Author(s): Nick Porter

Originally published on Towards AI.

Hiya! My name is Nick and I love writing Web apps! I have been writing in various languages both front-end and back-end professionally since ’98. My workflow, needless to say, has evolved considerably over that time.

There are incredible workflow improvement announcements coming out of Anthropic. Initially I was excited about the 1m context window but after digging deeper there are more meaningful impacts from other recent improvements to Claude Code.

we were months ahead

Anthropic continues to roll improvements to their product which is incredible and sometimes a little overwhelming. Don’t worry I stay on top of this and look for the latest workflow improvements coming out. In fact we were months ahead of these recent announcements with our own system which gracefully transitioned to the version in this article that is using more native features of Claude Code.

Anthropic announcing a larger Claude context seemed inevitable but it will not solve all the issues with long-running coding session and (for now) will not replace the importance of session files. It does however make redundant one part of that old system and that is the gotcha’s.

With our adjustments we also went from 4 commands to 1 in each coding session. This process makes the long-running coding session feel more refined, a little less bloated by commands and smoother from start to end. One of the greatest parts is that it kinda removed some of the manual steps I had to take care of or do a little reasoning about.

Why Recent Changes Impact My Workflow

Shipping alongside Opus 4.6 were features in Claude Code that changed my workflow pretty significantly. If you have read my previous articles (thank you) you know that I would prepare Claude for certain phases of the workflow with commands. Many of them have been replaced by Anthropic making these workflow features native to Claude Code.

Well have you had times coding in Claude where an autonomous process takes over and it just feels … disjointed.

Over the last few months my workflow has looked something like this.

1. Discuss feature with Claude & start a session file use the/orchestrate command for agents
2. continue to update session files by calling /update-session
3. call /wrap-up to run quality standards and write a commit message
4. call /push to run the quality gates, make fixes, and push the code

What my workflow looks like after the latest Claude Code V2.1.37 improvements.

1. Discuss feature and session file with Claude
2. do work
3. call /build-pipeline

Did We Lose Anything?

No! Because Anthropic has made native the very features that engineers in and outside their ecosystem pushed for. I will show you how adopting these native features changed my workflow and how you can use them in yours to improve the quality of code as well as how the experience feels.

Or ideally I didn’t want to even have to think about it.

What do I mean by ‘how the experience feels’? Well have you had times coding in Claude where an autonomous process takes over and it just feels … disjointed. You don’t really know the results of that agent taking over checking your code. Much of the previous process still relied on prompting too. This meant me manually reminding Claude to do something like stay in Orchestrator mode.

I know the frustration you may be feeling. I wanted the feeling of building a new feature or running through the code quality gates while still in the middle of a long-running task to feel more natural. Or ideally I didn’t want to even have to think about it.

1) Auto-memory (V2.1.32) — Replaced /update-session
2) Agent Teams (V2.1.32, research preview) — Replaced manual fan-out fan-in orchestration templates
3) Delegate Mode (V2.1.32 ships as part of Agent Teams) — Replace ~100 lines of CLAUDE.md Orchestrator scaffolding
4) Project-level Hooks in settings.json — Moved hooks from command-scope YAML to always-on JSON

What is Auto-memory

So something called Auto-memory sounds like a super good idea but what exactly is going on under the hood?

Claude reads these files on startup now

As Claude works with you it identifies patterns, gotchas, and constraints that would be useful to remember across sessions. Claude then automatically writes these to persistent memory files — specifically MEMORY.md and topic specific files in your .claude/ directory. Claude reads these files on startup now and has immediate access to those memories and context.

What does Claude collect organically (no prompting needed)?
— Codebase specific constraints — banned coding practices
— Architecture patterns — jest.setup.js has 3 tRPC mock locations update ALL 3
— Key identifiers — MCP tool connection IDs
— File-level knowledge — NutritionCard.tsx is 1200 lines use custom hooks for new logic

These are real examples that are in my system already!

What is Agent Teams

Agent Teams is a native multi-agent coordination system where the lead agent delegates work to teammate agents that work in parallel, coordinating through a shared task list. This is currently a research preview (V2.1.32). Sound familiar?

TDD Kills Claude Context: Using Orchestrator Pattern With Strict TDD Adherence

Heya, my name is Nick and I have been developing Web applications professionally since 98. Working with HTML and…

ai.plainenglish.io

Agent Teams formalized what we had hand built way back in 2025. This is a really huge improvement and gives incredible power to any first time user right out of the box.

What is Delegate Mode

At the beginning of each session I would still need to remind Claude to always be in Orchestrator mode. Even within a single session Claude will still tend to jump straight into code. I would have to stop Claude and make sure they stayed in their lane.

When you have ~100 lines of CLAUDE.md instruction telling Claude how to behave and it still doesn’t stick, that’s a sign that you need something stronger than prompting.

Delegate Mode (Shift+Tab in Claude Code) solves this natively. It tells Claude to delegate it’s work to sub-agents instead of writing code directly. The lead agent stays focused on the big picture while teammate agents handle the implementation in parallel.

By enabling Delegate mode as the default we removed ~100 lines from CLAUDE.md that were trying to describe the exact behavior through instruction. That’s been a consistent theme for this article, enforce through platform not prompting.

// To enable Agent Teams by default
// add this to your .claude/settings.json file
"env": {
 "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
},

Auto-memory can remind Claude about coding patterns but it can’t run your test suite for you.

How Do Project Level Hooks Help

Project level hooks have been around a while but they help wire together the remaining pieces into the consolidated single command that I use with every Claude session.

The big one is quality-gates.sh. Before this change our quality gates only ran when I called /push. Now they fire as a PreToolUse hook on every git push attempt. If tests fail, type-check fails, or lint fails the push is blocked. Auto-memory can remind Claude about coding patterns but it can’t run your test suite for you. Hooks can.

 // .claude/settings.json
 {
 "hooks": {
 "PreToolUse": [
 {
 "matcher": "Bash",
 "hooks": [{
 "type": "command",
 "command": "\"$CLAUDE_PROJECT_DIR\"/.claude/hooks/quality-gates.sh"
 }]
 }
 ],
 "PostToolUse": [
 {
 "matcher": "Edit|Write",
 "hooks": [{
 "type": "command",
 "command": "\"$CLAUDE_PROJECT_DIR\"/.claude/hooks/coding-standards.sh"
 }]
 }
 ]
 }
 }

We also kept a lighter coding-standards.sh hook on PostToolUse as a safety net. Auto-memory handles most of what it checks but Claude can still slip up after compaction. This is low cost insurance.

The Build Pipeline Consolidation

This is where it all starts to get wired in together. We had 4 commands that each did a part of the end-session work.
—/orchestrator — manage multi-agent coding and research
— /update-session — save what we learned and how far we got
— /wrap-up — coding standards check, create commit message
— /push — run quality gates then push

Because of recent advancements some of the instruction provided by these commands became native to Claude Code and came out of my workflow entirely! We moved the rest to one command. The /build-pipeline command looks like this.
ANALYZE →CHECK →GATE →SESSION →STAGE →COMMIT →PUSH →REPORT

Much of the ceremony that was in my workflow, a good chunk being the management of session files, became redundant by the recent releases from Anthropic. Workflow managed tools were replaced by tools such as:
— Delegate Mode
— Agent Teams
— Auto-memory

We Must Adapt To Survive

The ever changing landscape of using AI tools in your engineering workflow has been given a couple turbo chargers. I’m sure you’re like me and expect more changes on the horizon.

Anthropic has shown us that the feedback loop reward early adopters. We built patterns and hammered them out over time, over the last few months. Anthropic just made those into standards and shipped them as native features.

It has been fantastic learning and putting the findings down here on Medium. I hope you will continue to learn and build new things with me!

Footnotes

Where 1m Context Is Available Now

Opus 4.6 launched Feb 5, 2026 with a 1m context window. Anthropic’s first Opus-class model with this capability. It’s described as “in beta” in the official announcement.
— Claude.ai (web/app) — Yes, available now
— Anthropic API — Yes, with the model ID claude-opus-4–6
— Cloud platforms — Yes (AWS, Bedrock, Azure Foundry, Google Vertex)
— Claude Code CLI — No, still 200k

Sometimes The Hare Wins

In software engineering while using AI tools being ahead can help you prepare for when your early thinking becomes standard. Anthropic has recently codified concepts, patterns and architectures we built together MONTHS ago. As a reminder here are some of the things we built ahead of the curve that are now codified by Anthropic.

 ┌───────────────────────────────────────────────────┬───────────────────────────────────────────────────────────┐
 │ What We Built (2025) │ What Anthropic Shipped (Feb 5, 2026) │
 ├───────────────────────────────────────────────────┼───────────────────────────────────────────────────────────┤
 │ CLAUDE.md "You are the ORCHESTRATOR" (~100 lines) │ Delegate Mode (Shift+Tab) │
 ├───────────────────────────────────────────────────┼───────────────────────────────────────────────────────────┤
 │ /orchestrate fan-out/fan-in templates │ Agent Teams with shared task list │
 ├───────────────────────────────────────────────────┼───────────────────────────────────────────────────────────┤
 │ /hand-off session transfer protocol │ 1M context window (fewer handoffs needed) │
 ├───────────────────────────────────────────────────┼───────────────────────────────────────────────────────────┤
 │ /update-session manual learning capture │ Auto-Memory │
 ├───────────────────────────────────────────────────┼───────────────────────────────────────────────────────────┤
 │ quality-gates.sh as PreToolUse hook │ Hooks in settings.json (same concept, better integration) │
 ├───────────────────────────────────────────────────┼───────────────────────────────────────────────────────────┤
 │ coding-standards.sh only during /wrap-up │ Always-on PostToolUse hooks │
 └───────────────────────────────────────────────────┴───────────────────────────────────────────────────────────┘

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI