Practical patterns

Frameworks

These are working patterns from building with AI tools in real engineering work. They are not rules for trusting AI. They are ways to keep human judgment, testing, and review inside the loop.

Current framework

AI-Assisted Development Loop

My practical workflow for using Claude Code without handing over engineering responsibility.

AI-assisted development

The loop I use before AI-written code reaches a branch.

The point is not to make Claude Code perfect. The point is to make the workflow honest: the agent can draft, review commands can help, but manual testing and human judgment still decide what ships.

  1. 01 Open Claude Code in the terminal with a clear task list.
  2. 02 Point it at one task and let it execute the first pass.
  3. 03 Run /review to catch implementation, integration, and context issues.
  4. 04 Run /security-check to look for risky code paths and unsafe assumptions.
  5. 05 Manually test the developed feature like a user and like an engineer.
  6. 06 Review the changeset before committing or opening a pull request.
  7. 07 Let PR review, CI, and teammates stay in the loop before merge.
AI-Assisted Development Loop A workflow from task list to Claude Code, review checks, manual testing, pull request review, and merge into the develop branch. Task list one task at a time >_ Claude Code Execute the task Run /review Run /security-check Manual test prove the feature works Review changeset Commit or create PR develop branch Human judgment trust, verify, override PR + CI review teammate stays in loop

Start from a task list

I write the work down first, then point Claude Code at one task. This keeps the agent away from vague, sprawling changes.

Let the agent produce a first pass

Claude Code can move fast through implementation, but that output is treated as a draft until the review loop finishes.

Force review gates before trust

The work goes through review commands, security checks, manual testing, and changeset review before it reaches a pull request.

Keep human judgment in control

I trust useful output, verify behavior, and override the agent when the code misses context.

Review checklist

5-Check Framework

5-Check is the checklist inside the loop. I use it when the agent has already written code and I need to decide whether that code deserves to move forward.

There is no YouTube video for this yet. For now, this page is the source of truth.

  1. 1

    Intent check

    Does the change solve the actual task, or did the agent add extra behavior that was never requested?

  2. 2

    Integration check

    Does the code fit the existing architecture, naming, data flow, and project conventions?

  3. 3

    Failure-mode check

    What happens when data is missing, APIs fail, permissions change, or the happy path breaks?

  4. 4

    Security check

    Did the agent touch auth, secrets, validation, access control, storage, or external calls in a risky way?

  5. 5

    Maintainability check

    Will another engineer understand, test, debug, and safely extend this code later?