creation, assignment, completion, dependencies)
  • Default view shows speech and DMs only; tool calls, thinking, and system events are togglable
  • Dark theme (GitHub-dark style)

Output files go to .claude/output/ by default.

Sponsors

Details

Marketplace
kenneth-liao/ai-launchpad-marketplace
Source
agent-teams
Category
Community

More from kenneth-liao/ai-launchpad-marketplace

All 5