AI engineering workflow

How I use coding agents in day-to-day development.

I use Claude Code, Codex CLI, Pi, Cursor, and supporting orchestration tools for repository research, implementation, debugging, and review.

I remain responsible for the requirement, architecture, final diff, tests, browser behavior, and deployed result. Agents shorten parts of the work; they do not approve their own output.

2: Daily-driver models
5: Workflow stages
16: Reusable agent skills
4: Working surfaces

01 · Models

The models I use most.

I choose between them based on the task, repository context, and quality of the result.

Daily driver

GPT-5.6 Sol

Repository implementation, code analysis, debugging, test planning, and review.

Daily driver

Sonnet 5.0

Product requirements, architecture discussions, technical writing, and interface review.

02 · Workflow

A five-step path from request to production.

Each stage produces something concrete that I can inspect before the work continues.

01
Frame
Write down the user need, constraints, existing evidence, open questions, and acceptance criteria before editing code.
Scope and acceptance criteria
02
Challenge
Check assumptions against the repository and documentation, then identify edge cases, failure modes, and unnecessary scope.
Reviewed approach
03
Build
Break the plan into bounded changes, keep responsibilities clear, and review each diff before moving to the next step.
Reviewable changes
04
Verify
Run the relevant tests, type checks, browser flows, accessibility checks, and production-like scenarios.
Verification evidence
05
Ship
Review the complete diff, record important decisions, commit the change, and check the deployed result.
Verified release

03 · Setup

Where the work happens.

Terminal agents handle repository work, desktop tools support research, and editors keep direct code inspection close.

Orchestration

Run parallel agent sessions and keep each task, terminal, and result visible.

cmux
HERDR

Terminal agents

Inspect repositories, edit code, run commands, diagnose failures, and review changes from the terminal.

Claude Code
Codex CLI
Pi

Desktop agents

Research documentation, compare approaches, review plans, and work through decisions outside the editor.

Claude app
Codex app

Editors

Edit and navigate code directly, use completions where they help, and inspect every change in context.

Cursor
VS Code
GitHub Copilot

04 · Skills

Reusable instructions for recurring engineering tasks.

Each skill defines when it should run, the steps it follows, and the evidence required before the task is complete.

Product and interface

impeccable
Design, audit, polish, or optimize frontend interfaces and UX.
frontend-design
Create distinctive, intentional visual designs for new or existing UIs.
prototype
Build throwaway prototypes to explore logic, state models, or UI.
web-design-guidelines
Audit UI code for accessibility, UX, and web-interface best practices.

Architecture and quality

code-review
Review changes against repository standards and the originating spec.
codebase-design
Design deeper, clearer module interfaces and architectural seams.
diagnosing-bugs
Diagnose difficult bugs, failures, and performance regressions.
domain-modeling
Define domain terminology, models, and architectural decisions.
tdd
Develop features and fixes test-first using red-green-refactor.
resolving-merge-conflicts
Resolve active Git merge or rebase conflicts.
vercel-react-best-practices
Optimize React and Next.js code using Vercel’s performance guidance.

Research and decisions

grilling
Relentlessly stress-test a plan, decision, or idea.
research
Research topics using high-trust sources and record findings in Markdown.

Agent operations

wizard
Create interactive Bash wizards for human-only setup or migration steps.
writing-for-agents
Create or edit agent instructions, skills, AGENTS.md, and CLAUDE.md.
agent-browser
Automate websites, browser testing, scraping, screenshots, and Electron apps.

05 · Guardrails

What I verify myself.

01Define the requirement and acceptance criteria before choosing a model or agent.
02Give each agent a bounded task with an output that can be reviewed.
03Check repository code and primary documentation before accepting a claim.
04Run tests, type checks, browser checks, and direct code review before release.
05Keep employer, client, credential, and private repository data out of prompts and portfolio material.