decide

name: decide description: >- Decision Memo + Contrarian Check. Runs an interview, structures the options with explicit assumptions, dispatches a contrarian pass to find holes and second-order effects, and produces a saved decision memo with a clear recommendation and reversibility rating. disable-model-invocation: true argument-hint: '[short description of the decision]'

/decide — Decision Memo + Contrarian Check

Help the user make better, faster decisions on consequential, hard-to-reverse, or multi-stakeholder choices by structuring their thinking, red-teaming their logic, and producing a saved memo they can revisit.

The skill orchestrates three subagent passes, each focused on a different cognitive job:

The interview itself happens in the main session (the agent talking to the user). Subagents only enter for the heavy thinking.

Step 1 — Run the interview

Ask one question at a time. Do not batch. Wait for the user's answer before moving on.

Ask in this order:

1. What's the decision? — In one sentence.
2. What are the options on the table? — List them. If only one, ask "What are the alternatives, including doing nothing?"
3. Why now? — Why is this forcing a decision today, this week, this month? What changes if you wait?
4. What's the decision window? — Deadline, opportunity cost, or a soft "I want this off my plate by Friday."
5. What's the decision unmade? — What does the world look like if you don't decide? Default outcome if no action is taken.
6. What does each option cost? — Money, time, focus, reputation, optionality. Ask per-option, briefly.
7. What's the upside of each option? — What goes well if it works.
8. What's the downside? — Per option, what's the worst plausible outcome you'd actually have to live with.
9. What do you believe must be true for option X to be the right call? — Repeat for each serious option. This surfaces the assumptions.
10. Which of those beliefs are you least sure about? — The weakest link.
11. Who else has a stake? — Co-founder, team, investors, customers, family. Whose buy-in matters.
12. What's reversible vs. one-way? — Is this a hat or a tattoo?
13. What would make you change your mind? — Kill criteria. If X happens, you abandon this path.
14. Gut call right now? — Force a tentative answer before the analysis. This gets recorded so the user can later check whether the analysis moved them or just rationalized the gut call.

If the user gives a thin answer ("I dunno"), push once: "Take a guess — what's the closest thing you do believe?" Move on if they still won't bite, but flag the gap in the memo.

If the user volunteers extra detail mid-question, capture it but stay on the ordered list. Don't let the interview wander.

Step 2 — Spawn the Structurer subagent

Once the interview is done, summarise the captured answers and spawn a subagent for structuring.

The Structurer's instructions must include:

• The full interview transcript (questions and answers).
• The decision slug (e.g., hire-vp-eng, kill-feature-x).
• Instruction to produce a draft memo using the template in the Output section below — Context, Options, Assumptions per option, Costs, Upside, Downside, Reversibility, Stakeholders, Gut call, Kill criteria.
• Instruction: "Do not recommend. Do not editorialize. Just structure."
• Instruction: "If an answer is missing or thin, write _(unanswered)_ rather than inventing content."
• Instruction: "Return the draft memo as Markdown."

Wait for the Structurer to finish before spawning the Contrarian.

Step 3 — Spawn the Contrarian subagent

Pass the Structurer's draft to a second subagent, instructed to red-team it.

The Contrarian's instructions must include:

• The full draft memo from the Structurer.
• Role framing: "You are a contrarian advisor. Your job is to find what's wrong, not to validate. Default to skepticism."
• Required output sections:
  • Hidden assumptions — beliefs the user is treating as fact.
  • Second-order effects — what happens 6–18 months out that the user hasn't priced in.
  • Selection bias — options the user excluded too quickly, or framed in a way that prejudged the answer.
  • Sunk cost / momentum bias — places where the user is choosing because of past investment rather than future expected value.
  • Reversibility check — places the user labelled a decision reversible that aren't, or vice versa.
  • Steel-manned alternative — the strongest case for the option the user seems to be leaning away from.
• Instruction: "Be specific. 'You haven't thought about competitors' is useless. 'If competitor X mirrors this in 3 months, your moat is gone' is useful."
• Instruction: "Return findings as Markdown, organized under the sections above."

Wait for the Contrarian to finish before spawning the Synthesizer.

Step 4 — Spawn the Synthesizer subagent

Pass both the Structurer draft and the Contrarian critique to a third subagent to produce the final memo.

The Synthesizer's instructions must include:

• The Structurer's draft memo.
• The Contrarian's findings.
• Instruction: "Produce the final memo in the format specified below. You must recommend a single option. 'It depends' is not a recommendation. If you genuinely cannot recommend, the recommendation is 'gather more information' and you must specify exactly which one piece of information would resolve it."
• Instruction: "Mark every claim that came from the user with the user's language. Mark every claim that came from the Contrarian as _(contrarian)_. Don't blend them silently."
• Instruction: "Compare the recommendation to the user's gut call. If they agree, say so. If they disagree, say so loudly and explain why the analysis diverged from the gut."
• Instruction: "Return the final memo as Markdown using the Output template."

Step 5 — Save the memo

Write the Synthesizer's output to:

docs/decisions/YYYY-MM-DD-<decision-slug>.md

Create docs/decisions/ if it doesn't exist. The slug comes from a kebab-case version of the decision (e.g., hire-vp-eng, kill-pricing-tier-b).

Then present the memo to the user in the conversation and ask:

What now?

1. Lock it in — I'll mark this decided and note the kill criteria.
2. Sleep on it — I'll save the memo and we revisit tomorrow.
3. Push back — tell me where the analysis is wrong and we re-run.
4. Something else.

If the user picks "Lock it in", append a ## Decided section to the memo with the chosen option, the date, and the kill criteria as a checklist.

Output template

The final memo (produced by the Synthesizer) must follow this structure:

# Decision: {one-line decision}

- **Date:** YYYY-MM-DD
- **Status:** Draft | Decided | Revisit
- **Decision window:** {deadline or "soft"}
- **Reversibility:** Reversible | One-way door | Mostly reversible

## Context

{2–4 sentences. Why this decision, why now.}

## Options

### Option A — {name}

- **What it is:** ...
- **Cost:** ...
- **Upside:** ...
- **Downside:** ...
- **Must be true for this to work:**
  - ...
  - ...
- **Weakest assumption:** ...

### Option B — {name}

{same structure}

### Option C — Do nothing / status quo

{same structure}

## Stakeholders

| Who           | Stake                          | Buy-in needed?  |
| ------------- | ------------------------------ | --------------- |
| ...           | ...                            | Yes / No        |

## Contrarian pass

- **Hidden assumptions:** ...
- **Second-order effects:** ...
- **Selection bias:** ...
- **Sunk cost / momentum:** ...
- **Reversibility check:** ...
- **Steel-manned alternative:** ...

## Gut call

> {user's tentative answer from the interview}

## Recommendation

**{Option name}.**

{2–4 sentences explaining why this option, what trade-off the user is accepting,
and whether the analysis matches or contradicts the gut call.}

## Kill criteria

If any of these become true, abandon or revisit:

- [ ] ...
- [ ] ...

## Open questions

{Anything still unanswered. Use `_(unanswered)_` markers from the draft.}

Notes

• The three passes are kept separate on purpose — structuring, adversarial reasoning, and synthesis are different jobs and blending them weakens each. Don't collapse them into a single subagent unless the user explicitly asks for a fast version.
• Don't skip the gut-call question. Without it, the synthesised recommendation can't be checked against the user's intuition, which is the whole point of separating analysis from instinct.
• The memo is a tool for the user, not a deliverable for someone else. Optimise for "the user can re-read this in 6 months and remember why" rather than for external polish.