get-job

Internship.skill · get-job.skill

"You don't need a perfectly matching background; you need to translate the background you already have into the language your target role understands."

A general-purpose job-seeking skill. Cross-platform usable (Claude Code, Codex, or any environment that supports SKILL.md agent environments).

Reliability Protocol (run first)

See references/reliability-playbook.mdfor the complete method. Before starting any task, do an input checkup first; do not go directly into research or rewriting.

Status	Handling
✅ Complete	Target role, market/JD, resume, and application status/round are clear enough; go directly to the corresponding Stage.
⚠️ Workable	There are gaps but they are not blocking: infer automatically, fill in with WebSearch, mark low-confidence items, and continue producing.
🛑 Blocked	Hard inputs such as the target role or resume body are missing, so it cannot proceed safely; ask only the minimum necessary questions and do not start imagining details.

Automatic recovery priority:

If it can be extracted from user input, do not ask again. For example, infer the target market from the JD language/location, and infer the round from the interview invitation.
If it can be filled in by searching, search and fill it in first. For example, if the user only gives the company + role but does not paste the JD, first find the official recruiting page/same-role JD.
If it can be produced in degraded form, produce it in degraded form first. For example, when the round is unknown, generate 00-总览.md + 01-简历bullet逐条深挖.md, and list only subsequent rounds as to be confirmed.
If it cannot be verified, do not write it as fact. All low-confidence content goes into the "gaps/assumptions/to be confirmed" section.

Reliability goal: when input is incomplete, search sources are insufficient, file parsing fails, the round is unclear, or the user's claim has no evidence, first perform automatic checkup and degradation; do not push every issue to manual confirmation.

Quality Gate (must run before delivery)

See references/quality-gates.mdfor the complete method; see references/regression-evals.mdfor regression testing after major changes. Before any final delivery, run the quality gate internally first; do not by default write the self-check process into the user deliverable.

Level	Requirement
P0 Red Line	Fabrication, altering hard background-check information, presenting sources as facts, hard-coding rounds, missing bullet items, tag leakage, over-packaging: if any occurs, stop delivery.
P1 Core Quality	Role research, resume rewriting, and interview preparation must meet the corresponding completion standards; if they do not, fill them in, degrade them, or clearly mark the gaps.
P2 Experience Quality	File names, language, structure, and scannability should feel like a publishable product, with no work notes or template remnants.

Before delivery, internally check four coverage matrices:

JD core competencies → resume bullets: every role competency is supported by real experience.
Resume bullets → interview preparation: every final bullet has a narrative, follow-up questions, evidence, and fallback.
Claim → evidence: Every number/strong verb has an evidence level and an interview handling strategy.
Round → Source: Every independent round file has an invitation, interview experience, or process basis.

In an environment that can run scripts, use python3 scripts/quality_check.py <output-folder> to scan the structure and red-line words of public deliverables; a script passing does not mean semantic perfection, you still need to follow quality-gates.md to do a semantic self-check.

Chain (core)

Job hunting is not just making your resume look like that position. The complete chain has three segments, none of which can be missing:

① 调研岗位        →   ② 改简历          →   ③ 准备面试
   WebSearch            目标定位+迁移翻译       分轮次+逐bullet深挖+表达训练
   ↓产出                ↓产出                  ↓产出
   岗位调研.md          改后简历.docx          面试准备/

The output of each segment is the input of the next segment:First figure out what the position requires, then you know where to revise the resume; once the resume is set, you know which landmines to dismantle in the interview. If the direction is wrong, no matter how beautifully the resume is revised, it is useless.

examples/ Under it, each folder is one complete finished product—named after the position, containing position research, revised resume, and interview preparation folders.

Honesty bottom line (runs through the whole process, cannot be crossed)

Translation ≠ fabrication. Telling a true thing you have done from a different angle to the target position is translation; writing down something you have not done is fabrication, and it will collapse as soon as the interview probes it.
Skill transfer must have a real anchor. Every "transferable skill" corresponds to a thing you have really done; do not transfer a skill that does not exist at all.
User input is not a guarantee of fact. The Skill can only do evidence-based probing and risk labeling, and cannot prove on the user's behalf that he has not padded his credentials; any claim without evidence is uniformly marked as "pending user confirmation/low confidence".
Vagueness ≠ lying. For numbers whose source cannot be explained clearly, change them to "double-digit improvement" or "the calculation method is no longer remembered accurately"; this is honestly admitting you do not remember clearly.
Never touch the background-check red line. Hard information that will be verified, such as education, employment dates, and job titles, cannot be changed by a single word.
"Completed with AI assistance" is a plus. Using AI to quickly build an MVP and making product/architecture decisions yourself, if told honestly, is a hundred times better than pretending to be a senior engineer and being probed until you collapse.

Stage 0: Anchoring

Nail it down before starting work, ask everything at once, do not squeeze it out bit by bit:

Anchor	Function
Target position + department + direction	The entire chain revolves around it. "Product manager" in the "technology and product direction" and in "product operations" are two different sets of preparation.
Target country/region + language	Positions in China, the United States, the United Kingdom, Singapore, etc. come from different sources; if you do not confirm the market, it is easy to search in the wrong direction.
Existing resume + JD	Drop the resume in (PDF/Word); if there is a JD, provide it together.
Whether the resume can be revised	Not yet submitted → go to Stage 2 to revise the resume; already submitted → skip it and go directly to Stage 3.
Rounds + timing	The interview rounds the user has already received, invitation information, and how much time is left; when unknown, Stage 1 research infers the real process, and does not add rounds that do not exist.

If the information is incomplete, automatically recover according to the reliability protocol:

Missing JD: When there is a company + position, first use WebSearch on the official recruitment page/same-position JD; if it cannot be found, use a general position profile and mark it low confidence.
Missing market: Infer from the company, JD language, location, and recruitment platform; if it cannot be inferred, mark it as "market pending confirmation" and at the same time search using the most likely market.
Missing rounds: infer from invitations/interview experiences/company process; if it cannot be inferred, only do an overview and bullet-point deep dive.
Resume cannot be parsed: ask the user to paste the main text, but you can still first output a list of materials that need to be supplemented.

Stage 1: Research the role (WebSearch)

The true starting point of the whole process. Use WebSearch to get a clear picture of the target company/role and produce 岗位调研.md。

First determine the target market; if unsure, ask the user. Research must use local keywords for the target market and local recruiting/interview experience sources; do not search only in English or only in Chinese.

Source quality priority:

Grade A: first-hand/official: JD provided by the user, company recruiting pages, official website, product documentation, official news, team blogs.
Grade B: first-hand experience: interview experiences obtained by candidates themselves, referral information, first-hand experience with specific roles and dates on Nowcoder/Maimai/Xiaohongshu, etc.
Grade C: reliable second-hand: mainstream media, industry reports, public launch events, competitor updates.
Grade D: SEO/reposts/course advertorials: can only be used as clues and cannot support conclusions on their own; if only this type of source is found, it must be marked "low confidence/to be confirmed".

Market source coverage (choose according to target country/region; not all need to be searched, but gaps must be explained):

Market	JD / recruiting sources	Interview experience / reputation sources	Business / industry sources
Mainland China	Company official website, campus recruitment/social recruitment official website, BOSS Zhipin, Lagou, Liepin, 51job, Zhaopin	Nowcoder, Maimai, Xiaohongshu, KanZhun, offershow	Company official website, WeChat public account/press releases, 36Kr/LatePost/industry reports
United States	Company careers、Greenhouse、Lever、LinkedIn Jobs、Indeed	Glassdoor, Blind, Reddit, Levels.fyi, interview experiences provided by alumni/users	Company official website, product documentation, SEC/investor materials, TechCrunch/mainstream media
Other markets	Company official website + leading local job board + LinkedIn	Local forums/Glassdoor/alumni interview experiences/user-provided materials	Company official website, local media, industry reports

Search these dimensions:

Core competency keywords (3-5): infer from the JD and recruiting page what abilities this role truly tests → this is the bullseye the resume must align with
Business / product line / competitors: what this business does, who it serves, what it recently launched, how competitors are competing (second-round interviews often ask "do you understand us?")
Hidden thresholds / round process: how many rounds this company/role usually has, who interviews in each round, what each round emphasizes; if there is no evidence, mark "unknown/to be confirmed"; do not write it as fixed first round, second round, third round, HR interview
Industry / competitors / latest developments: industry judgment questions, "what do you think of our recent X"

⚠️ Capturelatestinformation; do not fabricate the company's recent situation based on training corpus; being outdated is a hard flaw. Do not be led astray by advertorials at the top of search results; key conclusions should be cross-verified with Grade A/B sources as much as possible; if they cannot be cross-verified, write them as assumptions. For multinational companies, distinguish global roles from local roles; roles with the same name may have different competency models in China/the United States.

The output must include a short "source coverage statement": target market, which local sources were used, how many A/B/C/D sources there are, which necessary sources were not found, and which conclusions are only leads or still to be confirmed. For the output format, see examples/*/岗位调研.md。

Stage 2: Revise the resume (do this when the resume has not been sent out yet)

For the complete method, see references/resume-playbook.md. Five steps:

Target positioning: use the competency keywords from Stage 1 as the bullseye.
Inventory real experience: for each experience, ask "what did I actually do + what evidence/details can explain it clearly + which part maps to the bullseye," and build a mapping of "real experience -> target competency -> evidence confidence."
Write a job intention header: add a target-oriented positioning sentence at the top of the resume so the interviewer can see in 5 seconds that "this person is coming for our role."
Competency transfer translation (the most core part): rewrite item by item = real event -> restate in the target role's language -> when necessary, naturally point out transferability. Transfer expressions must be restrained; prioritize weaving them into bullets, do not mechanically attach one sentence to every item, and do not output the label word "transfer sentence:".
Reorder + quantify + cut irrelevant content: put the most relevant items first; blur numbers whose source cannot be explained clearly; delete content unrelated to the role.

After finalizing, use scripts/generate_resume.py to generate a unified template docx (for usage, see scripts/README.md）。

✅ Checkpoint: have the user confirm the positioning, key transfer narrative, and high-risk claims - whether they are true, whether there is evidence, and whether there is exaggeration. Content the user cannot confirm should be downgraded, blurred, or deleted.

Stage 3: Prepare for interviews

For the complete method, see references/interview-playbook.md. Output the 面试准备/ folder. Six steps:

Build a question bank + discover the round map: based on the JD, product line, competitors, interview experience posts, company process, and user invitations, determine how many rounds actually need to be prepared; mark confidence and basis for each round.
Deep-dive each resume bullet: number every bullet in the final resume, and for each one prepare "how to explain this bullet, how the interviewer will follow up, what the evidence standard is, and how to fall back if you cannot answer." Stage 3 is not complete unless every bullet has been covered.
Project verbatim script + knowledge points tied to project deep-dive: for core projects, write a presentable script in the order "background -> what I did -> step-by-step how I did it -> results -> difficulties -> next optimization"; technical/fundamental/AI concepts must be tied back to the user's own project preparation, not piled into a question bank detached from experience.
Expression state + self-introduction training: prepare oral expression that is neither self-effacing nor arrogant, and natural and poised; the self-introduction output should be "spoken version + keyword card + practice actions," avoiding reading from a script, listing everything flatly, over-flattering, or being arrogant and perfunctory.
Evidence-based audit + risk map: for each claim, ask "can you explain clearly if asked about the source"; classify as ✅true / ⚠️cannot remember clearly / 🚨false; pair each item with a handling strategy and time budget; for questions you do not know, prepare honest boundaries rather than guessing or making things up.
Produce files by actual rounds + post-interview review question bank: must produce 00-总览.md、01-简历bullet逐条深挖.md、02-表达状态与自我介绍.md and 99-面后复盘题库.md; other files should only be generated for rounds that are "user-confirmed" or "high-confidence from research," for example 03-业务一面-简历深挖.md、04-产品二面-业务判断.md. If no HR interview can be found through search, do not generate an HR interview file.

✅ Checkpoint: Give the user for confirmation the bullet-by-bullet deep-dive table, expression status/self-introduction, round map, risk map, post-interview review question bank, and confirmed-round talking points. For bullets the user cannot explain clearly, if the resume has not been submitted yet, go back to Stage 2 to rewrite/delete them; if it has already been submitted, prepare honest downgrade and transition talking points.

Taste Guidelines (Quick Reference)

Principle	In One Sentence
Research first, then act	Revising your resume without figuring out what the role requires = blind revision
Translation > Piling Up	Translating real experience into the language of the role beats piling up keywords
Transfer needs an anchor	Every "transferable skill" corresponds to something you have actually done
Truth > Perfection	1-2 small things you have actually done beat 3 things fabricated seamlessly
Vagueness > Lying	"I can't recall the exact precision anymore" is always safer than forcing out a fake number that collapses under follow-up questions
Proactive > Passive	Proactively steer the pace to your strongest cards; don't wait for the interviewer to dig into where you feel insecure
Explain Clearly > Recite Scripts	An interview is not reading a script aloud; it is explaining real experience clearly while keeping your composure
Validate before delivering	Output that has not passed the quality gate does not count as complete

Things never to do

Fabricate experience or data that does not exist, or transfer a skill you simply do not have
Touch hard background-check information such as education/in-employment/job title
Teach the user to pretend to be a senior engineer and then collapse under follow-up questions
Revise the resume in a vacuum without researching the role

Creator Attribution

This Skill is provided by "Internship.skill / get-job.skill". Creator:agentenatalie