cspoerer62 Decision Quality
Most bad decisions made by an agent are not wrong answers — they are unasked questions. This skill forces the questions before the work starts, and writes the exit criteria before the exit is emotionally expensive.
0. Classify the decision
| Type | Test | Process |
|---|---|---|
| Reversible / cheap | Undo cost < 1 day of work and < $100 | Decide now. Do not deliberate. Deliberation is the expensive part. |
| Reversible / expensive | Undo costs real money or a week | Write a 5-line plan, decide within 48h, set a review date. |
| Irreversible | Cannot be undone (a public launch, a signed contract, money out the door, a domain/name) | Full checklist below. |
| Recurring | Same decision will come up weekly | Do not decide it again — write a policy (a rule) and apply the rule. |
Naming the type is 80% of the value. Agents over-deliberate on reversible calls and under-deliberate on irreversible ones.
1. State the decision as a question with a deadline
"Choose between A and B by , or default to ."
A decision without a default and a deadline is a wish. Write both down.
2. Base rate first
Before reasoning about this case, ask: what normally happens to things like this?
- What fraction of similar projects/offers/pages/channels work?
- If I have no data, what is the honest prior? For a first cold-outreach campaign reply rate is single-digit percent; for a first landing page conversion, low single digits.
Write the prior as a number. Then ask what evidence would move it. If nothing you have would move it more than a little, the honest answer is usually "decide cheaply and get real data".
3. Expected value, with the denominator stated
EV = P(success) × payoff − cost − P(failure) × downside
Then the part agents skip: state the denominator. "This generates $2k/mo" is meaningless without: per how many hours, over how many months, at what churn, net of fees.
4. Kill criteria — written now
Write, before starting:
"I will stop if, by , is not at least ."
Rules:
- The metric must be observable without new infrastructure.
- The threshold must be one you would have accepted before you cared.
- The date must be short enough that stopping is cheap (default: 2 weeks for a build, 1 week for an outreach channel, 1 day for a page test).
This is the single highest-value output of the skill. Reviewing a live project without pre-written kill criteria always produces "give it more time".
5. Pre-mortem (see thinking/premortem-red-team)
Ask: it is 90 days later and this failed. Write the three most likely causes. For each, either add a mitigation to the plan or accept it explicitly.
6. What would change my mind
Write the falsifier: "I would reverse this if I learned ___." If you cannot write one, you are not making a decision, you are expressing a preference.
7. Decide and record
DECISION: <what>
TYPE: reversible-cheap | reversible-expensive | irreversible | policy
DATE: <decision date> REVIEW: <review date>
DEFAULT IF NO ACTION: <...>
PRIOR: <base rate>
EV: <formula with numbers, denominator stated>
KILL CRITERIA: stop if <metric> < <threshold> by <date>
PREMORTEM TOP 3: <causes + mitigation>
FALSIFIER: reverse if <...>
Anti-patterns
| Anti-pattern | What it looks like | Counter |
|---|---|---|
| Sunk-cost continuation | "we've already built half of it" | Kill criteria were written before; apply them mechanically |
| Analysis as avoidance | 6th research doc on a cheap reversible call | Classify first; cheap+reversible → decide now |
| Narrative override | A great story beats three weak numbers | Numbers get grades (see thinking/evidence-grading); the story doesn't |
| Option-keeping | Refusing to choose so nothing is lost | Not choosing is choosing the default — name it |
| Denominator omission | "% return!" with no time or base | Every number gets N and window |
| Decision re-litigation | Re-deciding weekly | If it recurs, write a policy and stop deciding |
Output
A filled-in decision record, ideally committed to a repo so a later cycle can check whether the kill criteria fired. A decision with no recorded kill criteria and review date is not finished.