AI 面试陪练官
This is a local AI Skill for interview practice. The formal AIPC-compatible
entry point is scripts\run.ps1; it starts a short-lived client, talks to a
resident local server over a named pipe, and keeps model and answer data on the
machine. Do not fall back to a cloud service.
Usage
| Intent | Command |
| --- | --- |
| 查看岗位和题库 | scripts\run.ps1 roles |
| 评价第一次回答 | scripts\run.ps1 coach --role product-manager --answer "<回答>" |
| 对比第二次回答 | scripts\run.ps1 coach --role product-manager --answer "<首答>" --second-answer "<二答>" |
| 导出练习报告 | scripts\run.ps1 report --input examples\interview-coach\pm-second-answer.json --output build\interview-practice-report.md --overwrite |
| 查看本地服务状态 | scripts\run.ps1 status |
If first-run model preparation exceeds the host timeout, rerun:
scripts\run.ps1 --continue
For repository development and CI on non-Windows platforms, the equivalent offline verifier is:
python3 scripts/verify_interview_submission.py
Interpreting Output
The reply is JSON with Chinese user-facing fields:
roles: supported interview roles and question IDs.result.first_answer: six-dimension score, priority issues, follow-ups, and a safe rewrite that keeps missing facts as placeholders.result.second_answerandcomparison: score deltas after the retry.markdownandfilename: portable local practice report.
Important
scripts\run.ps1is the supported AIPC host interface. Do not ask the host to call helper scripts directly.- The first call may prepare a local environment and model. Use
--continuewhen the client exits with code3. - On unsupported hardware, the entry script prints a platform error and exits
with code
1. - The package contains no API key, no model weights, and no virtual environment.
- The AIPC contract mapping and host-specific adaptations are documented in
references/aipc-local-skill-standard.md.
Verification Matrix
| Operation | Expected postcondition | Valid evidence | Failure and recovery |
| --- | --- | --- | --- |
| V1 - List roles | Roles and questions are available | JSON from scripts\run.ps1 roles | Stop if the requested role is missing |
| V2 - Score first answer | Six scores, issues, follow-ups, and rewrite are returned | JSON from scripts\run.ps1 coach ... | Ask for a longer answer if the input is too short |
| V3 - Compare second answer | The result contains comparison.total_delta and dimension deltas | JSON from --second-answer output | Report no improvement when the score does not increase |
| V4 - Resume first-run setup | Pending request resumes after model download timeout | Exit code 3, then scripts\run.ps1 --continue | Preserve the pending request and rerun |
| V5 - Cross-role evaluation | All eight predefined interview-question tests pass across four roles | JSON from python3 scripts/interview.py evaluate | Inspect the failed case instead of changing thresholds silently |
| V6 - Export practice report | Markdown contains both attempts, six-dimension comparison, follow-ups, next practice actions, and usage boundaries | File from python3 scripts/interview.py report ... | Keep the JSON result when report writing is not requested |
| V7 - Offline submission verification | Roles, coaching, report, localhost API, Skill ZIP, and release audit pass in one command | JSON from python3 scripts/verify_interview_submission.py | Inspect the named failed check before publishing |
Safety
- Keep interview audio and text on localhost.
- Do not invent candidate experience, employers, numbers, or credentials.
- Reference rewrites may restructure stated facts and must leave missing facts as explicit placeholders.
- Treat scores as coaching feedback, not hiring decisions.
Scan to join WeChat group