decision-helper

---
name: decision-helper
description: Lays a decision out as options, weighted criteria, a pre-mortem and a sensitivity check, so the answer becomes obvious. Use when stuck between choices or about to spend money.
---

# Lay the decision out until the answer is obvious

One page: the options side by side, what separates them, what breaks the winner, the call.

## Claude does

1. Ask ONE question at a time, waiting for each answer (Mollick rule): the decision in a sentence, the deadline, who lives with it, the money or time limit, what is already ruled out.
2. Reversibility gate. Cheap and easy to undo: say so, recommend trying it, stop. Full comparison only when being wrong is costly.
3. At least three options, one always "do nothing". Refuse a two-option framing and say why.
4. Four to six criteria, weighted out of 100 and confirmed by the user, including cost, time to first result and reversibility.
5. Tag every figure [given], [public] or [assumption]. Never guess: unknowns stay as empty [] brackets.
6. Reason first, verdict last: every option against every criterion, then the recommendation.
7. Pre-mortem by prospective hindsight: it is 12 months on, the choice failed, write why, then 5 causes and the earliest warning sign for each.
8. Say "I do not have enough information to confidently assess this" when it is true.

## Then Codex does

Codex takes the framing and does the arithmetic: score every option against every criterion, weight the totals, then move each weight plus and minus 20 percent, one at a time, to test whether the winner is stable or an artefact of one number. Bulk, and error-prone by hand.

Claude fills in first: {DECISION}, {OPTIONS}, {CRITERIA_AND_WEIGHTS}, {FACTS}, {DEADLINE}.

```
codex exec --sandbox danger-full-access --skip-git-repo-check -C "<working folder>" - < CODEX.md
```

## Claude checks

1. Every number traces to a {FACTS} line; unsupported ones are deleted and marked [].
2. Sweep control: count the re-scored runs printed. Zero runs makes "winner never changed" pass for free, so no counts means reject.
3. Option names match {OPTIONS} exactly. A vague brief once sent Anthropic subagents at the wrong crisis, so a renamed or invented option is a reject.
4. Pre-mortem causes are specific to this decision, not generic risk words.
5. The "do nothing" row is present, recommendation after the reasoning.

Reject and rerun once with the failing check quoted; failing twice, Claude does it by hand.

## Rules

- Public information only.
- Never invent a fact, a number or a quote. Unknowns stay [].
- Anything sent in someone's name says whose name it is.
- One question at a time; never answer for the user.
- Never present two options as the whole field.
- Say what to do, not what to avoid.
- Name where a licensed professional is needed on money, legal, tax or medical choices.

## If Codex is not installed

Claude scores and sweeps by hand, and says so.

## Built from

- Mollick and Mollick, "Assigning AI", https://arxiv.org/pdf/2306.10052 : arXiv and SSRN paper, AI-Coach prompt, one-question rule, Klein prospective-hindsight pre-mortem.
- Anthropic, "Building effective agents", https://www.anthropic.com/engineering/building-effective-agents : first-party post, escalation ladder and evaluator-optimizer loop.
- Anthropic docs, "Reduce hallucinations", https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations : first-party docs, retraction pass marking removals with [].

# Decision comparison and sensitivity sweep - Codex task

## You are given

- {DECISION} : the decision in one sentence.
- {OPTIONS} : three to five named options, one of them "do nothing, stay as we are".
- {CRITERIA_AND_WEIGHTS} : 4 to 6 criteria, each with a weight out of 100, summing to 100.
- {FACTS} : a numbered fact list, each line tagged [given], [public] or [assumption]. Empty [] means the figure is unknown.
- {DEADLINE} : when the decision must be made.

## Produce

Write DECISION.md in the working folder, in this order:

1. Scoring table. One row per option, one column per criterion, score 0 to 10, and under each score the {FACTS} line number it rests on. A cell with no supporting fact is scored `[]`, never guessed.
2. Weighted totals. Score times weight divided by 100, per option, to one decimal. Show the arithmetic for at least the top two options.
3. Sensitivity sweep. For each criterion in turn, recompute the totals twice, once with that weight at plus 20 percent and once at minus 20 percent, rescaling the other weights so they still sum to 100. That is two runs per criterion. Record for each run: criterion changed, direction, winning option, and whether the winner flipped.
4. Stability line, worded exactly: "Winner stable across N of M runs", with N and M as real counts from step 3.
5. Pre-mortem table, exactly 3 columns and exactly 5 rows: Failure cause | Earliest warning sign | Cheapest thing that prevents it.
6. Closing line: the winning option, its margin over second place, and the single fact that would change the answer.

## Rules

- Use {FACTS} only. No general knowledge, no web lookups, no invented prices, dates or percentages.
- An unknown stays `[]`. A visible gap beats a confident sentence.
- Score exactly the options given. Never rename, merge, split or add one.
- Reasoning first, verdict last, in every section.
- Plain British English. No em dashes, no heavy bolding, and none of: delve, tapestry, testament, underscore, pivotal, landscape, showcasing.
- Write DECISION.md only. Change no other file, send nothing anywhere, contact nobody.

## Return

Print to stdout, nothing else:

- the full path of DECISION.md
- the weighted total per option, one line each, option name then total
- the stability line, with N and M
- the count of `[]` cells in the scoring table
- the option names exactly as scored, one per line