# Start here: pick your path through agentic engineering

URL: https://developertoolkit.ai/en/start/
Description: Where to start on AI Developer Toolkit: jump to the Cursor, Claude Code, or Codex quick start, or pick the reading path for your role.

AI Developer Toolkit teaches developers, tech leads, CTOs, and executives to build software with coding agents (Cursor, Claude Code, and Codex) while keeping quality provable. The site has two entry points: a quick start for each tool, ending in a first real task, and four role tracks, each opening with a free scorecard that sets a baseline.

You have three agents to evaluate, several hundred pages, and a real job to get back to. This page gets you to the right first page.

## Jump to your tool

- [Cursor quick start](/en/cursor-ide/quick-start/) — Install, set up project rules, and ship a first feature in the editor
- [Claude Code quick start](/en/claude-code/quick-start/) — Install, authenticate, and run a first project from the terminal
- [Codex quick start](/en/codex/quick-start/) — Set up the app, the CLI, and the IDE extension, then run a first task

Then pick the next page by where you are coming from:

| If you are… | Read next |
| --- | --- |
| New to agents | [Quick wins](/en/start/quick-wins/) |
| Coming from GitHub Copilot | [Copilot compared with the three agents](/en/comparison/vs-github-copilot/) |
| Adding a second tool | [Cursor vs Claude Code](/en/comparison/cursor-vs-claude-code/) or [Codex vs the others](/en/comparison/vs-codex/) |
| Already fluent in one tool | [Power-user tips](/en/start/power-user-tips/), then the [Cookbook](/en/cookbook/) |
| A platform or DevOps engineer | [CI/CD with Claude Code](/en/claude-code/ship/ci-cd/), [Codex automation](/en/codex/automate/), and [MCP servers](/en/ecosystem/mcp/) |

## Which reading path fits your role?

Each track is an ordered list of pages with an exit criterion per step. Take the scorecard first (about eight minutes): your answers show which steps to skip.

**Developer**

    You want agents to ship work you can trust. The track runs from a first real task to reviewing evidence instead of every diff.

    [Developer Scorecard](/en/scorecard/) · [Developer track](/en/start/developer/)

**Tech lead**

    You need one workflow for the whole team and a review queue that is not the bottleneck: a shared harness, verification design, and review capacity.

    [Tech Lead Scorecard](/en/lead-scorecard/) · [Tech lead track](/en/start/tech-lead/)

**CTO or VP Engineering**

    You decide the operating model, the guardrails, the vendors, and how to prove any of it works. Ten decisions, each ending in an artifact with an owner.

    [C-Level Scorecard](/en/cto-scorecard/) · [CTO track](/en/start/cto/)

**Executive**

    You need the effect on cost, speed, risk, and headcount, with a source on every number, and the decisions leadership owns.

    [C-Level Scorecard](/en/cto-scorecard/) (shared with CTOs) · [Executive track](/en/start/executive/)

**What is free**

Everything in **Start here** and the **Autonomy Ladder** is free, and so are the three scorecards. Every other section needs a subscription; every track marks each step Free or Subscribers.

## Where do you sit on the autonomy ladder?

The site is organized around one map: [Dan Shapiro's autonomy ladder](https://www.danshapiro.com/blog/2026/01/the-five-levels-from-spicy-autocomplete-to-the-software-factory/), Level 0 to Level 5, published 23 January 2026. At Level 3 you review every diff; at Level 4 you write the specification and check the evidence; at Level 5 a pipeline ships while you run the factory. Shapiro puts 90% of self-described "AI-native" developers at Level 2 and says almost everyone tops out at Level 3. The useful next page is the one that moves you up one rung.

- [The Autonomy Ladder](/en/ladder/) — The six levels, and how to place each of your loops on one

## What to do in your first hour

1. Take your role's scorecard and keep the result as your baseline.
2. Finish your tool's quick start on a real repository, not a toy project.
3. Run the two prompts below in that repository: the first gives the agent its context, the second shows which checks stand between it and a bad merge.
4. Commit the instructions file, and put the missing check from prompt 2 on your backlog.

**Copy-paste prompt: draft the instructions file your agent reads**

```text
Analyze this repository and draft the instructions file the coding agent reads at the start of
every session (CLAUDE.md for Claude Code, AGENTS.md for Codex, project rules for Cursor; use
whichever this repo already has). Include:
1) the architecture in five lines, 2) the exact build, test, and lint commands,
3) coding conventions with one real example each, 4) the known gotchas,
5) what "done" means: the checks a change must pass before it is reported as finished.
Keep it under 150 lines. Show me the draft before writing the file.
```

**Copy-paste prompt: audit how a change gets verified here**

```text
Do not modify any files. Report how a change to this repository is verified today:
1) the exact commands for tests, type checks, lint, and build, and whether each passes on this branch now;
2) which of them CI runs, naming the workflow file;
3) what no automated check covers (for example UI behaviour, migrations, or performance).
End with the single missing check that would most reduce how much code a human has to read
before merging, written as a task with acceptance criteria.
```

A faster agent without automated checks only moves work into review; [evidence, not diffs](/en/ladder/evidence-not-diffs/) explains what a reviewer checks instead of every line.

## Where everything lives

The sidebar is ordered by use; every group but Start here starts collapsed.

| Section | What it holds | Start at |
| --- | --- | --- |
| **Start here** | This page, the four tracks, quick wins, and why it matters | [Why agentic engineering](/en/start/why-ai-coding-tools/) |
| **Autonomy Ladder** | Each level and the human's job at it | [The ladder](/en/ladder/) |
| **Cursor · Claude Code · Codex** | One tree per tool, cut by task; the same task in another tool is a one-segment URL swap | [Cursor](/en/cursor-ide/) · [Claude Code](/en/claude-code/) · [Codex](/en/codex/) |
| **Workflows & method** | The tool-independent lifecycle and cross-tool workflows | [Method](/en/method/) · [Workflows](/en/workflows/) |
| **Ecosystem & comparisons** | Choosing a tool, models, MCP servers, skills, and frameworks | [Ecosystem](/en/ecosystem/) · [Comparison](/en/comparison/) |
| **Cookbook** | Prompts by language, framework, and stack | [Cookbook](/en/cookbook/) |
| **Lead a team** | Rollout, the shared harness, review, and people | [Lead a team](/en/teams/) |
| **Engineering organization** | Operating model, controls, security, measurement, and spend | [Engineering organization](/en/org/) |
| **Strategy & business** | Economics and evidence for executives | [Strategy](/en/strategy/) |
| **Reference & help** | Commands, configuration, glossary, and troubleshooting | [Reference](/en/reference/) · [Help](/en/resources/) |

## Why readers stall, and how to get moving again

- **Learning three tools at once.** Nothing sticks. Finish one quick start, and add a second tool only for work the first does badly.
- **Reading without doing.** Nothing changes. Run each prompt on your real repository the day you read it, and keep what works in the instructions file.
- **No instructions file.** The agent repeats the same mistakes every session. Run the first prompt above and commit the result.
- **No verification gate.** Review becomes the bottleneck, or bugs reach production. Run the second prompt, add the missing check it names, and re-take your scorecard after four weeks to see which answers moved.