Claude-led / Sol-challenger workflow

The operating principle

This is most useful when the question has material tradeoffs: strategy, architecture, research synthesis, prioritization, operating-model design, or a consequential change. It is overkill for routine drafting.

Roles and boundaries

The workflow

1. Establish the decision contract

   Claude records the decision, desired outcome, constraints, evidence standard, and the decision deadline. It distinguishes facts, inferences, assumptions, and unknowns from the outset.

2. Create a shared source packet

   Place the question, source material, constraints, and decision criteria in working/source-packet.md. Keep Claude's preliminary recommendation out of this packet for the first Sol pass.

3. Have Claude form a preliminary position

   Claude writes its thesis separately in working/claude-position.md. This preserves a genuine independent point of comparison rather than asking Sol to merely edit Claude's thinking.

4. Commission a blind Sol challenge

   Use a fresh Sol task with the source packet only. Ask for a recommendation, the strongest contrary view, material risks, missing evidence, and explicit owner decisions.

5. Reconcile in Claude

   Claude compares the two positions. It classifies disagreements as factual, inferential, or values-based, then states which challenger findings materially strengthen or weaken its original thesis.

6. Run one focused resolution pass, if needed

   Only if a material disagreement remains, ask Sol to evaluate that specific point. Do not reopen settled ground. Two challenger rounds are normally the limit.

7. Publish the decision brief

   Claude writes the final brief: recommendation, evidence, material dissent, residual uncertainty, reversible next step, and conditions that would cause reconsideration.

Copy-ready prompts

1. Claude decision contract

You are the lead analyst for this session.

We need to decide:
[decision or research question]

Practical outcome:
[what this enables]

Constraints:
- [constraint]
- [constraint]

Evidence standard:
- Distinguish observed facts, inference, assumptions, and unknowns.
- Cite important claims.
- Do not resolve material uncertainty through confidence alone.

Create working/source-packet.md and develop your preliminary position
separately in working/claude-position.md. You may commission Sol as an
independent challenger, but you own the final synthesis.

2. First Sol challenge

/codex:rescue --fresh --model gpt-5.6-sol --effort high
Read working/source-packet.md. Do not edit any files.

Act as an independent strategic challenger. Produce:

1. Your recommended conclusion.
2. The strongest reasoning supporting it.
3. The three most important risks or unsupported assumptions.
4. The strongest credible alternative.
5. Evidence that would change your conclusion.
6. Decisions requiring owner judgment rather than model judgment.

Do not assume the lead analyst's unstated conclusion. Return a concise,
evidence-ranked analysis.

3. Claude reconciliation

Compare working/claude-position.md with Sol's challenge.

Create a decision table with:
- Point of agreement or disagreement
- Claude's position
- Sol's position
- Evidence supporting each
- Whether the conflict is factual, inferential, or values-based
- Proposed resolution
- Confidence and remaining uncertainty

Do not treat agreement as proof or disagreement as a problem to eliminate.
Identify which Sol findings materially improve or weaken the original position.

4. Focused Sol resolution pass

/codex:rescue --resume
The lead analyst disagrees with your conclusion about [specific issue].

Its reasoning is:
[concise reasoning]

Evaluate only this disagreement:
1. Is the difference caused by facts, assumptions, or values?
2. What evidence would best resolve it?
3. Does the new reasoning change your conclusion?
4. If not, state the owner decision plainly.

Do not reopen settled portions of the analysis.

5. Final Claude brief

Produce the final owner-ready brief.

Include:
- Decision and recommendation
- Why this is the recommended path
- Evidence supporting it
- Material contribution from Sol
- Genuine remaining disagreement
- Risks and mitigations
- Reversible next step
- What would cause reconsideration

Do not imply consensus where none exists. Clearly identify any decision
that still requires owner judgment.

How to use it by work type

For ordinary work, use Terra at medium effort for the delegated pass. Reserve Sol at high effort for consequential choices, difficult contradiction resolution, or a genuine red-team assignment.

Guardrails

• Keep one lead: Claude owns continuity and synthesis.
• Use one canonical source packet, not scattered chat context.
• Keep the first Sol pass blind to Claude's conclusion.
• Ask Sol not to edit the canonical artifact unless you deliberately authorize a writing task.
• Require Claude to explicitly account for material Sol findings.
• Limit the process to two Sol rounds.
• Do not use an automatic review gate for this workflow.
• Do not treat model agreement as validation; check sources and real-world evidence.
• Keep API-key environment variables absent when using subscription-only authentication.
• End with a reversible next step whenever uncertainty is still material.

Definition of a good run

A good run produces a decision artifact that is stronger than Claude alone, without multiplying work. You should be able to point to: the original decision, the independent challenge, the specific insight that changed or strengthened the recommendation, the remaining uncertainty, and the next action.