Claude Code best practices are changing fast

A conference talk on Claude Code reveals a lesson far bigger than coding: every rule your business built for an old problem deserves a second look once the problem is gone.

A year ago, getting real work out of an AI coding agent meant building it a cage: long system prompts, strict file-access rules, a step-by-step playbook for almost every task. Skip the cage, and the agent wandered off script. That trade-off felt unavoidable — until the underlying models got noticeably better. The core claim behind a recent conference talk on Claude Code best practices is blunt: the cage built for last year’s model is now, for many teams, the very thing holding this year’s model back.

The talk, titled Rethinking Claude Code Best Practices, was delivered by Lydia Hallie, a Member of Technical Staff at Anthropic who works on developer relations for Claude Code — the company’s AI agent built to write and edit code directly inside a terminal or editor. It was posted to the AI Engineer YouTube channel on 12 November 2026, runs about 18 minutes, and is pitched squarely at working developers rather than casual viewers.

AI agents got good enough to embarrass our old habits

Hallie’s job at Anthropic, by her own description, is to translate how Claude Code actually behaves into guidance developers can put to use, through talks, workshops and technical writing. The premise of this particular talk is that the models have gotten really good, and most of what developers added to their Claude Code setups over the past year is now holding them back. The newer Claude 5.5 models, her argument goes, can carry far more of that weight unassisted.

That is a meaningful distinction, not a reason to assume every old habit is dead weight. A rule written to patch a weakness earns its place only as long as the weakness is still there. Once the underlying problem is solved, the rule does not expire automatically — it just sits there, quietly slowing people down, until someone bothers to ask whether it is still needed.

Inside the talk: Lydia Hallie on Claude Code’s shifting best practices

This is where the talk becomes more than an internal engineering note. Hallie is explicit that her goal is to separate what is still worth keeping from what developers no longer need to think about at all — a distinction every team using a fast-moving AI tool eventually has to make, whether the tool in question writes code, drafts emails or answers customer calls.

Tools and conditions move on; the rules rarely get a second look unless someone forces the question.

The real audience for this lesson reaches beyond software teams

Few InnoHub readers spend their afternoons tuning an AI agent’s configuration file. But the pattern Hallie describes shows up anywhere a business built a rule, a workaround or a manual check to cover for a weakness that has since been fixed — by better software, a calmer market, or simply more experience. A spreadsheet built to catch errors a now-reliable system no longer makes. An extra sign-off added after one bad month that still slows down every order today. Businesses that keep habits adopted under one set of constraints rarely stop to check whether those constraints are still in force.

The same drift shows up in how companies use AI for customer-facing work. A year ago, getting usable marketing copy out of a model took careful prompt engineering and heavy editing. Many teams now get better results simply by asking more directly and trusting the tool with more judgment — provided they resist the pull toward the flattened, interchangeable output flooding what one recent piece calls the generic content era of AI marketing. The skill that mattered twelve months ago, wrestling a weak model into shape, simply is not the skill that matters now.

Staffing decisions follow the same logic. A hiring freeze adopted to survive a difficult year can quietly turn into a permanent ceiling on growth if nobody ever revisits why it was put in place, long after the original danger has passed.

The takeaway: audit your AI habits before they calcify

Hallie’s talk is aimed at engineers tuning Claude Code, not at business owners deciding how to run their companies. But the instruction underneath travels well outside software. Every rule, every safeguard, every extra approval step was a response to a specific weakness at a specific moment. When that weakness disappears, the rule does not announce its own retirement — someone has to go looking for it.

For a small business leaning on AI tools of any kind, that argues for scheduling the second look on a calendar rather than waiting for it to happen by accident. Ask, every few months, which safeguard was built for a problem that the tool, the market or the team has since outgrown. Keep what still earns its place. Drop what has quietly turned into drag.

Claude Code best practices will keep changing as the underlying models keep improving — that is the entire point of Hallie’s talk. The founders who benefit most will not be the ones with the most elaborate setup. They will be the ones willing to simplify it the moment it stops paying for itself.

▶

Sources

• Rethinking Claude Code Best Practices — Lydia Hallie, Anthropic (AI Engineer) — Conference talk published 2026-11-12, ~18 minutes; speaker is a Member of Technical Staff at Anthropic working on developer relations for Claude Code.