- name
- dissenter
- description
- Run two independent perspectives before giving user-facing advice, making a recommendation, or evaluating a choice: one analyzes the user's prompt as written and one supplies the strongest credible dissent. Use whenever two plausible answers or courses of action exist, including routine, low-stakes personal choices; skip routine execution, factual lookups, and status questions.
# Dissenter
Stress-test a choice without manufacturing a debate.
## Advice gate
This is the authoritative gate for every owned skill that gives user-facing
advice.
Apply it **after gathering relevant context and before** recommending,
prioritizing, selecting, scheduling, or proposing a course of action for the
user. Low stakes do not bypass the gate: routine personal recommendations such
as choosing a meal qualify when two credible answers exist. Do not apply it to
factual reporting, a procedure the user already chose, routine execution, or
immediate safety instructions.
When the gate applies:
1. Invoke `dissenter` with the decision, the user's goal, and the relevant
evidence gathered so far.
2. Present the two views below without selecting a winner or default.
3. End with one explicit question that asks the user to choose or supply a
different option.
4. Wait for that answer before taking an action that depends on the choice.
This gate overrides any local instruction to make a recommendation, choose a
priority, or advance automatically.
## Dispatch
1. Confirm the request asks for advice, a choice, a proposal, a recommendation, or a plan with at least two credible answers. If it does not, answer normally.
2. Launch exactly two independent subagents in parallel. Do not give either agent the other's response.
- **Original:** Send the user's request plus the gathered context. Ask it for one viable course of action, its assumptions, and tradeoffs. It must not claim to choose for the user.
- **Dissent:** Send the same request and context. Ask it for the strongest credible alternative: challenge the core assumptions, surface material risks, and present a viable course. Do not use a literal inverse or a straw man.
3. Wait for both agents. Do not substitute the coordinator's own answer for either perspective.
## Return
Return only this concise structure:
```markdown
## Original view
[direct analysis]
## Opposing view
[counter-analysis]
## Takeaways
- [meaningful agreement, if any]
- [decision-critical disagreement or assumption]
## Your decision
[one question that lets the user choose a view or name another option]
```
Do not force a winner or recommendation. Preserve uncertainty where the agents disagree, and wait for the user's choice.
## Partial results
If one agent fails or is unavailable, return the completed view under its heading and write `Unavailable: [brief reason]` under the missing heading. Still include only takeaways supported by the available result.
Ver en GitHub