My AI Coding Workflow, January 2026

I wanted to capture how I work right now. A snapshot. In a year this will probably look outdated, maybe embarrassing. That’s the point.

Here’s the setup, the tools, the habits. January 2026.

The Setup

I run Claude Code with --dangerously-skip-permissions. That’s the only practical way to use it. Otherwise you’re approving tasks all day and the agentic loop breaks down.

Yes, there’s risk. My machine is set up to recover at any point. The trade-off is worth it.

For models: Claude Opus 4.5 for agentic coding (€200/month subscription), Gemini Flash for everything else. Ran the whole last year on Gemini 2.5 Flash for chat and extraction, never had issues. Why use something mediocre when Opus is the best agentic model right now?

Also experimenting with open source: Kimi K2.5 and GLM 4.7 are surprisingly good agentic models when you want alternatives.

I haven’t touched ChatGPT in over a year.

Starting Projects

When I start something new:

1. Capture a big scope of related tasks
2. Write up plans in markdown
3. Talk through the whole thing using voice
4. Let Claude explore the codebase and problem space
5. Claude comes up with the first plan
6. Read the plan with voice on, talk through what’s not great yet
7. Spend considerable time making it a good plan
8. Revisit the plan often during implementation

The plan becomes your anchor. You track progress, add new tasks, and work your way through it across sessions.

Voice is a game changer here. I use VoiceInk with the Parakeet model (faster than Whisper, runs locally). You can dump so much context via voice, then Claude takes a first shot at organizing it.

Context Management

Start a new conversation when the context gets polluted. Quality falls off as the context fills up.

CLAUDE.md holds the important persistent context, but keep it small. Don’t overdo it. Progressive disclosure works better: save detailed context for when it’s actually needed.

What I Do vs What Claude Does

My job:

• High-level architecture. Know how everything connects.
• Create visibility and tooling. Build skills and CLIs so Claude can figure out why something doesn’t work.
• Break down tasks (sometimes with Claude’s help).
• Final QA testing. Hands-on, like a PM would do.
• Tech stack choices and integration decisions.
• Know the end goal because you talked to the client.

Claude’s job:

• Lower-level implementation
• Self-validation (tests, linters)
• First pass at organizing tasks from voice input

You are the puppet master on top. Everything below that level should be AI validating itself.

Handling Errors

Coding is so cheap now. If something goes wrong, just start again. You learn something and get better each time.

Git is your friend. Work with checkpoints a lot. You can always go back. This gives you room to experiment.

Always close the loop. Provide a self-validation loop: tests, linters, everything that matters. Be sure it works as expected before moving on.

Red Flags

Signs the output is going sideways:

• Claude starts doing weird stuff with new scripts to test something
• Excessive mocking in tests
• Random new files appearing

Keep an eye on it, especially at the start of a session, to see it’s going in the right direction. But cycles are short. You can also just validate the end result.

Skills

I’ve built a library of custom skills. Small CLI tools with a SKILL.md prompt that tells Claude how to use them. Gmail, Slack, n8n workflows, Kanban boards, time tracking, deployment, DNS, databases, transcription, and more.

The philosophy: build for AI agents, not humans . Structured output over pretty output. --json flags everywhere. No spinners or colors. Small scripts that do one thing. Easy to chain. Agents combine primitives.

At time of writing: 171 commits, 35 skills. Each one is a force multiplier. And they keep improving themselves .

Real Output (January 2026)

Personal projects, on the side:

Pulumi Infrastructure Automation. Migrated entire AWS and Google Cloud infrastructure to code. Didn’t know Pulumi before. Now Claude can manage infrastructure directly.

Invoice Validation Refactor. Had Claude write tests for the current structure first. Then refactored the complete validation logic against those tests. Made sure everything still passes.

n8n Workflow Reviewer. A linter for n8n workflows. Catches silent failures, memory issues, poor patterns before they cause problems. Works as CLI or library. Can review directly from n8n instances via API.

n8n Test Blueprint. TypeScript testing framework for n8n subworkflows using Vitest. Finally, proper tests for automation workflows.

Custom Report Generator. Client wanted specific design for reports. Hit the limit of n8n to Canva automation. Built a custom engine that outputs exactly the specified template. Haven’t touched the code. Literally only steered it with voice.

Ansible Server Setup. Spin up new machines for customers. Install open source software (started with n8n , now around 10 different apps). Security hardening, monitoring, backups. Can run commands on all servers at once. Manages around 10 client servers.

Meanwhile: 5 active clients. n8n automations, Rails backends, voice agents, Java development, Ghost development. All those personal projects are on top of active client work.

Non-Code Uses

Workout tracking and reporting to my coach. All workouts are in Hevy . Claude can search through them completely.

Podcast discovery. Following specific guests. Built a skill using Taddy for that.

Readwise is my central hub for articles. When I remember I read something relevant, Claude can search it because everything streams into Readwise.

Bookkeeping with LexOffice . Claude has access to my data. When questions from the accountant come in, it’s been helpful.

Analyzed my blood work recently. Brought in context from various books I’ve read. The recommendations were solid.

YouTube video summaries that go into Readwise for later reading. Slack voice messages transcribed and summarized. The skills chain together.

The Point

It’s a skill. You have to commit to it, actually use it. Whenever you hit friction in your day-to-day, ask: Why is this still hard? What can I build? How can I give Claude the tools to make it better?

How crazy is it that you can clone any repository from the internet, tell Claude to explore it, and it can explain what it does? I wouldn’t mind being thrown into any project anymore. There are no limits.

Ship it, see what happens.