Anti-Rationalization Rules Agents are excellent at inventing reasons to skip the hard parts. This file lists common rationalizations and required rebuttals.
Rationalization
Required rebuttal
"The user asked me to build, so I should build."
The user asked for an outcome. First check problem, evidence, risk, and simpler alternatives.
"This is obvious."
If it is obvious, write the assumption and the evidence. If you cannot, it is not obvious.
"We can research later."
Later research is how wrong products get polished. Do the minimum research now.
"This needs an agent."
Prove that a checklist, script, form, dashboard, cron job, or runbook is insufficient.
"The plan is enough."
A plan without verification and kill criteria is a story.
"Tests pass, so it works."
Tests prove selected behavior, not product fit, safety, or user value.
"The source is popular."
Popularity is a lead, not proof. Find source-backed claims.
"This skill is mostly guidance."
A skill must be executable: trigger, steps, output, verification.
"The user wants speed."
Speed means reducing scope, not skipping thinking.
"We can remember everything."
Memory pollution makes agents worse. Store only durable facts and reusable procedures.
"The risk is low."
Name the failure mode and blast radius. If you cannot, risk is unknown.
"Review will slow us down."
Review prevents expensive rework. Make the artifact smaller if speed matters.
Rationalization
Required rebuttal
"Everyone needs this."
Name the first user segment and their urgent pain.
"This is a platform."
Start with the narrowest wedge that proves value.
"The feature is cool."
Cool is not value. Name the user job and success metric.
"Existing tools are bad."
Identify exactly where existing tools fail.
"People are talking about it."
Public discussion is a signal. It is not validation.
Engineering rationalizations
Rationalization
Required rebuttal
"We can refactor later."
Later rarely comes. Keep the first slice small enough to design well.
"This edge case is unlikely."
Record it. Decide whether to handle, defer, or block.
"Manual testing is enough."
Manual evidence is acceptable only if captured and repeatable enough for the risk.
"The architecture is flexible."
Flexibility without a known change case is complexity.
"We need a framework."
First define the state, artifacts, and interfaces. Then choose runtime support.
Research rationalizations
Rationalization
Required rebuttal
"I found a few links."
Links are not synthesis. Produce claims, confidence, contradictions, and open questions.
"The sources agree."
Check whether they are independent or repeating the same claim.
"No one complains about this."
Absence of complaints is not evidence. Look for failure modes and maintenance cost.
"This repo has many stars."
Stars indicate attention, not fitness for this use case.
Learning rationalizations
Rationalization
Required rebuttal
"This was a one-off."
If the failure or success is likely to recur, capture it.
"The skill is good enough."
Run it against at least one example or eval.
"The wiki can be updated later."
If the decision depends on the knowledge, update it now.