Doubt-Driven Development

Overview

For most code, optimism is fine: try it, run the test, fix what breaks. For changes you cannot cheaply undo, optimism is a liability. Doubt-driven development inverts the default — you actively try to refute your chosen approach before you build it, and you require concrete evidence to clear each doubt.

Core principle: default to refuted when uncertain. A doubt you cannot disprove with evidence is a reason to choose a safer path, not a reason to proceed and hope.

This is review applied during the build, by you, before a single line is committed. It is distinct from [[validate]], the post-hoc multi-agent loop that runs after implementation — that catches what got built wrong; this catches what should never have been built that way.

When to Use

Apply when the change is high-stakes or hard to reverse:

• Schema / data migrations — backfills, column drops, type changes, anything that mutates existing rows
• Authentication / authorization — session handling, token validation, permission checks
• Payments / billing — charges, refunds, idempotency, money math
• Public API surface — endpoints or contracts other teams or users depend on
• Destructive / irreversible operations — deletes, overwrites, bulk updates, cache purges
• Concurrency — locking, ordering, race-prone shared state

Routine, easily-reverted CRUD does not need this. Reserve the ceremony for blast radius.

Process

1. Name the decision

State the high-stakes decision in one sentence, explicitly. "I will run the backfill in a single transaction over the whole table." Vague intent cannot be refuted. If you cannot name it crisply, you do not understand it well enough to commit to it.

2. Adopt the skeptic stance — refute, don't defend

Before implementing, switch jobs. Your job is now to break the chosen approach, not to justify it. Produce, in writing:

• How it fails — concrete failure modes, not "it should be fine"
• What it breaks — existing callers, data, invariants, in-flight requests
• What would disprove it — the specific observation that would prove the approach wrong (the row count mismatches; the lock is held > N seconds; the old client gets a 500)

A skeptic who lists zero failure modes is not trying. Push until you have real ones.

3. Demand evidence per doubt

Each doubt is cleared only by concrete evidence — a test, a doc/spec, or a spike (throwaway experiment against real-ish data). "I'm pretty sure" does not clear a doubt. If clearing a doubt requires a behavior, write the failing test first ([[test-driven-development]]) so the evidence is durable. Spike code lives in the task worktree (or a scratch worktree via [[use-git-worktree]]) and is never committed — never in the shared workspace, where it pollutes other work.

4. Resolve: clear, refute, or escalate

• Cleared — evidence disproves the doubt. Proceed on that point.
• Refuted — evidence confirms the failure, or you cannot get evidence after honest effort. Pick a safer, reversible path (expand/contract migration instead of in-place; feature flag; dry-run mode; smaller batches) or escalate to the human with the doubt and what you tried.

Uncertainty after honest effort counts as refuted, not as permission to proceed.

Rationalizations

Red Flags - STOP

• Listing only reasons the approach works, none of how it fails
• "Should be fine" / "I'm pretty sure" / "probably" as the basis to proceed
• Clearing a doubt with reasoning instead of a test, doc, or spike
• Treating "I couldn't find a problem" as "there is no problem"
• Running the irreversible step with no dry-run, no flag, no rollback
• Proceeding while a named doubt is still open

Verification

Before committing the high-stakes change:

• The decision is named in one explicit sentence
• At least one concrete failure mode and its disproving observation are written down
• Every doubt is either cleared by evidence (test/doc/spike) or marked refuted
• No open doubt remains; uncertain ones were treated as refuted
• If refuted, a safer reversible path was taken or the human was escalated to
• The irreversible step has a rollback, dry-run, or flag — or proof it needs none

Can't check every box? The doubt is not cleared. Do not commit.