backend-agent-resume-scout: a Codex Skill that clones the repo before it praises it

The install target is a skill directory, not a package

There is no pip install and no build step. The deliverable is a directory of prompts, rules and two Python scripts that you copy into the place Codex looks for skills. On macOS and Linux:

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

On Windows PowerShell the same copy is expressed with the profile path:

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

Note the naming: the repository is backend-agent-resume-scout, while the directory you install and the name you invoke are both backend-agent-project-selector. The README credits part of the idea to a separate internship tool and describes the author as focusing on backend work, AI agents, and resume project design. The repository is Apache-2.0 licensed and the primary language is recorded as Python, which refers to the two helper scripts rather than to anything you import.

Five modes, and the skill asks before it searches

The first gate is a question, not a search. One of five recommendation modes has to be chosen explicitly, and if you do not name one the skill asks you to rather than starting a search or producing a recommendation. agent-only returns complete business-grade AI agent projects. backend-only returns traditional software backend projects and by default excludes IoT and hardware integration work. mixed asks for one agent project plus one traditional backend project. safe-mode aims at projects that are easy to land, with few dependencies and low deployment cost. challenge-mode deliberately points at harder, more differentiated projects suited to deep modification. The distinction that matters in practice is between safe-mode and challenge-mode: one optimizes for finishing, the other for having something to talk about, and picking the wrong one wastes the shortlist.

README filters, local source is the only accepted evidence

The staged flow is fixed, and one rule runs through all of it: README pages and web search are for pre-filtering only, while the final responsible features and technical difficulties must come from verifying local source code. The stages are:

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

Evidence has to be locatable in the local source, configuration, tests, migration scripts or run entry points, which is a stricter bar than reading a project's own description. Candidate bucketing is deliberate too, spanning e-commerce transactions, CRM and ERP, collaborative office, ticketing and customer service, knowledge bases and agent workflows, and the tool is supposed to cover medium popularity and niche areas rather than chase star counts.

Three or four candidates survive before anything is cloned

The shortlist is a hard stop, and it happens before any repository is downloaded. After the search and a README level type screen, the skill outputs three to four shortlisted projects and waits for you to confirm the direction. Only then does it clone. That ordering is the difference from an ordinary list of project links, which hands you a pile of URLs and leaves the sorting to you. The same contrast runs through the elimination rules: demos, thin wrappers, pure framework projects, browser extensions, simple chatbots and shallow CRUD work are excluded by default, on the stated grounds that candidates otherwise mistake tutorial clones and tool shells for resume projects. The output is meant to carry a reason and a risk for each candidate, not just a name.

Four files land in your workspace and only two are the resume

The skill writes into the current working directory, and the four artifacts are separated by purpose. backend-agent-project-shortlist.md is the confirmation draft, holding the candidates, the reasons for choosing them and the main risks. backend-agent-project-resume-pack.md is the final package with source evidence, responsible features, technical difficulties and modification suggestions. backend-agent-project-resume-pack.pdf is the same thing in a shareable form, and the PDF plus the Markdown are both produced by default. repo-source-manifest.json records what was pulled locally and the repository state. Inside the pack, the shape is fixed as well: a conclusion-first recommendation, the candidate pool with elimination reasons, the mode and diversity notes, per project positioning with existing capabilities and risks, a local source verification summary, an 80 to 120 word project introduction, five or six responsible features, and a follow-up modification plan with interview angles.

The rules live in references/, and two scripts do the fetching

SKILL.md is the entry point and the core rules, and the detail is pushed down into a references/ directory so it can be read on demand. The reference set covers the user input template for mode and profile, the execution flow for search, screening, confirmation, verification and output, the screening and scoring criteria, the resume writing rules for bullets and technical difficulty framing, the output template, and an index of the rule documents. Two Python scripts sit in the same directory: search_github_candidates.py for building the candidate pool from GitHub, and pull_github_repos.py for cloning sources and writing the manifest. There is also an agents/ directory holding a single openai.yaml described as Codex Skill UI metadata. The practical consequence is that the scoring thresholds are not summarized in the entry file, so judging whether a recommendation was filtered fairly means opening the scoring reference yourself.

Suggested changes are labeled so they cannot pass as achievements

The design principles include one rule aimed squarely at resume inflation. The skill separates existing capability, suggested modification, and what is writable into a resume, and anything not implemented stays in the middle column: it can only be written down after you have done the work. The stated boundary is explicit. The tool helps you choose and express projects, it will not fabricate experience on your behalf, and before anything goes on a resume you still have to complete the corresponding reading, deployment, modification or implementation. A second principle pushes back on keyword resume padding: a project is not qualified because it used Redis, a message queue, Elasticsearch or an LLM, but because you can explain its business data, state transitions, exception handling, permission boundaries and user value. A third puts business closure ahead of technology names.

A skill with a Chinese entry point and an English summary

The documentation is bilingual in an uneven way. The main body is written in Chinese, with section anchors for the problem it solves, the recommended modes, the workflow, the project structure and the quick start, and the README carries a Chinese to English summary link near the top. All five usage examples are given as Chinese prompt text invoking the skill by name with a mode prefix, for example asking for one business-grade agent project and one traditional backend project under mixed mode, or asking for Java backend projects deeper than an ordinary mall project. That makes the skill usable as shipped for a Chinese-reading user and usable in English only after rewriting the prompts yourself, since the reference documents, the output templates and the examples are all Chinese. The repository also carries an assets/ directory alongside the skill folder and the README, and publishes no GitHub releases.

Editorial conclusion

Use this if you are a student or early-career engineer who can pick up a real codebase in a couple of weeks, because its whole value is the clone-and-verify step and the shortlist gate before that. Two limits are worth stating before you rely on it: the five modes and the scoring thresholds live in references/ files you have to read yourself, and the tool explicitly will not fabricate experience, so anything it marks as a suggested change is work you still owe. The last push was on 2026-06-22, there are no GitHub releases to pin, and the skill directory is named backend-agent-project-selector even though the repository is not.

Frequently asked questions

Official sources

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