Skip to content

About

TechTwoForty: AI. The finished Accelerate the Build, Verify the Output project, a template and cheat sheet.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

7 Commits

Folders and files

Repository files navigation

TechTwoForty: AI

TechTwoForty: AI

Accelerate the Build, Verify the Output · the finished project

← Previous: Security · Course map · Next: Capstone →

Tip

This repo is the finished template for the AI project. Treat it as your cheat sheet. It is your devsecops-capstone repo after the ai-changes pull request is merged: the /health endpoint with its test, and the dependency upgrade that closes item 5 from the Security triage. Your exact version numbers may differ from these. What matters is that trivy is clean and pytest says 2 passed.

Section 9 of 10 · AI tooling
Time About 45 minutes, 5 phases
Runs on Claude Code (or an agentic IDE) in your capstone repo. Optional: AWS Bedrock
Deliverable Two real changes made with AI help, both read by you and verified by tests and scanners before they were accepted

The big picture

TechTwoForty AI: scoped prompt, Claude Code proposes a diff, you read it, the tools decide, then a PR through every gate

Drawn in Excalidraw. To edit it, open assets/diagram.excalidraw at excalidraw.com.

Important

The one rule of the section: never accept a diff you have not read. Verify with tools, not with how confident the output sounds. And never put a secret in a prompt.

What changed since the Security project

File Change
app/app.py New /health route returning {"status": "ok"}, in the style of the existing route
app/test_app.py New test_health. pytest now says 2 passed
app/requirements.txt flask, werkzeug, requests and urllib3 upgraded to current versions that work together, plus pytest
extras/bedrock_call.py Phase 5, optional: one call to a hosted model through Bedrock

Everything else (pipeline, risk register, gates.md, tools/, infra/) is unchanged from the Security template.

Cheat sheet, phase by phase

Phase 1 · Point the agent at the repo

cd ~/devsecops-capstone && git switch -c ai-changes && claude

On a branch, never on main. Claude Code asks before every command it runs. That approval is the safety model.

Phase 2 · A well-scoped task

The prompt, word for word:

In app/app.py, add a /health endpoint that returns {"status": "ok"} as JSON with a 200 code. Add a pytest test for it in app/. Follow the style of the existing endpoint. Don't change anything else.

A named file, an exact behaviour, a required test, a style to match and a boundary. Compare it with "add a health check", which might touch the Dockerfile, rename things or add a dependency.

Phase 3 · Review, then verify

Read the diff with three questions: does it do what I asked? did it touch anything else? is anything unsafe? Then let the tools decide:

pytest app/                    # 2 passed
semgrep --config auto app/     # no findings
git add -A && git commit -m "Add /health endpoint with test"

Phase 4 · Close the scheduled finding (item 5)

git checkout vulnerable -- app/requirements.txt && trivy fs --scanners vuln --severity HIGH,CRITICAL app/

Only requirements.txt comes across from the vulnerable branch. This repo has the vulnerable branch too, exactly as the Security project left it, so the command works here as well. Then the prompt:

trivy flags app/requirements.txt for vulnerable flask, werkzeug, requests and urllib3 versions. Upgrade all four to current stable versions that work together, keep pytest, run the tests, and show me the diff. Don't change anything else.
trivy fs --scanners vuln --severity HIGH,CRITICAL app/ && pytest app/    # clean, 2 passed
git add -A && git commit -m "Upgrade the vulnerable dependencies to patched versions" && git push -u origin ai-changes

Open a PR and watch every gate from the Security section run against the AI's work. Tests, scans, then a human merge.

Phase 5 · Optional: call a hosted model

python3 extras/bedrock_call.py

Bedrock bills per token, and model access has to be switched on in the console first. Credentials come from your IAM profile, never from the code. In the walkthrough this file sits wherever you created it. It is in extras/ here to keep it apart from the app.

Final handover check

  • Two changes made with AI help, both read and verified before accepting
  • The /health endpoint has a passing test
  • The dependency finding from the Security project is closed and trivy is clean
  • Nothing was merged without the pipeline running

Common mistakes

Mistake What happens The fix
Running Claude Code on main Changes land where branch protection cannot review them git switch -c ai-changes first
A vague prompt A diff that touches files you did not ask about Name the file, the behaviour, the test and the boundary
Approving a diff unread AI bugs ship with total confidence Read every line, then run the tools
Pasting a key or password into a prompt The secret leaves your machine Never. No exceptions
git checkout vulnerable -- . instead of the one file Every planted weakness comes back Only app/requirements.txt

What you can now do

  • Directing an agent: scoped tasks with a named file, exact behaviour, a test and a boundary
  • Reviewing AI output: read every diff, then let tests and scanners decide
  • Closing a backlog: taking a deferred security finding through to verified fixed
  • Hosted models: calling a model from your own code with boto3

TechTwoForty · AI · 9 of 10
← Previous: Security · Course map · Next: Capstone →

About

TechTwoForty: AI. The finished Accelerate the Build, Verify the Output project, a template and cheat sheet.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages