How to Actually Use Claude Code Well (September 2026 Update)
Back in July I read two pieces on how to use Claude Code well: one a feature-by-feature overview from Anthropic engineer Shrivu Shankar, the other a hands-on beginner’s guide from Eyad Khrais. Two months on, Claude Code has shipped dozens of version bumps, and some of their specific advice no longer quite holds. This post revisits the core principles from both, then adds what’s actually changed over the past six months.
What hasn’t changed: the core principles still hold
Think first, then act. Both pieces hammer the same point: going into Plan mode to work out the architecture beats handing the agent a vague idea and hoping for the best. That judgment hasn’t aged - model capability has gone up, but “unclear goals” still produces “overbuilt implementations.”
CLAUDE.md is a constitution, not a manual. Only write the project’s quirks, real workflows, and places the agent tends to screw up - don’t expect piling on docs to make the agent smarter. This advice deserves more attention now, because there are more context-management tools available, which makes it easier to get lazy about trimming CLAUDE.md and just assume “there’s room for it anyway.”
Wrapping a complex CLI in a thin layer and only documenting the wrapper itself - rather than writing long docs explaining the complex commands - is still just as true six months later, and remains a good test for “is this still worth maintaining this way.”
What changed: six months’ worth of shifts
The context window isn’t a tight 200k budget anymore. Sonnet 5 and Opus 4.6 now default to a 1M token context, and the old anxiety around /compact - “this could blow away half your context at any moment” - has clearly eased. But a bigger window doesn’t mean you should cram more into it: “write it down, then /clear and restart” is still more reliable than grinding through one long session for complex tasks - the urgency to do that has just relaxed a bit.
The concerns about custom subagents still stand, but Skills are now a real, productized feature. Both pieces pointed toward “use scripts instead of fiddly custom subagents/MCP servers,” and now Anthropic has paved that exact path officially: Skills has a marketplace and a skill-creator, and Plugins can bundle skills, hooks, MCP, and LSP into a single command distributed to a whole team, instead of everyone hand-rolling their own setup. The earlier call to “hand-roll scripted agents” turned out to be the right direction - there’s now a more convenient official version of it.
Claude Code SDK was renamed Claude Agent SDK. This September 2025 rename wasn’t just a logo swap: the docs moved out of the Claude Code section into their own Agent SDK branch, signaling a shift from “an SDK for writing code” to a general-purpose agent framework - consistent with use cases like batch scripts, internal chat tools, and rapid prototyping; it’s just that the official positioning now confirms that direction.
Remote takeover and permission checks got a lot more polished. You can now take over a session running locally straight from your phone or browser, branch off a remote session and have it keep running locally in the background, and on desktop, pop a panel out into its own window. The idea of gating commits with a hook hasn’t changed, but the triggers and signals available to hooks have grown quite a bit since six months ago.
The core call on GitHub Actions hasn’t changed, but there’s now more to grab onto for quality checks. Operationalizing Claude Code with GHA is still an underrated pattern, and now there’s claude plugin eval, which lets a plugin run test cases and get scored - giving teams a more standardized way to audit “is what the agent produced actually any good,” instead of relying purely on manually reading logs.
Updated checklist for myself
Keep writing CLAUDE.md as “guardrails, not a manual” - but go back and prune it periodically. Entries added six months ago to work around old-version bugs may not be needed anymore.
When you keep needing to integrate some third-party capability, check for an existing Plugin first, instead of hand-rolling an MCP server or subagent from scratch.
No need to over-/compact just to save context, but “write a plan doc, then /clear” is still worth doing for complex tasks - that habit matters more than window size.
Worth spending ten minutes looking at the Skills marketplace and skill-creator to figure out which of your own throwaway scripts could be upgraded into a proper, reusable Skill.
Reference
How I Use Every Claude Code Feature - Shrivu Shankar
The complete claude code tutorial - eyad_khrais on X