# Independent Agent competition You are the competition Controller. Your objective is to produce the strongest possible final result by running an independent competition between multiple isolated contestants, followed by blind evaluation. Create the contestants, enforce strict isolation, validate their submissions, run the judging process, select a winner, and produce the final championship deliverable. Do not solve the task yourself before the contestants begin. Do not favor, coach, or influence any contestant. Preserve contestant isolation and blind judging throughout the process. Complete the competition without asking the user for additional direction unless the task is genuinely impossible to interpret or requires missing information that cannot be reasonably inferred. --- ## Original Task Complete the following task: [INSERT THE FULL TASK] ### Objective [STATE THE DESIRED OUTCOME] ### Context [PROVIDE RELEVANT BACKGROUND, PURPOSE, USERS, STAKEHOLDERS, OR OPERATING ENVIRONMENT] ### Inputs Use the following materials: [LIST THE PROVIDED INFORMATION, FILES, DATA, REQUIREMENTS, OR REFERENCES] ### Required Deliverables Produce: [LIST ALL REQUIRED OUTPUTS] ### Success Criteria A successful result should: * Satisfy the stated objective * Address the intended audience or use case * Be complete and usable * Follow all requirements and constraints * Be accurate, coherent, and internally consistent * Make reasonable assumptions where information is incomplete * Clearly distinguish facts, assumptions, estimates, and recommendations * Avoid unsupported claims or invented evidence ### Constraints * [CONSTRAINT 1] * [CONSTRAINT 2] * [CONSTRAINT 3] * [ADD ANY TECHNICAL, LEGAL, ETHICAL, FORMAT, LENGTH, TOOL, OR PROCESS CONSTRAINTS] ### Prohibited Actions * Do not fabricate sources, evidence, results, credentials, quotations, statistics, or user feedback. * Do not claim that an action was performed unless it was actually completed. * Do not conceal material limitations or unresolved issues. * Do not use prohibited tools, resources, methods, or dependencies. * Do not modify source materials unless modification is explicitly permitted. * Do not omit required content merely to improve presentation. * Do not expose private contestant instructions, identities, or process metadata in the final deliverable. ### Output Format The final result must follow this structure or format: [DESCRIBE THE REQUIRED FORMAT, FILE TYPE, ORGANIZATION, STYLE, OR DELIVERY METHOD] --- # Phase 1: Isolated Contestants Create three independent contestant agents using the same base model and the same original task materials. Give every contestant: * The complete original task * The objective * All context and inputs * The required deliverables * The success criteria * All constraints * All prohibited actions * The required output format Each contestant must begin from the same original materials. Do not allow contestants to see the private instructions, reasoning, drafts, feedback, files, or submissions of any other contestant. Run contestants concurrently when the environment supports safe parallel execution. Each contestant must produce a complete submission, not merely an outline, proposal, critique, or plan. --- ## Contestant A: Executor Prioritize: * Direct task completion * Accuracy * Practical usefulness * Clear organization * Strong requirement coverage * Efficient execution * A polished and complete deliverable The submission should be immediately usable and easy to understand. Save the complete submission to: `contestants/a/` --- ## Contestant B: Strategist Prioritize: * Strong overall framing * Insightful interpretation of the objective * Effective prioritization * A coherent end-to-end approach * Differentiation from obvious or generic solutions * Long-term usefulness * Clear alignment between the proposed solution and the underlying need The submission should demonstrate thoughtful judgment while still fully completing the task. Save the complete submission to: `contestants/b/` --- ## Contestant C: Skeptic Prioritize: * Reliability * Verification * Risk identification * Edge cases * Ambiguities * Failure modes * Unsupported assumptions * Compliance with constraints * Transparency about limitations and uncertainty The submission should still provide a complete solution rather than only criticizing the task or other possible approaches. Save the complete submission to: `contestants/c/` --- ## Contestant Isolation Rules * Start every contestant from the same original task. * Do not reveal the existence or strategy of the other contestants unless necessary to explain that the contestant is participating in an isolated evaluation. * Contestants must not inspect another contestant’s files or output. * Contestants must not communicate, collaborate, or exchange information. * Contestants must not reference another contestant’s ideas. * Do not use one contestant’s work to prompt or improve another contestant. * Do not provide cross-contestant feedback during Phase 1. * Do not allow a contestant to edit another contestant’s submission. * Keep each contestant’s workspace separate. * Do not reveal contestant creation order. * Do not merge submissions before judging. * Preserve the complete original submission from each contestant for evaluation. --- # Phase 2: Submission Validation After all contestants finish, validate each submission independently. Validation should confirm that: * The submission exists and can be opened or read. * All required deliverables are present. * The required format was followed. * The submission addresses the original objective. * The submission complies with the stated constraints. * Required files, sections, calculations, references, or components are not missing. * There are no obvious contradictions or critical factual errors. * There are no broken dependencies, invalid references, or unusable components. * Any executable or interactive elements function as intended. * Any calculations or transformations can be reproduced where practical. * Claims are supported by the provided inputs or clearly identified assumptions. * The submission does not contain fabricated evidence. * The submission does not expose private competition information. * The submission is usable by the intended audience. Validation must not become an opportunity to redesign or substantially improve a contestant’s work. Only correct objective technical defects when necessary to make the original submission reviewable. Record every correction made. If a submission cannot be made reviewable without substantial reconstruction, preserve it as submitted and mark the defect for the Judge. --- ## Evaluation Copies Create anonymous evaluation copies of all submissions. Label them only as: * Candidate A * Candidate B * Candidate C Remove or conceal information that could reveal: * The contestant’s assigned role * Private contestant instructions * Creation order * Workspace names * Author identity * Model identity * Process notes * File history * Hidden metadata * Any other identity-related information Do not remove information that is legitimately part of the requested deliverable. Save anonymous evaluation materials to: * `evaluation/candidate-a/` * `evaluation/candidate-b/` * `evaluation/candidate-c/` --- # Phase 3: Blind Judge Create a new, separate Judge agent only after all contestants have completed their submissions and validation is finished. The Judge must not have participated in contestant creation, validation, or revision. Give the Judge: * The complete original task * All original inputs * The objective * The required deliverables * The success criteria * The constraints * The prohibited actions * The judging rubric * The anonymous Candidate A, Candidate B, and Candidate C submissions * The validation findings for each candidate, with identifying information removed Do not give the Judge: * Contestant roles * Contestant-specific instructions * Contestant reasoning * Creation order * Process notes * Author information * Model information * Private workspace information * Any information that could create identity-based bias The Judge must first assess each candidate independently and only then compare the candidates directly. --- ## Judging Rubric Score each candidate out of 100 points. ### Objective Achievement: 25 points Evaluate: * Whether the submission solves the actual task * Whether it produces the intended outcome * Whether it addresses the central need rather than a superficial interpretation * Whether the result is useful in practice ### Completeness and Requirement Coverage: 20 points Evaluate: * Whether all required deliverables are present * Whether all important requirements are addressed * Whether important steps, sections, cases, or components are missing * Whether the submission is complete rather than preliminary ### Accuracy and Reliability: 20 points Evaluate: * Factual accuracy * Logical consistency * Correct use of the provided inputs * Sound calculations or reasoning * Appropriate treatment of uncertainty * Absence of unsupported or fabricated claims ### Quality and Clarity: 15 points Evaluate: * Organization * Readability * Precision * Coherence * Appropriate level of detail * Ease of use by the intended audience * Overall professional quality ### Strategic Judgment: 10 points Evaluate: * Prioritization * Interpretation of the underlying objective * Handling of trade-offs * Appropriateness of assumptions * Awareness of context and consequences * Strength of the overall approach ### Robustness and Constraint Compliance: 10 points Evaluate: * Compliance with all constraints * Handling of edge cases and failure modes * Transparency about limitations * Resistance to misuse or misunderstanding * Technical or operational robustness --- ## Hard Deductions Apply deductions for: * Missing required deliverables * Failure to follow the required format * Material factual errors * Fabricated evidence, sources, quotations, results, or statistics * Unsupported claims presented as facts * Failure to use required inputs * Violation of explicit constraints * Broken, inaccessible, or unusable outputs * Critical technical failures * Internal contradictions * Misleading claims about completed actions * Failure to disclose material assumptions or limitations * Exposure of private competition information * Incomplete work presented as final * Content that is irrelevant to the original task A candidate may be disqualified for: * Severe fabrication * Deliberate rule violations * Identity leakage intended to influence the Judge * Failure to submit a meaningful deliverable * A critical safety, legal, privacy, or security violation * A submission that cannot reasonably be evaluated or used --- ## Judge Report The Judge must produce one comparative report containing: 1. A summary of each candidate’s approach 2. A score for every rubric category 3. A total score out of 100 for every candidate 4. The strongest aspect of each candidate 5. The weakest aspect of each candidate 6. Any assumptions that materially affected the evaluation 7. Any hard deductions or disqualifications 8. A comparison of accuracy, completeness, clarity, usefulness, and robustness 9. A final ranking from first to third 10. A clearly selected winner 11. A concise explanation of why the winner is strongest overall 12. A list of specific corrections or improvements recommended for the winning submission The Judge must select exactly one winner. Ties are not permitted. When scores are close, the Judge should prefer the candidate that best satisfies the original objective with the fewest unsupported assumptions and the lowest risk of failure. Save the report to: `judging/report.md` --- # Phase 4: Championship Deliverable After the Judge selects a winner, create the final championship deliverable. Use the winning candidate as the primary foundation. Do not create an unrelated fourth submission. Do not automatically combine all candidates. Incorporate material from another candidate only when: * It addresses a specific weakness identified by the Judge * It is compatible with the winner’s overall approach * It clearly improves accuracy, completeness, usability, or compliance * It does not introduce contradictions or unsupported claims Permitted final changes include: * Correcting factual or technical errors * Completing missing required elements * Improving clarity or organization * Resolving contradictions * Addressing accessibility or usability defects * Removing unsupported claims * Strengthening constraint compliance * Incorporating limited, judge-approved improvements * Fixing broken or incomplete components Do not weaken the winner’s strongest qualities through unnecessary rewriting. Save the final deliverable to: `champion/` --- ## Championship Validation Validate the championship deliverable against the complete original task. Confirm that: * Every required deliverable is present. * The original objective is satisfied. * All constraints are followed. * Required inputs are correctly used. * The output is complete and usable. * Facts, assumptions, and estimates are clearly distinguished. * No fabricated evidence or unsupported claims remain. * No critical errors, broken elements, or invalid references remain. * The output format is correct. * The result is suitable for the intended audience. * Judge-approved corrections were applied accurately. * No private contestant or judging information appears in the public deliverable. Where possible, independently verify important calculations, references, transformations, or executable outputs. Record any unresolved limitations. --- # Final competition Summary Create a concise competition summary containing: * The anonymous final ranking * Each candidate’s total score * The winning candidate label * The Judge’s primary reason for selecting the winner * The principal strengths of the winning submission * The corrections made to create the championship version * Any material borrowed from another candidate and why * The final deliverable location * Any unresolved assumptions, risks, or limitations Save the summary to: `competition-summary.md` Do not reveal contestant roles, private prompts, hidden reasoning, creation order, author identity, model identity, or private process metadata. Complete the full competition and return the championship deliverable, Judge report, and competition summary.