Use this Codex skill beside your usual PDF reader. Keep reading in Zotero, Office, Preview, or a browser; Codex finds the local PDF behind a copied title, explains the paper with you, tracks related sources, writes concise HTML notes, helps you build a scoped survey, and can deposit selected work in Feishu/Lark.
Built and maintained with Codex.
- Install the skill once.
- Choose a local folder for your papers.
- Drop downloaded PDFs there without renaming them.
- Copy a title from your PDF reader and talk to Codex.
You do not need to make folders, rename numbered downloads, build an index, or edit a registry by hand.
Requires Python 3.9 or later. Optional PDF text/rendering tools are checked by Codex during setup.
The easiest installation is to give Codex the repository link and say:
Install or update the Codex skill from https://github.com/hwang847/codex-paper-reader.git as literature-workflow. Check for an existing installation first, preserve local changes, validate SKILL.md, and tell me when it is ready.
Codex can place and validate the skill for you. If you prefer a terminal, install it manually:
mkdir -p ~/.codex/skills
git clone https://github.com/hwang847/codex-paper-reader.git ~/.codex/skills/literature-workflowmkdir -p only ensures that the parent skill directory exists; it is harmless when the directory is already present. An installed skill is normally available on the next Codex turn. If the current task does not detect it, open a new Codex App task or a new CLI session.
Then open your chosen paper folder, or cd into it, and say:
Use $literature-workflow to set up this literature workspace.
Codex creates the workspace, checks available PDF tools, and writes a compact local AGENTS.md from your preferences. Say that you want automatic filing if you want new papers organized automatically.
After setup, your everyday loop is simple:
- Put a new PDF in the workspace root. Names such as
2508.05002v1.pdf,2025.findings-naacl.245.pdf,download.pdf, orPaper.PDFare all fine. - Open it in your normal PDF reader and copy its title.
- Paste the title into Codex:
Use $literature-workflow. Read this paper with me:
APEX-SQL: Talking to the data via Agentic Exploration for Text-to-SQL
Codex refreshes the local index when needed, identifies the corresponding file, starts from the abstract, then explains the thesis and implementable method pipeline. Ask follow-up questions naturally, for example:
Walk me through the pipeline as if I had to implement it.
What state does each agent keep, and how does feedback change the next step?
Relate this paper to the literature review I am writing.
Generate a concise HTML note for this paper.
The HTML note reflects the paper and your discussion. Later questions that change the interpretation should update the same note; supported local renderers can add native PDF screenshots.
Choose the papers and tell Codex the question and boundary in ordinary language:
Use $literature-workflow. Build a survey scope on execution-feedback training for Text-to-SQL.
Use ExCoT, SQL-Trail, and MTSQL-R1 as the core papers.
Treat my existing integrated_survey.md as read-only source material.
Codex creates a small, traceable work pack under surveys/<topic>/: a source manifest and a reading pack. It can register your existing survey draft, or move it into surveys/sources/ when you ask, without changing its contents. Then ask Codex to compare papers, shape the narrative, identify gaps, or write/update survey prose in your chosen style.
Automatic filing is opt-in: enable it while setting up the workspace, or ask Codex to organize the new PDFs.
When enabled, the workflow:
- detects the real title from PDF text and ignores generic conference metadata when possible;
- renames a download to a normalized paper-title filename;
- places it in one primary
papers/<category>/folder using editable local rules; - preserves duplicate files with a suffix instead of overwriting them;
- updates the search index, source registry, BibTeX
Sourcecomments, and links in local notes when paths change.
The starter categories are deliberately general. Tell Codex how you want your library grouped, and it will adapt references/organization_rules.json and your workspace-local AGENTS.md. Ambiguous papers stay in the fallback category. The workflow never deletes PDFs.
When a note, survey artifact, or sharing path is ready, give Codex the exact Drive-folder or Wiki link and say what to publish:
Use $literature-workflow to publish recommended_readings.md to this Feishu folder: <link>.
Codex checks for the official lark-cli, explains the exact install before asking permission to install it, guides the one-time authorization, previews the write when supported, and then creates the requested cloud artifact. It will not sync, overwrite, move, delete, or change sharing permissions unless you explicitly ask.
your-workspace/
├── papers/ # organized PDFs
├── notes/ # HTML notes
│ └── assets/ # native PDF screenshots
├── references/ # search index, source registry, filing rules
├── surveys/ # scoped survey manifests, packs, and managed source drafts
└── AGENTS.md # your durable reading preferences
One reading target can have many sources: a local PDF, arXiv, DOI, project page, repository, slides, or technical documentation. Codex records these as one entity, so a copied title does not need to match the downloaded filename.
Tell Codex preferences in ordinary language:
Use Chinese for reading and notes.
Focus on implementation details and pipeline reconstruction.
Skip experiments unless I ask.
I am writing a survey, so emphasize related work and positioning.
Track project pages and GitHub repositories with each paper.
Codex keeps durable preferences compactly in that workspace's AGENTS.md; the public skill supplies the mechanics, while your workspace defines the research context.
python3 tests/smoke_test.py
python3 tests/privacy_scan.py