| Claude (Anthropic) | Codex (OpenAI) | |
|---|---|---|
| Chat | claude.ai | chatgpt.com |
| Workspace agent | Cowork | Codex app |
| Code agent (VS Code) | Claude Code extension | Codex extension |
| Instructions file | CLAUDE.md | AGENTS.md |
| Reusable workflows | Skills (.md files) | Skills (.md files) |
| External tools | MCP servers | MCP servers |
| Subagents | Claude Code subagents | ✓ Just shipped (Mar 16) |
Live data connections
APIs & databases
Reusable workflows
Just markdown
Parallel autonomous work
Now live in Codex
Drop a dataset. Ask it to build a dashboard. Watch it reason, plan, write code, and execute — without you touching a line.
Let's do this
together.
code.visualstudio.com — if you don't have it yet
Search "Codex" in the VS Code Extensions panel
New file in root — this is where we're going next
An AI without context is just a chatbot. Structure gives it your world.
The AI reads
this first,
every time.
If it gets something wrong — that's your cue to improve the file.
Write it once,
use it forever.
A skill is a markdown file with instructions the AI follows when you invoke it.
Open skills/daily-digest.md from the starter kit
Four parts.
All plain English.
Then test it: "Give me my daily digest"
The other half
of your OpenAI
subscription.
Run parallel tasks across projects, review results in one place
Review diffs, stage changes, commit — all without leaving the app
Scheduled background tasks: scan for bugs, triage issues, generate briefs
Extend your agent with reusable workflows and external tool connections
You have both. Use them for different things.
| ChatGPT | Codex App | |
|---|---|---|
| Best for | General tasks, research, writing | Workspace-aware coding tasks |
| Data analysis | Upload CSV/PDF, build charts | Analyze your codebase and project files |
| Automation | Scheduled prompts, email output | Recurring dev tasks, results in app inbox |
| Context | Conversation only | Reads your whole workspace (AGENTS.md, skills, files) |
| Execution | Sandboxed code, web browsing, canvas | Local terminal, Git, parallel agents |
You just customized the skill. Now make it run automatically.
Tomorrow morning: your digest is waiting in the Codex inbox when you open your laptop.
Create skills/interview-prep.md
Now you know the pattern — trigger, steps, output, rules.
Customize for
your beat.
Your interview prep skill works for everyone on the team
More context = more relevant help. Beat notes, source lists, past research.
Ask it to self-correct: "That missed X — update the skill."
Anthropic cataloged hundreds of internal skills. They cluster into 9 categories.
How to use tools correctly
Test that output is correct
Connect to data sources
Recurring workflows
Generate boilerplate
Enforce standards
Ship and monitor
Symptom → investigation → report
Maintenance with guardrails
Insider tips from
Anthropic's team.
The AI already knows how to code. Focus on what makes YOUR workflow unique.
Highest-signal content. Every time the AI fails, add the fix to the skill.
Include reference files, templates, examples. The AI reads them when needed.
Give the AI information and flexibility — don't railroad every step.
It tells the AI WHEN to use the skill, not what it does.
Use your assistant at least once for an actual task. Note what works and what's missing.
Add one thing you wish it knew about you or your beat. Add a context file.
Try to make your assistant useful for your capstone project. Bring context, data, or sources you'll need.
Try automation / scheduled tasks.