AI SDLC: The AI-Native Software Engineer in the Agentic Era

, , , , ,
AI SDLC: The AI Native Software Engineer

What an extraordinary time to be in tech! Coding agents are changing everything. This course positions you for success at the forefront of this change.

These intensive 2 weeks prepare you to be industry ready as an AI-Native Software Engineer delivering software hand-in-hand with coding agents.

The course repo includes comprehensive instructions for your coding agent; you can ask it questions and it should be fully briefed. If not, then ask me instead!

Four ways to take the course

  1. Follow me exactly. I use Claude Code in the terminal in VS Code, on a Claude paid plan (claude.com/pricing).
  2. Use your own paid plan with another coding agent. The repo includes guides on all the popular coding agents.
  3. Use a free or cheap plan: OpenCode with its free models, Pi with OpenRouter, or Antigravity CLI on a free Google account. Follow the guide for your agent. Free models will need tons of patience. In the repo, instead of following the instructions at course/instructions, use course/cheap_llm_instructions. This reduces the scope to ensure it is achievable for a small model.
  4. Go free and live vicariously through me. Switch to my code with git each day, using the branches below, and get my results without spending tokens or time.

Main links

Git branches on the course

I make heavy use of git branches on the course. Your coding agent is fully briefed – ask it to clarify at any point.

Your main holds your own application as you progress. My day-end branches contain my results. You should commit your changes before switching branches.

(If you’re using cheap or free models with reduced scope, stay on your own build rather than switching to my branches. Week 2 Days 2 to 4 are lectures only on that route.)

From Week 2 Day 2, the checkpoint instructions use npm run seed after a branch switch, followed by npm run dev. Seeding updates the local database and replaces its records with sample data after confirmation.

Week 2 Days 2 to 4 use checkpoint walkthroughs as the main route. Their instructions also contain the optional full builds. Day 5 returns to building a feature in your own application. Running the cloud factory is optional and uses a paid Claude route; I give you the branches with all the results.

I’d love to see your work! To share your progress, follow the contribution guide and submit the requested documents in your own folder under community_contributions/. Or just ask your coding agent!

Week 1: Ready for industry today

Week 1 Day 1: The AI SDLC

  • I use Claude on the free plan for the opening sales-tracker exercise. Accountable setup begins on Day 2.
  • Before Day 3, skim Anthropic’s AI-native SDLC playbook. We’ll apply its six stages to Accountable on Day 3.
  • Edit: I say junior roles fell from 30% to 10% of job openings. The figure is the share of engineering hires at 28 large US tech companies that were recent graduates: nearly 30% in 2023 and 10% in 2025.
  • Edit: when I say PRs take five times as long to review, the figure is the wait before a review starts, not the time spent reviewing. In LinearB’s data, PRs written by agents waited 17.6 hours, against 3.4 hours for PRs written without AI (at the 75th percentile).

The prompt for the opening exercise, to paste into Claude:

Build me a simple sales pipeline tracker as an artifact that I can open on my phone. Each deal has a company name, a contact, a value in dollars and a stage: Lead, Qualified, Proposal, Won or Lost. Show the deals in columns by stage, let me move a deal to the next stage with one tap, and show the total value of each stage at the top. Start it with five example deals so I can see how it works, and make it easy to use on a small screen.

Check the example deals, move a deal between stages and check the totals. Try adding a deal of your own. To open the result on your phone as I do, choose Copy link from the artifact menu. Open that link on your phone and sign in with the same Claude account.

Sales tracker with a spec and a plan

In a new Claude conversation, put this line above the original sales-tracker prompt:

This is my intent:

Add this below the prompt before sending it:

Do not start development yet. First, please write a detailed Spec and get my OK.

Read the specification and ask for any changes you want. When you’re happy with it, send:

OK that’s approved. Don’t start the build yet. First, please write a detailed Plan with success criteria.

Review the plan and request any changes before approving the build:

OK – please start the build

Check the deal controls, stage totals, Add Deal form and mobile layout. Your generated app may differ from mine.

Week 1 Day 2: Coding agents crash course

  • Setup and coding-agent guides
  • The setup includes Google Chrome for Playwright and the browser tests.
  • Reference branch: week1day2-end. This contains my Claude Code setup. If you use another agent, follow its guide’s steps for adopting this checkpoint.

If you’re new to LLMs, this is my YouTube playlist explaining what on earth all these ‘parameters’ are:

And here’s my playlist about Coding Agents including how to use them for free and cheaply:

Week 1 Day 3: Anthropic AI-SDLC

  • Hopefully you completed your homework and skimmed Anthropic’s AI-native SDLC playbook!
  • The review policy
  • Reference branch: week1day3-end, the local sales CRM.
  • Remember that if you’re using free or cheap models, use the instructions in the cheap_llm_instructions folder.
  • Fork the repository at the start of the day. The Git guide covers sharing your documents at the end.

Week 1 Day 4: Harness engineering

  • All prompts
  • Harness engineering reading: Mitchell Hashimoto’s My AI Adoption Journey, Ryan Lopopolo’s Harness engineering at OpenAI, and Vivek Trivedy’s The Anatomy of an Agent Harness at LangChain.
  • Reference branch: week1day4-end, with the harness and dark mode. Install gitleaks as the instructions describe; the commit hook uses it.
  • The coverage target in my prompt is 80%. You can lower it or omit that requirement, as described in the instructions.
  • Clarification: I say the push won’t complete until CI passes. CI doesn’t hold up the push: the push completes first, then GitHub Actions runs the checks on the commit you pushed. To stop a change merging until CI passes, you’d add required status checks to the branch, which we don’t set up here. And when I say there’s no way around CI, a workflow can still be disabled, and a commit message containing [skip ci] skips the run.

Here’s more on the Dynamic Workflows feature in Claude Code that I show off:

Week 1 Day 5: Subagents and Agent Teams

  • All prompts and deployment steps
  • Claude Code documentation: Run agents in parallel, custom subagents and agent teams.
  • Reference branch: week1day5-end, with the reviewer, UI checker, sign-in, search and user menu.
  • npm run dev uses local SQLite and sample teammates. The shared app and Render deployment require your own Supabase project and settings. Add those settings to .env yourself, following the instructions.
  • The assignment is a feature of your choice through all six stages. A phone or tablet layout is one option.

Week 2: Positioned for the industry of the future

Week 2 Day 1: AI SDLC in Industry

Week 2 Day 2: Deliver as a Team

Week 2 Day 3: Large Codebases, Security, Testing

Week 2 Day 4: Cloud Agents, Software Factory

  • Checkpoint walkthrough and optional factory setup
  • Checkpoints: week2day4-books and week2day4-end, with the books, accounting reports and quote-to-invoice feature.
  • The optional factory route includes the issue intents, setup and review steps. The lead runs the full local checks before merging; the factory’s quick checks omit end-to-end and speed tests.

Week 2 Day 5: Capstone

  • Copilot prompts, deployment and final assignment
  • On the cheap or free route, the capstone is a pipeline board in place of the Copilot, with no OpenRouter key. Use week2day5.md in the cheap_llm_instructions folder.
  • Start from your own shared application or a reference branch from week1day5-end onward. Adapt the Copilot intent to the features in that version.
  • Reference branch: week2day5-end. Add your own OpenRouter key to .env as the instructions describe. The default model setting is openrouter/free.
  • Edit: my estimate of less than a ten-thousandth of a dollar per call is too low. OpenRouter lists GPT-6 Luna at $0.10 per million input tokens and $0.50 per million output tokens, so a call with 1,000 input and 200 output tokens costs about $0.0002, and a Copilot conversation can make several calls. It’s still inexpensive, but set a small credit limit on your key and keep an eye on usage.
  • The final assignment is an import from Salesforce, Pipedrive or another CRM. Follow the instructions for exports, record mapping and handling repeated imports.

Please stay in touch!

Thank you for sticking to the very end! Remember that I come with the package.. please ask me questions anytime on Udemy. You can also connect with me on LinkedIn or subscribe on YouTube. I’d love to see your completed application and final assignment; please consider making a PR, and posting about it so I can share in your achievement. Congratulations!

Leave a Reply