Interview Scripts That Sound Human, Not Generated

Where this fits

A candidate is staring at a blank page before a product manager round. They have a story-bank.md full of project details, but every draft they write sounds like a LinkedIn summary. They need a script they can actually read aloud — with numbers that hold up under follow-up questions. The interview-script skill turns raw experience cards into a spoken-style 逐字稿, organized by modules and tagged with who asks each question.

Why agents benefit

• Writes for the ear, not the eye. The output is verbatim speakable language — contractions, pauses, natural rhythm — so the agent isn't producing a bullet list disguised as prose.
• Enforces a hard source rule. Every number in the script must trace back to the story bank or be explicitly flagged as "needs user input." The agent refuses to invent metrics, which kills the classic AI hallucination in interview prep.
• Applies anti-AI-tone constraints. A built-in style ban list (from references/style-zh.md) strips out clichés, over-polished transitions, and generic filler that screams "generated."
• Adapts to the user's voice. If the user has past interview transcripts or a profile style preference, the agent mimics their pacing and word choice — the script sounds like them, not like a template.
• Structures for quick recall. Each answer is modular with anchors, so the candidate can jump to "tell me about a conflict" or "why this company" during the actual interview without flipping pages.

Practical scenarios

First-round self-introduction under time pressure

Instead of a generic "tell me about yourself" outline, the agent reads the story bank and builds a 90-second spoken intro that leads with the most relevant project for the target role, cites one concrete number, and ends with a transition to the interviewer's question.

Behavioral question with a weak story

The user asks "这题怎么答" for a leadership question but the story bank lacks a matching experience. The agent doesn't fabricate — it flags the gap and suggests either a qualitative answer or a prompt to add a new story card first.

Pre-round tuning with intel

The agent reads jobs/<slug>/intel.md to see which matching-table items are flagged ⚠️/❌, then prioritizes those topics in the script. It also adapts tone based on the interviewer profile — more direct with a data-driven VP, more narrative with a startup founder.

Add it to your agent workflow

Install the skill in your agent environment:

npx skills add YunyueLi/greenroom --skill interview-script

Then trigger it by asking for a script directly:

帮我写自我介绍逐字稿，岗位是资深产品经理，明天二面。

The agent will check prerequisites (story bank, intel, style prefs), ask for missing numbers, and produce jobs/<slug>/script.md — a ready-to-read script with source tags on every figure.

Compare before adopting

Before you settle on this skill, evaluate against three things: quality signals — does the output pass the "read-aloud test" and avoid AI clichés? Maintenance freshness — is the story bank being updated, and does the skill handle new experiences gracefully? Alternatives — a generic prompt can produce bullet points, but it won't enforce traceability or voice matching. Also check workflow fit: this skill sits after story-bank and before mock-interview, so it assumes a pipeline, not a one-shot.

Why it is worth tracking

This is a niche but well-scoped skill — 11 stars, AGPL-3.0, actively shaped by a single author who clearly understands the interview-prep problem. The real signal is the design: it refuses to hallucinate numbers, forces spoken style, and modularizes answers for real-time use. That's the kind of constraint-driven agent behavior worth watching. If you build interview-prep agents, evaluate it now; if you're just curious, track it as a reference for how to encode domain rules into a skill. See the full skill page at OpenAgentSkill for updates and community feedback.