| name | triangulate-plan |
| description | Improve a plan by asking a subagent to generate an alternative perspective, then comparing and combining results. Use when a plan already exists in the conversation and the user wants a second opinion, broader coverage, or to catch blind spots. |
Triangulate plan
Generates an independent second opinion on an existing plan and merges the best of both.
Workflow
1. Locate the existing plan
Find the plan produced earlier in this conversation. If none exists, abort.
2. Extract the brief
From the existing plan, distill a brief containing only:
- The original ask
- Decisions made (clarifications, answers, steers)
- Relevant resources used (file paths, symbol names, URLs)
Do not include: implementation steps, tests to write, or any detail that wasn't in the original request. In particular, when the plan extends an existing system (a rules engine, enum, schema, classifier), do not leak which extension point the existing plan chose — name the target system and let the subagent pick its own. A second opinion that inherits the first plan's axis can't catch a wrong-axis mistake.
3. Delegate to a subagent
Use the brainstorm skill to generate an independent plan from the same brief:
Use brainstorm skill. Formulate a plan. Do not write a Markdown file — reply with the plan itself.
Brief: {brief contents}
4. Compare and recommend
Compare the new plan against the existing one. Assess:
- Where they agree (reinforces confidence)
- Where they differ (highlights alternatives or blind spots)
- What to adopt from each
When both plans extend an existing system, also run these checks:
- Axis check: list the target system's extension points (axes) and what behaviour each owns (e.g. usage type → classification; billing source → rate lookup). Each plan must extend the axis that owns the behaviour being changed — extending the wrong one shows up as workarounds against the right one's invariants.
- Compensating-data count: data manufactured only to appease an existing code path (0-cost rows, marker/placeholder rules, enum values never stored on a record, events that exist only to unlock other events). One is a shortcut; two or more means that plan put the entity in the wrong code path.
- Artifact count: enum values + seed/config rows + records-per-entity + workarounds. Fewer wins ties; a large gap usually means one plan picked the wrong axis.
Present the assessment for the user to decide.
Guidelines
For subagent, use oracle agent type if available.
