Harness Interview Workflow
Overview
Use this skill to slow Codex down before expensive work. The skill turns a vague or complex request into a structured interview: discover unknowns, ask only the highest-leverage questions, propose routes, and wait for explicit approval before execution.
Operating Contract
- Restate the user's goal and likely deliverable.
- Do a blindspot scan before proposing execution.
- Ask at most 1-3 questions per round.
- Prioritize questions whose answers change architecture, risk, deliverable shape, or verification.
- Give 2-3 route options when enough context exists.
- Stop after the plan if the user asked for a plan or has not approved execution.
- If the user confirms execution, create a normal implementation plan and use approval gates for risky actions.
- Maintain implementation notes when execution uncovers edge cases or deviations.
Blindspot Scan
Classify the request into four unknown types:
Known knowns:
Known unknowns:
Unknown knowns:
Unknown unknowns:
Highest-risk misunderstanding:
Smallest next move:
Use plain Chinese by default when working with Anita unless the user asks otherwise.
Interview Mode
Ask only the next 1-3 questions. Prefer this order:
- What answer would change the architecture or core route?
- What answer would change risk boundaries or approval requirements?
- What answer would change the final artifact format?
- What answer would change verification or acceptance criteria?
- What answer only affects style?
Avoid long forms. If a question is optional, continue with a conservative assumption and label it.
Reference Handling
When the user provides an article, repo, screenshot, code, old project, template, or example, classify the reference before using it:
must reuse: preserve structure, semantics, names, or content.style reference: borrow feel or pattern, do not copy.do not imitate: use only to understand constraints.fact source: cite or preserve factual provenance where needed.
If the classification is unclear, ask one short clarification question.
Route Proposal
After enough context, provide options like:
方案 A:最小可行版
适合:
会做:
不会做:
风险:
验证:
方案 B:稳妥完整版
适合:
会做:
不会做:
风险:
验证:
方案 C:高质量扩展版(可选)
适合:
会做:
不会做:
风险:
验证:
推荐:
等待用户确认:
Do not execute until the user chooses a route or says a clear confirmation such as 确认执行, 继续, or 按推荐方案执行.
Execution After Approval
When the user confirms execution, use this plan shape:
Goal:
Success criteria:
Current context:
Constraints:
Risks:
Approval required:
Workflow artifact path:
Work packets:
Integration policy:
Verification:
Reusable artifacts:
For complex or reusable work, create a workflow artifact:
.workflow/<slug>/
|-- plan.md
|-- state.json
|-- orchestration.md
|-- packets/
|-- results/
`-- final-report.md
If codex-dynamic-workflows is available and the work is broad, use it together with this skill.
Implementation Notes
For execution tasks, keep notes in the task's natural location, usually implementation-notes.md or .workflow/<slug>/results/implementation-notes.md:
# Implementation Notes
## Original Plan
## Decisions
## Deviations
## Edge Cases
## User Confirmations
## Verification
## Follow-up
When an edge case forces a deviation, choose a conservative path, record why, and continue only if it remains inside the user's approval boundary.
Final Response
For plan-only work, return:
- blindspot scan summary
- questions asked or assumptions made
- route options
- recommended route
- clear approval question
For executed work, return:
- completed artifacts and paths
- verification evidence
- skipped checks
- remaining risks
- how to reuse the workflow
For complex changes, add a short after-action quiz so the user can confirm they understand what changed.