---
name: sherlock-company
description: Use when a job seeker wants to investigate a company, role, JD, Offer, or HR claims before deciding whether to continue interviewing or accept an offer, especially when evidence is incomplete or employee experience is hard to obtain.
---

> 📦 **套件索引说明**：本 Skill 已收录于 [`ouxxyy/ouba-classmate`](https://github.com/ouxxyy/ouba-classmate) 套件。完整执行脚本、`references/` 与示例数据位于子仓库 [`ouxxyy/sherlock-company`](https://github.com/ouxxyy/sherlock-company)。

# Sherlock Company

Help the user answer three questions: what is verified, what remains unknown, and what must be checked before accepting the Offer. The deliverable is a decision aid, not a corporate verdict.

## Start from the decision

Collect or infer only what the user supplied: company/brand, role, city or team, decision deadline, concerns, and available JD/Offer/HR materials. Search public evidence by default. Ask the user only when company identity remains ambiguous or a missing preference would materially change the decision.

Anchor the legal employer before investigating. Treat brand, parent company, recruiting entity, contract entity, payroll entity, social-insurance entity, city, and team as separate fields. A famous brand never proves the user's contracting entity.

If identity remains unresolved, still produce a clearly labeled **stage report** containing candidate entities, missing identifiers, generic material checks, and the next question. Do not attach company-specific findings to an unconfirmed candidate. Keep the exact research identity separate from the display name when the user asks for an anonymized report.

## Build one evidence record

Read [references/evidence-contract.md](references/evidence-contract.md) before research or analysis. Keep every claim in one of four states: `verified`, `signal`, `conflict`, or `unknown`. Record access failure separately; zero results never mean zero risk.

Use primary sources first, then reliable reporting, then user-supplied experience material. A company's own site can prove what it promises, not that it consistently delivers. Search snippets, AI summaries, scraped aggregations, and anonymous posts are leads until their original context is checked.

When legal or employment rules matter, verify current authoritative sources and keep them separate from company-specific facts. Do not turn a legal minimum into proof of actual company practice.

## Search for decision-relevant experience

Before online research, read [references/search-playbook.md](references/search-playbook.md). Search by concern, grounded aliases, city/role/team and time; reserve capacity for original posts and relevant comments. Separate firsthand accounts, interview reports, recruitment promotion and recycled summaries. Record actual access attempts and deduplicate shared origins.

When authenticated reviews are requested or needed and a browser is available, read [references/authenticated-reviews.md](references/authenticated-reviews.md). Use the user’s existing authorized session through Tabbit when requested; login increases visibility, not truthfulness. Do not mark unattempted channels blocked or treat an accessible anonymous post as a verified workplace fact.

## Check the three default warnings

For every company investigation, read [references/company-warnings.md](references/company-warnings.md) and check these alongside the user's concerns, even when not explicitly requested:

- **规模与初创预警**：the company has fewer than 100 people **or** is currently a startup. Either condition triggers a warning independently.
- **感恩与业绩标语预警**：the workplace or company materials show many or repeatedly used gratitude/loyalty or performance-pressure slogans, such as “感恩公司”“冲刺××业绩”.
- **居民楼办公预警**：the actual workplace is in a residential building. If only a residential registered address is known, issue an address warning pending workplace verification.

Warn when supported; label leads or unresolved contradictions **待核实预警**. With no usable evidence, explicitly mark the check unknown and give the next check. These are prompts to investigate scale, stability, management and workplace conditions, not proof of misconduct or an automatic rejection. Preserve evidence status, entity, date, scope and counterchecks; follow the reference for threshold, proxy and image-context boundaries.

## Choose tools narrowly

Use ordinary web research for public sources, then browser reading for relevant dynamic or authenticated pages when available. If an official QCC MCP or CLI is available, read [references/qcc.md](references/qcc.md) and use it only for exact entity, corporate, operating, public-risk, or legal-record facts. QCC is optional and never a substitute for employee experience.

If the user supplies JD, Offer, contract drafts, HR chat, or review screenshots, read [references/user-materials.md](references/user-materials.md). Ask for redaction of phone numbers, IDs, addresses, bank details, signatures, and unrelated names before analysis.

If employee experience cannot be obtained, read [references/employee-experience.md](references/employee-experience.md). Do not manufacture an experience score, completeness percentage, or sample-size threshold.

## Decide without fake precision

Default to a plain-language state:

- `可以继续聊`：no confirmed deal-breaker, but material unknowns remain.
- `核实后再接受`：contract, compensation, probation, work pattern, or team evidence is still decision-critical.
- `已足够决定`：the user's concerns have current, relevant evidence and the remaining gaps are not material.
- `先不接受`：a verified fact crosses a boundary the user actually stated.

Never derive a star rating or numeric confidence/completeness score unless the user explicitly requests a defined scoring model and the record supports it. Unknown is not negative evidence. The default warnings above can surface supported facts or labeled leads; they do not establish misconduct. A reason to reject requires the decision evidence below, not merely missing web coverage.

Move from `核实后再接受` to `先不接受` only when a verified fact crosses the user's stated boundary, written terms conflict on a decision-critical condition and the conflict is not resolved, or the responsible party refuses to confirm a decision-critical item before the deadline. Move to `已足够决定` only when each stated concern is answered with relevant evidence and remaining gaps would not change the user's choice.

If the user explicitly requests scoring, define dimensions, weights, hard boundaries, and missing-value treatment before calculating. Keep unknowns unknown, show the unweighted facts alongside the score, and include sensitivity to any subjective weight. Do not silently reuse a company-level score as an Offer-level score.

## Deliver the report

Lead with one sentence the user can act on. Then show:

1. the exact company/role/scope checked;
2. verified facts, leads/conflicts, and unknowns in separate groups;
3. the three default warning checks followed by one card per additional user concern: known, unknown, impact, and next check; merge overlapping concerns without dropping any check;
4. an Offer-stage checklist with copyable questions and desired evidence;
5. employee-experience coverage and the next feasible evidence route;
6. source ledger, dates, scope, limits, and blocked or unattempted channels with actual reasons.

For HTML, read [references/html-report.md](references/html-report.md), fill the case schema in [references/case-schema.md](references/case-schema.md), then run `scripts/render_report.py`. The renderer loads `assets/report-v2.css` by default; standalone sample pages are visual references, not executable templates. Keep Markdown and HTML derived from the same record. Use the detective IP as a visual guide, not as a source of theatrical claims.

Mention triggered warnings, including pending ones, in the opening conclusion and put their detail in the first `concerns` cards and matching `checklist` items. Keep all three checks visible even when unknown. Use existing evidence states; “预警” is a presentation label, not a fifth status. Do not upgrade `decision.state` solely because a warning exists.

Write visible interface text and case conclusions in natural Chinese. Prefer “职位说明”“录用通知”“招聘人员” over unexplained `JD` / `Offer` / `HR`. Keep English only when it is part of a legal entity name, source title, URL, product name, or a necessary term that is explained on first use. Do not add decorative English labels.

Stop when the identity is resolved or explicitly blocked, each concern and default warning check has a result or named gap, decisive claims have sources and counterchecks, and the next actions are specific enough to use before the deadline.