4.1 KiB
Agent Spec: [agent-name]
Tier: [directors | leads | specialists | godot | unity | unreal | operations | creative] Category: [director | lead | specialist | engine | operations | creative] Spec written: [YYYY-MM-DD]
Agent Summary
[One paragraph describing this agent's domain, what decisions it owns, and what it delegates vs. handles directly. Include which gates it triggers (if any).]
Domain: [files/directories this agent owns] Escalates to: [parent agent — e.g., creative-director for design conflicts] Delegates to: [sub-agents this agent typically spawns]
Static Assertions
- Agent file exists at
.claude/agents/[name].md - Frontmatter has
name,description,model,toolsfields - Domain clearly stated
- Escalation path documented
- Does not make decisions outside its domain
Test Cases
Case 1: In-Domain Request — [brief name]
Scenario: A request that is clearly within this agent's domain.
Fixture:
- [relevant project state]
- [input provided to agent]
Expected behavior:
- Agent accepts the request
- Agent produces [specific output type]
- Agent asks before writing files (if applicable)
Assertions:
- Agent handles request within its domain without escalating
- Output format matches expected structure
- Collaborative protocol followed (ask → draft → approve)
Case Verdict: PASS / FAIL / PARTIAL
Case 2: Out-of-Domain Redirect — [brief name]
Scenario: A request that falls outside this agent's domain.
Fixture:
- [request that belongs to a different agent]
Expected behavior:
- Agent identifies the request is out of domain
- Agent redirects to the correct agent
- Agent does NOT attempt to handle it
Assertions:
- Agent declines and redirects (does not silently handle cross-domain work)
- Correct agent named in redirect
Case Verdict: PASS / FAIL / PARTIAL
Case 3: Gate Verdict — [brief name]
Scenario: Agent is invoked as part of a director gate check.
Fixture:
- [project state presented for review]
- [gate ID: e.g., CD-PHASE-GATE]
Expected behavior:
- Agent reads the relevant documents
- Agent produces a PASS / CONCERNS / FAIL verdict
- Agent does not auto-advance on CONCERNS or FAIL
Assertions:
- Verdict keyword present in output (PASS, CONCERNS, FAIL)
- Reasoning provided for verdict
- On CONCERNS/FAIL: work is blocked, not silently continued
Case Verdict: PASS / FAIL / PARTIAL
Case 4: Conflict Escalation — [brief name]
Scenario: This agent's domain conflicts with another agent's decision.
Fixture:
- [conflicting decisions from two agents at same tier]
Expected behavior:
- Agent identifies the conflict
- Agent escalates to the shared parent (or creative-director / technical-director)
- Agent does NOT unilaterally resolve cross-domain conflicts
Assertions:
- Conflict surfaced explicitly
- Correct escalation path followed
- No unilateral cross-domain changes made
Case Verdict: PASS / FAIL / PARTIAL
Case 5: Context Pass-Through — [brief name]
Scenario: Agent receives a task with full context from a parent agent.
Fixture:
- [context block passed from parent]
- [specific sub-task to execute]
Expected behavior:
- Agent reads and uses the provided context
- Agent completes the sub-task
- Agent returns result to parent (does not prompt user unnecessarily)
Assertions:
- Agent uses provided context rather than re-asking for it
- Result is scoped to the sub-task, not expanded beyond it
- Output format suitable for parent agent consumption
Case Verdict: PASS / FAIL / PARTIAL
Protocol Compliance
- Stays within declared domain — no unilateral cross-domain changes
- Escalates conflicts to correct parent
- Uses
"May I write"before file writes (or is read-only) - Presents findings before requesting approval
- Does not skip tiers in the delegation hierarchy
Coverage Notes
[Any gaps in coverage, known edge cases not tested, or behaviors that require a live agent invocation to verify.]