LLMInternSkill: A Codex Skill That Grades Resume Claims Before an Interviewer Does

The Problem Is Not Wording, It Is Unbacked Claims

Most resume tools operate on sentences. LLMInternSkill operates on claims and their evidence. The README states the one-line principle directly: polishing can make a real experience clearer and more technical, but it cannot turn an experience without evidence into a fact. That distinction drives everything else in the repository. The target user is a student applying to large-model internships in China or abroad, working in Chinese or English, who has some combination of a RAG demo, a LoRA experiment, a course project, and a paper or two, and who is worried about being questioned into a corner during a technical interview. The README's own framing is that the toolkit exists so that every line of the resume can survive follow-up questioning, not so that the resume sounds impressive. The supported directions table lists eleven tracks, including RAG and knowledge-base QA, Agent and tool use, agentic RL, post-training and alignment, pretraining and mid-training, LLM application engineering, LLM algorithms, search and recommendation and ranking, AIGC, multimodal, and AI backend. For each track the README names what gets checked. RAG is checked on chunking, retrieval, rerank, citation, evaluation, refusal, and permissions. Search and ranking is checked on query-doc pairs, recall, coarse ranking, rerank, NDCG and MRR, and document quality. That list is the real specification of the tool: it is a checklist of the things an interviewer would probe, applied line by line to your resume.

Five Verdicts and an Evidence Contract

The core judgement in LLMInternSkill is a five-way classification: can write, write with caution, write after adding evidence, cannot write, and cannot determine. Every resume line is meant to land in one of those buckets. The pipeline in the README runs from a raw resume plus a materials folder plus target_jd.txt, through resume polish, JD matching, a truth boundary, an Evidence Contract, a tailored resume, interview grilling, answer cards, an upgrade plan, Project Scout, and finally a LaTeX draft. The Evidence Contract is the piece that makes the rest coherent. It is the artifact that records which claims are permitted by the supplied materials, so the later stages have something to check against rather than regenerating plausible-sounding prose each time. The full mode writes eleven numbered files into output/, from 01_jd_analysis.md through 11_final_pack.md. The README's before-and-after table shows the mechanism in practice. A line reading that the candidate is familiar with large models and RAG and optimized search results becomes a description of organizing query-doc samples and long-tail bad cases, analyzing ambiguous queries, freshness problems, and low-authority documents. The stated reason is that without metrics you do not write that you optimized results; you write the problem analysis you actually did. The same pattern applies to a claim of building an enterprise-grade knowledge base QA system, which is downgraded to a demo with document splitting, vector retrieval, and prompt template configuration, plus recorded retrieval errors, citation misalignment, and unsupported answers. The rule the README gives is that without deployment, permissions, monitoring, and user records, you do not write enterprise-grade system.

Installing It as a Codex Skill

The repository is Markdown, so there is no build step and no runtime dependency to install. Method three in the README is the installation path: create the skills directory, clone the repository into it, and start a new Codex session. The commands are mkdir -p ~/.codex/skills followed by git clone https://github.com/couragec/llm-intern-skill.git ~/.codex/skills/llm-intern-skill. After restarting Codex or opening a new session, the invocation is a plain-language request to use LLMInternSkill and read the materials folder. The second method assumes a directory layout rather than an install: materials/ containing target_jd.txt, resume.md, and the subfolders projects/, code/, notes/, papers/, awards/, and other/. The prompt in that mode asks for resume polish, a JD fit verdict, a targeted resume, interview grilling, answer cards, an evidence upgrade plan, and project scout recommendations in one pass. The first method is the lightest: paste a job description and a resume into the chat, ask for polish without fabrication, and request the before-and-after, the list of writable and downgradable lines, a technical-expression version, a JD-tailored version, and likely interview follow-ups. Note the naming inconsistency, because it will trip you up: the repository is wanyichen06/LLMInternSkill, but the clone URL in the README points at couragec/llm-intern-skill, and the badge in the header links to a repository under a third account, couragec/llm-intern-skill. If you are scripting the install, use the clone URL exactly as written; if you are citing the project, use the repository you actually cloned from.

The Doubao Seed Example Is the Honest Part of the Repo

The flagship example is a deliberately hard case: a Doubao large-model team Seed search, ranking, and document-understanding research internship, paired with candidate materials the README describes as weak to medium. The stated purpose is to show that the toolkit does not force a strong-fit verdict when the evidence does not support one. The output verdict is risky fit, with the reason given that the candidate has RAG and a small search demo but lacks ranking metrics, strong algorithm internship evidence, and document-understanding experiments. The recommended fastest upgrade is to turn the mini search demo into a BM25 versus embedding versus rerank comparison with NDCG@10 and MRR plus bad-case analysis. The example ships as four files: examples/doubao-seed-search-ranking-jd.md, examples/doubao-seed-materials-input.md, examples/doubao-seed-final-pack.md, and a simulated LaTeX resume PDF at examples/seed-resume-latex/seed-topseed-resume.pdf. This is the section of the documentation worth reading before you trust the tool, because a verdict of risky fit on a strong job description is the behaviour that distinguishes an evidence-bound system from a rewriter. It also sets the expectation correctly: the output is a conservative, submittable resume, not a competitive one, and the gap is filled by a one-day, three-day, or one-week evidence plan rather than by stronger adjectives.

Project Scout Recommends Work Instead of Inventing It

When the materials are too thin to write anything defensible, LLMInternSkill switches modes and recommends projects to build. The README names five categories: a MiniMind-style from-scratch LLM project, a RAG evaluation and citation-accuracy project, a search and rerank baseline project, an agent tool-calling workflow project, and an LLM inference, serving, or quantization project. Each recommendation is required to carry eight fields: why it fits, why not and what the risk is, the minimum run path, what to modify, what evidence to collect, the resume-safe claim available after completion, and the interview grilling questions that follow. The boundary is stated plainly: studying an open source project cannot be packaged as work experience. Only after real reproduction, understanding, modification, and evidence collection does it become a personal project or open source practice. That constraint is what keeps the module from becoming a suggestion engine for fabrication. It also means Project Scout has a cost the README does not quantify. A minimum run path for a from-scratch LLM project is not an afternoon, and the toolkit is honest that the alternative to a weak resume line is days of work, not a better sentence. There is a MiniMind example at examples/project-scout-minimind.md and a search and rerank example at examples/project-scout-search-rerank.md if you want to see the eight-field format filled in.

Where the Toolkit Stops Being Useful

The limitations follow from the design. First, output quality is bounded by input materials. A resume passed in as a chat message with no folders, no code, and no logs gives the Evidence Guard nothing to verify against, so most lines will land in cannot determine or write after adding evidence, which is accurate but not helpful if you needed a resume today. Second, the toolkit is built for a narrow applicant. The supported directions are all LLM-adjacent, and the checks assume you can talk about chunking, rerank, NDCG, SFT, DPO, tokenizers, or tool schemas. A frontend or mobile candidate would get a checklist aimed at someone else's interview. Third, the output format is an eleven-file directory. That is thorough for one carefully prepared application and slow for the twenty-application volume strategy many students actually run; the README's first method exists precisely because the full pipeline is too heavy for casual use. Fourth, the repository is Markdown only, so there is no programmatic validation of the Evidence Contract. Nothing stops a model from drifting past the contract on a later turn, and the guard depends on the session honouring the earlier artifact. Fifth, the README describes the status badge as final MVP, and no releases were retrieved for this repository, so there is no versioned artifact to pin and no changelog to check for behaviour changes in the prompt logic. Finally, the LaTeX export targets one template, the Chinese variant of Bill Ryan's elegant resume under templates/resume-latex/bill-ryan-elegant-zh_CN/. If your field expects a different format, you are editing the template yourself.

Against a General Resume Rewriter

The obvious alternative is a general-purpose resume rewriter or a career-focused prompt you keep in your own notes. The difference is not tone, it is the presence of a refusal. A general rewriter asked to improve the line about optimizing search results will produce a stronger sentence, because that is the only task it has been given. LLMInternSkill is instructed to ask whether the sentence has evidence, how an interviewer would follow up, and what evidence would make it writable, in that order. The README's own contrast is that an ordinary polisher turns a plain sentence into a pretty one, while this toolkit takes three additional steps. In practice that means the general rewriter is faster and produces a better-looking document, and LLMInternSkill produces a worse-looking document that you can defend. For a candidate with genuinely strong evidence, the general tool is probably sufficient and cheaper in time. For a candidate whose resume is a mix of real work and inflated framing, the classification step is the whole value, because the inflated lines are exactly the ones that collapse under questioning. The other real alternative is to skip tooling and have a senior engineer read your resume line by line. That is higher quality per line than either approach, and it does not scale past a few readers, which is the problem a Skill like this is trying to solve.

Licence, Maintenance, and What to Check Before Adopting

The repository is MIT licensed, which permits commercial and private use, modification, and redistribution provided the copyright notice and licence text are retained. The LaTeX template it ships under templates/resume-latex/ is a separate artifact derived from Bill Ryan's elegant resume, and the README attributes it to an Overleaf source; the licence of that template is not stated in the material available here, so if you plan to redistribute a modified version of the template itself rather than just generate a PDF from it, verify the upstream template's terms independently. Nothing here is legal advice. On maintenance: the last push recorded is 2026-08-04, the repository is not archived, and no releases were retrieved, so there is no tagged version to depend on. Because the project is Markdown and prompt text rather than code, an upgrade means re-reading the changed instructions and re-checking that your saved prompts still match the current output filenames in output/. Nothing will break at runtime if you skip an update, but the verdicts can shift silently, which is worse. The concrete thing to verify before you commit to this workflow is the calibration of the evidence guard on your own material: run the full pipeline once against a materials folder you know well, read 03_truth_boundary.md and 04_evidence_contract.md, and check whether the lines it marked writable are lines you would defend in an interview. If it lets through a claim you would not defend, the guard is not doing its job for your case, and the fix is to add the missing artifacts to the materials folder rather than to argue with the output.

Editorial conclusion

Adopt LLMInternSkill if you are applying for RAG, Agent, post-training, or search-ranking internships and you already have a materials folder of code, notes, and logs to feed it. Skip it if you need a keyword-stuffed one-page resume for volume applications, or if you have no artifacts at all, since the toolkit will mostly tell you to go build something first. Before relying on it, open examples/doubao-seed-final-pack.md and check whether its downgrade of a mid-strength candidate matches your own judgement of that candidate; that single file tells you more about the tool's calibration than any feature list.

Official sources

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