LLMInternSkill: Evidence-Bound Resume Polish for LLM Internship Candidates

What LLMInternSkill Does and Who It Is For

Resume polish tools typically improve phrasing without checking whether the improved phrasing can be defended in an interview. LLMInternSkill is designed to close that gap. Its core constraint is explicit: the tool polishes wording to make real experience clearer and more technical, but it will not help the candidate write claims that are not supported by evidence they can actually produce in an interview.

The README states this directly: "润色可以让真实经历更清楚、更有技术含量；但不能把没有证据的经历润色成事实." The English translation is: polishing can make real experience clearer and more technically detailed, but it cannot turn experience without evidence into fact.

The tool is specifically designed for the LLM internship market in China, targeting positions in RAG and knowledge-base Q&A, Agent and tool use, post-training and alignment, pretraining, search and recommendation ranking, LLM application engineering, LLM algorithms, AIGC, multimodal, and AI backend engineering. The README documents what evidence the tool checks for each track: for RAG, it checks for chunk design, retrieval, rerank, citation, evaluation, rejection, and permission handling. For Agent, it checks for tool schema, parameter validation, trace, retry, state management, and human review.

It is a Markdown repository that functions as a Codex Skill. The primary users are students and early-career engineers applying for LLM-related internships at Chinese technology companies.

The Evidence-First Philosophy and the Truth Boundary

The README includes a before/after table that shows how the tool transforms resume bullet points. The transformation pattern is consistent across examples: vague claims that imply more than the evidence supports are replaced with specific descriptions of what was actually done.

The before column shows: "熟悉大模型与 RAG，优化搜索效果" (familiar with LLMs and RAG, optimized search performance). The after column becomes: "围绕搜索相关性场景整理 query-doc 样例与长尾查询 bad case，分析歧义查询、时效性不足和低权威 DOC 对检索结果的影响" (organized query-doc examples and long-tail query bad cases around search relevance scenarios, analyzed the effects of ambiguous queries, insufficient timeliness, and low-authority DOC on retrieval results).

The key change is the removal of the outcome claim ("optimized performance") and its replacement with a specific description of the actual work done. The README's annotation explains: when there are no metrics, do not write "optimized effectiveness"; describe the real evidence as problem analysis ability instead.

The output the tool labels as the Truth Boundary classifies each resume claim into one of five verdicts: can write, write with caution, can write after adding evidence, cannot write, or cannot determine. This verdict structure is the mechanism by which the tool enforces the evidence constraint systematically rather than statement by statement.

Setting Up and Running LLMInternSkill

Three operating modes are documented in the README.

The simplest path is pasting a JD and resume directly into the Codex chat and invoking the skill by name:

使用 LLMInternSkill。
请帮我做简历润色，但不要编造经历。

The full materials folder mode organizes inputs into a directory structure:

materials/
├── target_jd.txt
├── resume.md
├── projects/
├── code/
├── notes/
├── papers/
├── awards/
└── other/

With this structure in place, the invocation tells the skill to process the full folder:

Use LLMInternSkill on ./materials.

To install as a Codex Skill permanently:

mkdir -p ~/.codex/skills
git clone https://github.com/couragec/llm-intern-skill.git ~/.codex/skills/llm-intern-skill

After restarting Codex or starting a new session, the skill is available by name.

JD Tailoring, the Evidence Contract, and Interview Drill

The JD tailoring module takes a target job description and the candidate's materials and produces three outputs: a match table comparing the JD's requirements against the candidate's evidence, a verdict (strong fit, borderline fit, or risky fit), and a customized resume section.

The README's flagship example uses a Douyin/ByteDance Seed search-ranking research internship JD. The output for that example includes: Verdict: risky fit; the explanation cites RAG and a small search demo as present but notes missing ranking metrics, strong algorithm internship evidence, and DOC understanding experiments; the fastest upgrade path suggests turning the mini search demo into a BM25 vs. embedding vs. rerank comparison with NDCG@10 and MRR evaluation plus bad-case analysis.

The Evidence Contract is a document the tool generates that lists what claims in the final resume depend on what evidence. If the evidence is lost, challenged, or misremembered in an interview, the contract identifies which resume lines become indefensible.

The Interview Drill module generates questions that an interviewer would ask about each line of the resume, organized into three categories: dangerous answers (likely to expose a gap), adequate answers, and strong answers. The README calls this "逐行拷打" (line-by-line grilling). It also generates multi-step follow-up chains for high-risk claims.

Project Scout: Building Evidence When Materials Are Thin

When a candidate's existing materials are too weak to support the target JD, LLMInternSkill recommends open-source projects to study, reproduce, and extend to create new evidence rather than inflating existing claims.

The README lists the categories of projects the Scout module recommends: MiniMind-style from-scratch LLM training, RAG evaluation and citation accuracy, search and rerank baselines, Agent tool-calling workflows, and LLM inference, serving, and quantization projects.

For each recommendation, the Scout output must include: why this project fits the target role, why it might not or what the risk is, the minimum run path to get it working, what to modify to add original contribution, what evidence to collect while working on it, what resume claim can be written safely after completing it, and what interview questions will be asked about it.

The README includes an example for MiniMind (examples/project-scout-minimind.md) and one for search and rerank (examples/project-scout-search-rerank.md). These serve as concrete illustrations of what the Scout output looks like for a specific project and a specific target role.

A boundary condition is stated clearly in the README: learning an open-source project cannot be packaged directly as work experience. Only after actual reproduction, understanding, modification, and evidence collection can it be written as a personal project or open-source contribution.

Output Structure and LaTeX Export

In full materials mode, LLMInternSkill generates a set of output files:

output/
├── 01_jd_analysis.md
├── 02_materials_audit.md
├── 03_truth_boundary.md
├── 04_evidence_contract.md
├── 05_resume_polish.md
├── 06_targeted_resume.md
├── 07_interview_grilling.md
├── 08_answer_cards.md
├── 09_upgrade_plan.md
├── 10_project_scout.md
└── 11_final_pack.md

The numbered sequence reflects the intended reading order. The JD analysis and materials audit come first to establish context; the truth boundary and evidence contract come before the polished resume to enforce the evidence constraint before writing begins.

The LaTeX export module uses Bill Ryan's elegant LaTeX resume template in a Chinese version. The entry file is resume-zh_CN.tex and compilation requires XeLaTeX. The template source is credited to the Overleaf template library and the upstream GitHub repository at github.com/billryan/resume.

The README notes that large CJK fonts are kept locally but are not committed to the Git repository, which means they must be present on the machine used for compilation.

Maintenance, Licence, and Scope Limits

LLMInternSkill is licensed under MIT, which allows free use, modification, and distribution without copyleft requirements. The repository was last pushed on 2026-08-04, about eight weeks before this review. It is not archived.

The primary language of the repository is Markdown. The README is in Simplified Chinese by default, with README_EN.md available for the English version. The skill file (SKILL.md) defines the Codex Skill interface. The repository also includes agents, evals, and references directories suggesting ongoing development of the skill's capabilities.

The tool's scope is narrow by design. It targets LLM internship roles at Chinese technology companies. The job categories listed in the README (RAG, Agent, post-training, pretraining, search/ranking, LLM application engineering, AIGC, multimodal, AI backend) reflect the hiring landscape at companies like ByteDance, Baidu, Alibaba, and similar firms. A candidate applying to a software engineering internship outside AI would find the role-specific evidence checks inapplicable.

Comparable tools in the space are generic resume enhancers (Resumify is cited in the README's references) that improve phrasing without evidence auditing. LLMInternSkill's differentiation is the evidence guard mechanism, which makes it specific but also more honest about what it can and cannot do.

Editorial conclusion

LLMInternSkill is the right tool for LLM internship candidates who want their application materials to hold up under technical interview scrutiny rather than just read well at first glance. Its value is specific to the LLM and AI internship hiring context: the role-specific evidence checks (for RAG, Agent, post-training, search ranking, and similar tracks) require background knowledge that a generic resume tool does not have. Candidates applying to non-LLM roles will find little use for it. Before using it, read the Evidence Guard section of the README and the examples to understand the truth boundary concept, since the tool's value comes from working within that constraint rather than around it.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. Project website
4. README
5. wanyichen06/LLMInternSkill on GitHub