Turn "a JD I like + my resume" into an evidence-backed, actionable LinkedIn job shortlist.
This skill is step 0 of the Offer Toolkit:
种子 JD + 简历 → 搜索画像 → LinkedIn 候选池 → 匹配排序 → shortlist
↓
Job Description Skill 深度解码
Boundaries
- Only perform job search, read public job information, deduplicate, match-rank, and recommend.
- Never automatically click Apply, never fill out application forms, never send messages, and never apply on the user's behalf.
- Never request or handle LinkedIn passwords, verification codes, cookies, or session tokens.
- If encountering a login wall, CAPTCHA, rate limit, or robots restriction, stop that path and do not attempt to bypass it; instead use public search results or output clickable LinkedIn query links.
- Whether a job is still open can only be judged based on the page at the time of search; the report must state the search time.
Input process
Step 1 · Collect seed JD
First ask for a JD link or full text that the user is genuinely interested in. It is not the only target, but is used to extract role semantics: title, level, scope, domain, core competencies, and exclusions.
If the user has no seed JD, accept 2-3 target titles + a one-sentence direction description as degraded input, and mark "no seed JD, lower confidence in search profile".
Step 2 · Collect resume
Accept PDF, Word, plain text, or an already structured resume. Extract only from facts provided by the user:
- Current/most recent title and approximate level
- Years of experience and core scope of the most recent 3 years
- 3-6 competencies with direct evidence
- Industry, product stage, customer type, and team type
- Explicit constraints such as location, language, and visa
Do not write the seed JD's requirements back into the user profile unless there is direct evidence in the resume.
Step 3 · Fill in hard filters
Only ask about missing conditions that would significantly change the result set, and ask only one at a time:
- Work location, and whether remote / hybrid / relocation is acceptable
- Target level or acceptable upward/downward range
- Time range: 24 hours / 7 days / 30 days
- Companies, industries, contract types, or visa conditions that must be excluded
Defaults: most recent 7 days, full-time, target level with one level of upward/downward range. Defaults must be explicitly listed in the final report and must not be silently assumed.
Search execution
Step 4 · Generate search profile
First form this structure in memory:
target_titles: [] # 2-5 个,不堆同义词
level: ""
locations: []
workplace: [] # remote / hybrid / on-site
date_posted: "7d"
employment_types: []
core_capabilities: [] # 3-6 个,必须有简历证据
domains: [] # 0-3 个
company_preferences: []
exclusions: []
seed_signals: [] # 从种子 JD 提取,但不等同于简历能力
Echo a search profile of no more than 8 lines for the user to quickly correct. If the user does not correct it, continue and do not require a second confirmation.
Step 5 · Construct query matrix
Generate 6-10 complementary query sets, rather than cramming all terms into one query:
- Exact title: target title + location
- Adjacent title: adjacent title + location
- Capability-led: title + 1 core competency
- Domain-led: title + target domain
- Scope-led: title + scope signals such as 0→1 / platform / enterprise / growth
Each query set should contain only 1-2 discriminating terms. Overly long queries will miss jobs that use different wording but actually match.
LinkedIn query URLs use public Jobs Search parameters; prioritize setting keywords, location, date posted, workplace, and employment type. Record the query and URL actually executed for each group.
Step 6 · Collect candidate pool
Use a browser to open LinkedIn Jobs Search. Read the visible result cards group by group and collect:
- title
- company
- location / workplace
- canonical job URL or job id
- posted age
- salary (record only if present on the page)
- card summary or matched keywords
Deduplicate by canonical URL/job id; when there is no job id, deduplicate by company + normalized title + location deduplication.
Stop condition: whichever is met first:
- 30 unique candidates that pass the hard criteria already exist
- View the first 2 pages or first 25 items for each group
- The page requires login, shows a CAPTCHA, or restricts access
Do not keep results whose location, level, or job type clearly do not match just to reach 30.
If LinkedIn is inaccessible:
- Use public Web search, limited to
site:linkedin.com/jobs/view, and execute the same query matrix; or - Output clickable LinkedIn URLs for the query matrix and let the user open them;
- Clearly mark that real-time collection was not completed this time; do not fabricate jobs.
Step 7 · Two-stage scoring
Quickly screen all candidates:
- Hard criteria pass/fail: location, work mode, employment type, explicit visa threshold
- Title/level fit:0-30
- Core competency keyword matches: 0-30
- Domain/scope similarity: 0-20
- Freshness: 0-10
- User preference: 0-10
Only bring the top 10 from quick screening into deep scoring. For these 10, open the job page one by one and read the visible full JD; jobs whose full JD cannot be read may be kept, but must be marked Description unavailable, and can at most enter Tier B.
Top 10 for deep scoring: Read frameworks/ranking-rubric.md, and compare Must Have, Nice to Have, and Hidden Signals item by item using resume evidence. Give ranges, not false-precision single points.
Output
By default, output one Markdown shortlist; unless the user explicitly wants HTML, do not create a large report for a single search.
# LinkedIn Job Shortlist — YYYY-MM-DD
## Search profile
[目标、地点、时间范围、默认假设]
## Best bets
| Tier | Match | Role | Company | Location | Posted | Why it fits | Main risk | Link |
## Worth a look
[Tier B 表格]
## Skipped patterns
[被批量排除的模式和原因,不逐条堆岗位]
## Search log
[执行过的 queries、采集时间、LinkedIn/公开搜索、访问限制]
Tiering rules:
- Tier A · Apply first: deep match range lower bound ≥ 70%, with no missed threshold-type Must Have
- Tier B · Review: lower bound 55-69%, or incomplete JD but strong quick score
- Tier C · Skip: lower bound < 55%, or hard criteria failed; do not enter the main table, only summarize exclusion reasons
Each Tier A job must provide:
- 2 pieces of direct matching evidence from the resume
- 1 risk a hiring manager may worry about
- 1 next step: apply directly / seek a referral first / build portfolio first / hand to JD Skill for deep decoding
Finally recommend 5-10, and do not dump all 30 search results to the user as-is. Links must point to specific jobs, not just the search page.
Connection with the Toolkit
- After the user selects a job, hand that JD + original resume to
job-description-skillto generate a complete Offer Strategy Report. - After the user decides to apply, hand the target JD + original resume to
resume-skillto create a targeted resume. - Do not copy a complete JD report or rewrite the entire resume within this skill; job discovery, job decision-making, and resume creation are three different steps.