Agent Skills: Doubt-Driven Development

Doubt-Driven Development

Three principles

1. A confident answer is not a correct one. Long sessions turn assumptions into "facts" without anyone noticing.
2. The reviewer gets the artifact and the contract, never your reasoning or your conclusion. A prompt biased to disprove, not approve, catches what a summary-and-agree pass would wave through.
3. The loop is bounded, not recursive. It stops at trivial findings, three cycles, or user override — and three unresolved cycles is information about the artifact, not a reason to grind a fourth.

Overview

Doubt-driven development materializes a fresh-context reviewer, biased to disprove, not approve, before any non-trivial output stands.

This is an in-flight posture, not a post-hoc gate. A verdict on a finished artifact arrives too late to change direction cheaply. Doubt-driven cross-examines non-trivial decisions while course-correction still costs little.

When to Use

A decision is non-trivial when at least one of these holds:

• It introduces or modifies branching logic
• It crosses a module or service boundary
• It asserts a property the type system or compiler cannot verify (thread safety, idempotence, ordering, invariants)
• Its correctness depends on context the future reader cannot see
• Its blast radius is irreversible (production deploy, data migration, public API change)

Apply the skill when:

• About to make an architectural decision under uncertainty
• About to commit non-trivial code
• About to claim a non-obvious fact ("this is safe", "this scales", "this matches the spec")
• Working in code you do not fully understand

When NOT to use:

• Mechanical operations (renaming, formatting, file moves)
• Following a clear, unambiguous user instruction
• Reading or summarizing existing code
• One-line changes with obvious correctness
• Pure tooling operations (running tests, listing files)
• The user has explicitly asked for speed over verification

Loading Constraints

This skill runs in the main-session orchestrator, where Step 3 (DOUBT) can spawn a fresh-context reviewer.

• Do not add this skill to a subagent's skills: frontmatter. A subagent that follows Step 3 would spawn another subagent, the orchestration anti-pattern forbidden by references/orchestration-patterns.md ("subagents do not invoke other subagents").
• If you find yourself applying this skill from inside a subagent context (where the harness blocks nested subagent spawn): surface to the user that doubt-driven cannot run nested and let the main session handle it. As a last resort only, a degraded self-questioning fallback exists: rewrite ARTIFACT + CONTRACT as a fresh self-prompt with a hard mental separator from your prior reasoning, and walk Steps 1 to 5. This is not fresh-context review (you carry your own context with you), so flag the result as degraded and prefer escalation whenever the user is reachable.

The Process

Copy this checklist when applying the skill:

Doubt cycle:
- [ ] Step 1: CLAIM — wrote the claim + why-it-matters
- [ ] Step 2: EXTRACT — isolated artifact + contract, stripped reasoning
- [ ] Step 3: DOUBT — invoked fresh-context reviewer with adversarial prompt
- [ ] Step 4: RECONCILE — classified every finding against the artifact text
- [ ] Step 5: STOP — met stop condition (trivial findings, 3 cycles, or user override)

Step 1: CLAIM: Surface what stands

Name the decision in two or three lines. The format holds across stacks:

CLAIM: "The new caching layer is thread-safe under the
        read-heavy workload described in the spec."
WHY THIS MATTERS: a race here corrupts user data and is
                  hard to detect in QA.

CLAIM: "The Python migration script is idempotent — re-running
        after a partial failure converges the schema to one state."
WHY THIS MATTERS: a retried deploy must not double-apply or
                  half-apply the migration.

If you cannot write the claim that compactly, you have a vibe, not a decision. Surface it before scrutinizing it.

Step 2: EXTRACT: Smallest reviewable unit

A fresh-context reviewer needs the artifact and the contract, not the journey.

• Code: the diff or the function, not the whole file
• Decision: the proposal in 3 to 5 sentences plus the constraints it has to satisfy
• Assertion: the claim plus the evidence that supposedly supports it (separate from the Step 1 CLAIM block, which is the orchestrator's hypothesis under scrutiny)

Strip your reasoning. Hand over conclusions and you get back validation of those conclusions. The unit must be small enough that a reviewer holds it in mind in one read; a 500-line PR decomposes first.

Step 3: DOUBT: Invoke the fresh-context reviewer

The reviewer's prompt must be adversarial. Framing decides the answer.

Adversarial review. Find what is wrong with this artifact.
Assume the author is overconfident. Look for:
- Unstated assumptions
- Edge cases not handled
- Hidden coupling or shared state
- Ways the contract could be violated
- Existing conventions this might break
- Failure modes under unexpected input

Do NOT validate. Do NOT summarize. Find issues, or state
explicitly that you cannot find any after thorough examination.

ARTIFACT: <paste artifact>
CONTRACT: <paste contract>

Pass ARTIFACT + CONTRACT only. Do NOT pass the CLAIM. Handing the reviewer your conclusion biases it toward agreement. The reviewer must independently determine whether the artifact satisfies the contract.

Spawn the reviewer as a fresh subagent with isolated context. A general reviewer role may default to a balanced verdict (strengths alongside weaknesses), but doubt-driven needs issues-only output. The adversarial prompt above takes precedence over any default response shape: paste it verbatim so it overrides. If a reviewer's response shape cannot be overridden cleanly, fall back to a generic subagent with the adversarial prompt.

Cross-model escalation

A single-model reviewer shares blind spots with the original author; a colder, different-architecture model catches them. Doubt-driven is already opt-in for non-trivial decisions, so offering cross-model within that scope is part of the skill's value.

Interactive sessions: always offer. Never silently skip.

Step 1: Ask the user

After the single-model review in Step 3 above, but before RECONCILE, pause and ask:

"Single-model review complete. Want a cross-model second opinion? Options: Gemini CLI, Codex CLI, manual external review (you paste it elsewhere), or skip."

This question is mandatory in every interactive doubt cycle, even on artifacts that feel low-stakes. The user, not the agent, decides whether the cost is worth it. The agent's job is to surface the choice.

Step 2: If the user picks a CLI: verify, then invoke. Read references/cross-model-invocation.md once the user has named a specific CLI (Gemini, Codex, or another external reviewer) — it has the PATH/version verification steps, the mktemp+heredoc invocation template for both tools, and the shell-escaping caveat for passing ARTIFACT + CONTRACT without letting embedded backticks or $(...) execute.

Never interpolate the artifact into a shell-quoted argument. Code, markdown, and review prompts routinely contain backticks, $(...), and quote characters that will either truncate the prompt or execute embedded shell. Write the full prompt to a file and pipe it through stdin.

Step 3: If the CLI is unavailable or fails

Surface the failure explicitly. Offer: run it manually, try a different tool, or skip. Do not silently fall back to single-model. The user should know cross-model did not happen.

Step 4: If the user skips

Acknowledge the skip in the output ("Proceeding with single-model findings only") and continue to RECONCILE. Skipping is fine; silent skipping is not.

Non-interactive contexts (CI, autonomous loops, scheduled runs):

• Cross-model is skipped, and the skip must be announced in the output: "Cross-model skipped: non-interactive context."
• Never invoke an external CLI without explicit user authorization. This is a load-bearing safety property.

Cross-model adds cost, latency, and tool fragility. The agent surfaces the choice every cycle; the user decides whether this artifact warrants it.

Step 4: RECONCILE: Fold findings back

The reviewer's output is data, not verdict. You are still the orchestrator. Re-read the artifact text against each finding before classifying; rubber-stamping the reviewer is the same failure mode as ignoring it.

For each finding, classify in this precedence order (first matching class wins):

1. Contract misread: reviewer flagged something specifically because the CONTRACT you provided was unclear or incomplete. Fix the contract first, re-classify on the next cycle.
2. Valid + actionable: real issue requiring a change to the artifact. Change it, re-loop.
3. Valid trade-off: issue is real but the cost of fixing exceeds the cost of accepting. Document the trade-off explicitly so the user sees it.
4. Noise: reviewer flagged something correct under context it did not have. Note it, move on, and ask whether adding that context to the contract would have prevented the false flag.

A fresh reviewer can be wrong because it lacks context. Do not defer just because it is "fresh."

Step 5: STOP: Bounded loop, not recursion

Stop when:

• Next iteration returns only trivial or already-considered findings, or
• 3 cycles completed (escalate to user, do not grind a fourth alone), or
• User explicitly says "ship it"

If after 3 cycles the reviewer still surfaces substantive issues, the artifact may not be ready. Surface this to the user. Three unresolved cycles is information about the artifact, not a reason to keep looping.

If 3 cycles is "obviously insufficient" because the artifact is large: the artifact is too big. Return to Step 2 and decompose. Do not lift the bound.

Common Rationalizations

| Rationalization | Reality | |---|---| | "I'm confident, skip the doubt step" | Confidence correlates poorly with correctness on novel problems. Moments of certainty are exactly when blind spots hide. | | "Spawning a reviewer is expensive" | Debugging a wrong commit in production is more expensive. The check is bounded; the bug is not. | | "The reviewer will just nitpick" | Only if unscoped. Constrain the prompt to "issues that would make this fail under the contract." | | "I'll do doubt at the end" | A final pre-merge gate is too late. Doubt-driven catches wrong directions early, when course-correction is cheap. | | "If I doubt every step I'll never ship" | The skill applies to non-trivial decisions, not every keystroke. Re-read "When NOT to use." | | "Two opinions are always better than one" | Not when the second has less context and produces noise. Reconcile, do not defer. | | "The reviewer disagreed so I was wrong" | The reviewer lacks your context. Disagreement is information, not verdict. Re-read the artifact, classify, then decide. | | "Cross-model is always better" | Cross-model catches blind spots a single model shares with itself, but it adds cost and tool fragility. Offer it every interactive doubt cycle. The user decides whether the artifact warrants it. | | "User said yes once, so I can keep invoking the CLI" | Each invocation is its own authorization. The artifact, the prompt, and the flags change between calls. Re-confirm the exact command with the user before every run. |

Red Flags

• Skipping doubt under time pressure on a high-stakes decision
• Re-spawning fresh-context on an unchanged artifact (you get the same findings; you are stalling)
• Doubt theater (checkable signal): across 2 or more cycles where the reviewer surfaced substantive findings, zero findings were classified as actionable. You are validating, not doubting. Stop and escalate.
• Doubting only after committing. That is a post-hoc gate, not doubt-driven development

Verification

After applying doubt-driven development:

• [ ] Every non-trivial decision (per the definition above) was named explicitly as a CLAIM before standing
• [ ] At least one fresh-context review per non-trivial artifact (for behavioral claims, a failing disproof test can satisfy this)
• [ ] The reviewer received ARTIFACT + CONTRACT, not the CLAIM, not your reasoning
• [ ] The reviewer's prompt was adversarial ("find issues"), not validating ("is it good")
• [ ] Findings were classified against the artifact text (not rubber-stamped) using the precedence: contract misread / actionable / trade-off / noise
• [ ] A stop condition was met (trivial findings, 3 cycles, or user override)
• [ ] In interactive mode, cross-model was explicitly offered to the user (regardless of artifact stakes) and the response was acknowledged in the output
• [ ] In non-interactive mode, cross-model was skipped and the skip was announced
• [ ] Any external CLI invocation was preceded by a PATH check, a working-binary test, syntax confirmation with the user, and explicit authorization to run