dynamic-workflow

name	dynamic-workflow
description	Use when the user asks for a dynamic workflow, multi-agent workflow, task-specific harness, complex task orchestration, fan-out synthesis, adversarial verification, generate-and-filter exploration, tournament-style solution search, or loop-until-done execution in Codex.
metadata	{"short-description":"Dynamic task harness for Codex"}

Dynamic Workflow

Use this skill for complex tasks that benefit from a temporary task-specific execution harness.

Core Contract

Before execution, define:

• target outcome
• success criteria
• constraints
• workflow shape
• verification evidence
• stop condition

Keep the main thread responsible for integration, risk decisions, and final verification.

Workflow Shapes

Classify And Act

Use for mixed queues, issue lists, review comments, inboxes, and task triage.

Steps:

1. Classify each item by type, risk, owner, and required evidence.
2. Route simple items to direct execution.
3. Route codebase lookup to explore.
4. Route implementation to executor.
5. Route final claim checks to verifier.

Fan Out And Synthesize

Use for repo mapping, migration scoping, backlog analysis, research, and large comparison tasks.

Steps:

1. Split the task into independent questions.
2. Assign only independent sidecar work to subagents when the user requested multi-agent or parallel work.
3. Continue local integration while sidecar work runs.
4. Synthesize findings into ranked actions.
5. Verify the highest-value conclusion with direct evidence.

Adversarial Verification

Use for security-sensitive work, code review, architectural claims, data handling, auth, billing, migrations, and bug-fix closure.

Steps:

1. State the claim being verified.
2. Define a rubric before judging.
3. Ask critic or verifier for an independent attack when subagent use is authorized.
4. Fix issues found.
5. Report evidence and residual risk.

Generate And Filter

Use for architecture options, UX direction, naming, migration strategy, prompt variants, and technical design alternatives.

Steps:

1. Generate 3-5 candidate approaches.
2. Define a scoring rubric before selection.
3. Score each candidate against the rubric.
4. Select one path.
5. Execute or produce a decision record.

Tournament

Use for high-value design, algorithm, architecture, or prompt choices.

Steps:

1. Create several candidate strategies.
2. Evaluate each with the same rubric.
3. Advance the strongest candidate.
4. Run a final critic/verifier pass when authorized.
5. Record the selected strategy and rejected alternatives.

Loop Until Done

Use for failing tests, repeated build errors, flaky bugs, cleanup loops, and diagnostic repair tasks.

Steps:

1. Define the stop condition.
2. Run the smallest useful diagnostic.
3. Apply the smallest fix.
4. Verify.
5. Repeat until the stop condition is met or a hard blocker is identified.

Subagent Routing

Use subagents only when the user explicitly asks for subagents, delegation, parallel agents, or a multi-agent workflow.

Recommended roles:

• explore: repository search, file mapping, symbol lookup
• researcher: external official documentation and source-backed references
• dependency-expert: package or SDK evaluation
• executor: scoped implementation
• critic: plan, design, and claim challenge
• verifier: completion evidence and test adequacy
• test-engineer: test strategy and regression coverage
• writer: documentation and migration notes

Delegated tasks must be independent, bounded, and useful to the main thread. The main thread owns final integration and verification.

Completion Report

Return:

• workflow shape used
• actions taken
• evidence gathered
• result
• remaining risks
• concrete next step when action remains