escapeboy/ai-prompts/agent-team

agent-team

Spawns a coordinated agent team for tasks that benefit from parallel, independent exploration.

When to Use (and When NOT to)

Only spawn a team when ALL of these are true:

• Work can be split into independent parts (no shared file edits)
• Teammates benefit from communicating with each other (not just reporting back)
• The task is complex enough that 3+ perspectives add real value

Preset modes (pass as argument)

/agent-team pr-review

3 teammates reviewing the current branch changes:

• Security — IDOR, mass-assignment, cross-tenant leaks, rate limiting
• Logic — correctness, edge cases, missing error handling
• Coverage — test coverage for new code, missing assertions

/agent-team debug

3-5 teammates investigating a bug with competing hypotheses — each tries to disprove the others' theories. Pass the bug description as the argument.

/agent-team feature

Teammates each own an independent slice of a larger feature. Pass the feature description — Claude will decompose into non-overlapping file groups.

/agent-team custom

Free-form — describe the team structure you want directly.

Instructions to Claude

When this skill is invoked:

1. Confirm the team structure with the user before spawning (cost warning: each teammate = separate context window)
2. Spawn the appropriate team based on the preset or custom description
3. Brief each teammate with enough context: which files they own, what to focus on, what to avoid
4. Wait for teammates to finish — do not start doing work yourself while they are active
5. Synthesize findings from all teammates into a single structured report
6. Clean up the team when done (Clean up the team)

Cost awareness

Agent teams consume significantly more tokens than a single session. Always mention this before spawning if the user hasn't explicitly acknowledged it.

Rough guidance:

• 3 teammates × medium task ≈ 3-5× normal session cost
• 5 teammates × large task ≈ 8-10× normal session cost

Limitations to communicate

• No session resumption (/resume won't restore teammates)
• Task status can lag — nudge stuck teammates manually
• Each teammate owns different files — two editing the same file = overwrite risk
• One team per session — clean up before starting another

PR review example output

After all reviewers finish, synthesize as:

## Agent Team Review — [branch name]
 
### Security (teammate 1)
[findings]
 
### Logic (teammate 2)
[findings]
 
### Test Coverage (teammate 3)
[findings]
 
### Summary
- Blockers: [list]
- Concerns: [list]
- Approved areas: [list]

Boundaries

Always

• State the team structure and the rough cost multiplier before spawning.
• Give each teammate a disjoint set of files.
• Clean up the team when finished.

Ask first

• Before spawning more than 5 teammates, or a team on a large task (≈8-10× session cost).

Never

• Let two teammates edit the same file.
• Start a second team in the same session before the first is cleaned up.

Related skills

• Guide: `06-advanced-patterns/agent-teams-guide.md` — how Agent Teams work and when they pay off.
• `company` — uses this skill (depended-on-by) for the parts of a larger project where teammates must talk (debug, cross-review); reach for /company when the whole task needs clarifying, splitting and staffing.
• `code-research` — read-only parallel research/audit of a codebase or git repo (agents fan out over non-overlapping slices, write durable notes, and do NOT talk to each other). Use it instead of a team when the goal is to map / audit / understand a codebase rather than have teammates coordinate mid-task.

Task: $ARGUMENTS