Backend Agent Resume Scout: a Codex Skill that verifies resume projects against source code

Star counts make bad shortlists

The README opens on a precise observation: most job seekers are not short of open source projects, they just cannot tell which ones belong on a resume. The failure modes it lists are familiar to anyone who has screened juniors. Picking by star count lands on overheated, homogeneous repos that half the applicant pool also picked. Demos, tutorial replicas, shallow CRUD and plugin shells get dressed up as engineering work. Resume bullets collapse the moment an interviewer asks about the source, the request chain, the state transitions or the error handling. And among AI Agent projects, few have real business data, state flow or user value at all.

Backend Agent Resume Scout reframes project-picking as a verifiable decision flow: search a candidate pool, filter for business value, confirm a shortlist with the user, then verify the finalists against pulled source code before writing a single resume line. It is aimed at students hunting backend internships and campus offers, candidates upgrading from plain CRUD work, and people who want an Agent project that survives follow-up questions, across Java, Python, Go and Node.js backends.

A note on names: the repository calls itself 牛肉项目雷达 (Beef Project Radar), the skill inside is named backend-agent-project-selector, and the English summary uses Backend Agent Resume Scout. They are the same thing.

Five modes, and it refuses to guess

The skill demands an explicit recommendation mode before doing anything. agent-only recommends complete business-grade AI Agent projects. backend-only recommends traditional software backend projects and excludes IoT and hardware-integration work by default. mixed recommends one Agent project plus one traditional backend project. safe-mode favours projects that are easy to run, with few dependencies and low deployment cost, and challenge-mode favours harder, more differentiated projects suited to deep modification.

If the user does not name a mode, the skill asks first rather than starting a search. That single rule does more for output quality than any scoring tweak, because the mode determines the candidate pool before any filtering happens. A "find me a project" prompt and a "prepare me for a hard interview" prompt should not return the same list, and here they structurally cannot.

From search pool to source-verified pack

The workflow the README documents runs in a straight line:

用户画像与推荐模式
        ↓
联网搜索 GitHub / Web
        ↓
构建多样化候选池
        ↓
README probe 与项目类型初筛
        ↓
输出 3-4 个短名单项目
        ↓
用户确认方向
        ↓
拉取 GitHub 仓库到本地
        ↓
基于源码提取证据点
        ↓
生成 Markdown / PDF 简历项目包

In stages: the user's profile and chosen mode go in; GitHub and web search build a diverse candidate pool; a README probe does type pre-filtering; a shortlist of three to four projects comes out; the user confirms a direction; the chosen repositories are pulled to local disk; evidence points are extracted from source; and the Markdown and PDF resume pack is generated.

The candidate pool is deliberately bucketed across e-commerce transactions, CRM and ERP, collaboration and office work, ticketing and support desks, knowledge bases, and Agent workflows, covering medium-heat and niche projects rather than chasing whatever is trending. Shallow projects are dropped by default: demos, thin wrappers, pure frameworks, browser extensions, simple chatbots and shallow CRUD.

The load-bearing rule sits at the end of the workflow: README and web search are pre-filtering only, and the final "responsible features" and "technical difficulty" claims must come from local source verification. That is the difference between a prompt that recommends repositories and a pipeline that can defend its own output.

Installing into Codex and invoking it

The skill's name is backend-agent-project-selector, and installation is a directory copy into Codex's skills folder. On macOS or Linux:

mkdir -p ~/.codex/skills
cp -R backend-agent-project-selector ~/.codex/skills/

On Windows PowerShell:

New-Item -ItemType Directory -Force "$env:USERPROFILE\.codex\skills" | Out-Null
Copy-Item -Recurse -Force .\backend-agent-project-selector "$env:USERPROFILE\.codex\skills\"

Invocation is a prompt that names the skill and a mode. The README's full example, quoted verbatim (it is in Chinese, like most of the README):

Use $backend-agent-project-selector
推荐模式：mixed
技术栈：Java / Python
目标岗位：后端实习
时间预算：2 周
背景水平：普通本科，无实习
项目偏好：一个新奇 Agent 项目，一个有业务闭环的扎实后端项目
避开方向：浏览器插件、简单聊天机器人、IoT、硬件接入、太难部署
输出需求：推荐 + 简历写法 + 面试题 + 改造计划

Translated, it sets mode mixed, stack Java and Python, target role a backend internship, a two-week time budget, background a standard undergraduate with no internships, a preference for one novel Agent project plus one solid business-loop backend project, avoidance of browser extensions, simple chatbots, IoT, hardware integration and hard-to-deploy picks, and asks for recommendations, resume wording, interview questions and a modification plan. Nothing in that contract is exotic; the discipline is that the skill will not start without the mode line.

Output lands in the current workspace as four files: backend-agent-project-shortlist.md for the confirmation stage, backend-agent-project-resume-pack.md as the final pack, the same pack as a PDF for reading and sharing, and repo-source-manifest.json recording which repositories were pulled and their state.

What lands in the pack, and the honesty rules

The final pack is specified concretely: a conclusion-first recommendation, the candidate pool with rejection reasons, positioning plus existing capabilities and risks per project, a local source verification summary, a project intro of 80 to 120 characters, five to six responsibility or technical-difficulty bullets, suggested resume points, a second-pass modification plan, and interview follow-up directions. That 80-to-120 character intro is a Chinese resume convention, so English CV users will need to adapt the numbers rather than copy them.

Three honesty rules hold the whole thing up. Source evidence first: every final claim must be findable in local source, config, tests, migration scripts or the run entry points. Proposed modifications never masquerade as completed work: unimplemented ideas appear only as what you could write once you finish them. And the skill does not fabricate experience; the README states plainly that you still have to do the reading, deployment and modification before anything goes on the resume.

Those rules are what separate this from a prompt that hallucinates impressive bullets. The pack is designed to survive the exact interview follow-ups it also enumerates.

Apache-2.0, helper scripts, and the alternative

Under the hood the skill is documents plus scripts. The entry point is SKILL.md with the core rules, agents/openai.yaml carries Codex Skill UI metadata, and references/ holds the input template, execution flow, screening and scoring criteria, resume-writing rules, output template and a rule index, alongside three Python helpers:

search_github_candidates.py builds the candidate pool, pull_github_repos.py pulls repositories locally and generates the manifest, and markdown_to_pdf.py produces the PDF. The licence is Apache-2.0, the README credits inspiration from the shushu-internship-tool project it links, and the last push was on 2026-06-22.

The conventional alternative is an awesome-style curated list plus your own vetting. Lists rank by stars and stop at links; all the judgement about what is a demo, what is a wrapper and what survives questioning stays with you. This skill encodes that judgement: mode-gated pools, default shallow-project exclusions, forced local source verification before any bullet is written, and a fixed output contract. If you already know the ecosystem and trust your own reading, a list costs nothing and installs nothing. If you do not, the encoded process is the product.

Editorial conclusion

Adopt Backend Agent Resume Scout if you are a student or early-career backend or AI Agent candidate who wants project picks backed by cloned source instead of star counts, and you run Codex with skills installed. Skip it if you need English resume conventions left untouched, or a tool that does the reading for you, since the README is explicit that unimplemented modifications remain suggestions. Verify first: that the mode you pass matches the pool you want, that repo-source-manifest.json really lists the pulled repositories for your run, and that every bullet you keep traces to a file you opened yourself. The last push was on 2026-06-22.

Frequently asked questions

Official sources

1. Issues
2. License: Apache-2.0
3. lishuangqiang/backend-agent-resume-scout on GitHub
4. README