Monday, 7 September 2026
A rundown of 19 Claude Code habits that are now outdated or costly, most of them sourced from Anthropic's own docs and recent product changes. The big themes: persona prompts don't help (drop them for location, done-criteria and self-check instructions), context is cheaper to manage than you think if you configure connectors, auto-compact and CLAUDE.md correctly, and several "savings" moves like switching to a cheaper model mid-conversation or leaning on sub-agents actually cost more. It closes with a set of little-used slash commands: /doctor, /insights, /btw and /branch.
- A study of 162 personas across ~2,500 prompts found persona openings performed no better than asking directly, and the Claude Code team recently cut its own system prompt by roughly 80%; better to spend those words on where to look, what "done" looks like, and a self-check line.
- Write instructions positively — "write it as smooth flowing text paragraphs" rather than "do not return markdown" — because negative instructions conflict with the actual task.
- Connectors are added to new routines by default and tool definitions load up front unless you set tool access to "load tools when needed," which people have measured as costing thousands of tokens per session; conversely the old advice to delete rarely-used MCP servers is obsolete since tool search means only ~120 tokens of tool names load, with schemas pulled on demand.
- Sub-agents cost roughly seven times more tokens than standard sessions because each teammate runs its own uncached context, so they suit broad parallel searching but not context-heavy delegation — fork the conversation instead when full context is needed.
- CLAUDE.md files are read once at session start and held in memory, so mid-session edits are ignored until you clear, restart or compact; shorter files work better (Anthropic's guidance points to ~300–350 words, and Claude Code's own system prompt is about 50 instructions), and /doctor can propose trims.
- Anthropic's long-context benchmark on Opus 4.6 scored 93% retrieval at 256K tokens but dropped to 76% at 1 million tokens, meaning roughly one in four buried rules fails to be retrieved — bigger context is not more memory.
- Switching to a cheaper model mid-conversation can cost more than staying put, because each model has its own prompt cache and switching forces you to rebuild it in both directions; the same uncached-context trap applies to enabling fast mode mid-conversation.
- Anthropic calls verification the single most impactful practice, with four escalating levels: an in-prompt check, /goal, stop hooks that block the turn, and an adversarial review agent (which must be told to flag only correctness or stated requirements, or it will invent problems).
- Claude Code keeps only 30 days of resumable conversations by default — set cleanupPeriodDays in settings.json (0 wipes everything, it doesn't mean unlimited), though auto memory files are exempt from the sweep.
It kills the single most common prompt-writing habit with a study plus Anthropic's own 80% system prompt cut, and gives the three-part replacement.
They tested 162 different personas across 2 and 1/2 thousand prompts, and the prompts with a persona did no better than just asking the question directly.
A counterintuitive cost mechanic — per-model prompt caches — that reverses the obvious optimisation most users make.
If you're 100,000 tokens into a conversation with Opus and you want to answer a question that's fairly easy to answer, it would actually be more expensive to switch to Haiku at that point than to have Opus answer.
If Claude keeps doing something you don't want despite having a rule against it in your CLAUDE.md, the file is probably too long and the rule is getting lost.
The single most impactful tip in this guide is verification... if there's no check, you're basically acting as the verification loop for Claude.
A bigger context window isn't really more memory. It's basically then going to become harder to actually retrieve the right information.
Dylan explains why the old advice to "start fresh chats often" has partly changed: compaction (auto-summarizing long chats) and native memory in Claude and ChatGPT now make long conversations viable, but mainly in desktop agents like Claude Cowork and Codex rather than in the browser. He lays out two setups: Setup A, fresh chats in a task-dedicated folder with an instructions file (claude.md / agents.md) plus an externalized memory file, which covers ~95% of use cases; and Setup B, a pinned thread that runs for weeks or months, only for tasks with no finish line where yesterday's context helps today's work (inbox assistant, long-running project monitoring). Both setups rely on keeping memory files short (150-200 lines), pruning them periodically with approval, and, for pinned threads, generating a one-to-two-page handoff document when intelligence degrades and it's time to start a new thread.
- Two recent improvements enable longer chats: compaction, where the AI silently summarizes earlier conversation to itself, and native memory, which still only stores surface-level facts like name, location, job, and broad preferences.
- The benefits are unevenly distributed — in the browser you won't see much gain; compaction and memory work much better in desktop agents (Claude Cowork for Claude, Codex for ChatGPT), which are included with existing subscriptions.
- Setup A (recommended for ~95% of work): a folder per finite task with an instructions file that defines the purpose, points to a memory file, and tells the AI to append short dated lines whenever it's corrected — giving both fresh-chat intelligence and accumulated preferences.
- Setup B (pinned long thread) only passes two tests: the task has no finish line, and today's work benefits from knowing what happened yesterday. If you're hesitant, default to Setup A.
- Best Setup B use cases are an inbox assistant that triages, researches, and drafts replies (the one Dylan says OpenAI people discuss most) and monitoring a 6-12 month project with recurring updates.
- Don't build a thread around a client — a client is a topic, not a task. Use a parent client folder with subfolders for each task (proposals, contracts, questions).
- Even a well-run pinned thread eventually degrades — the AI forgets obvious facts and contradicts itself. Ask it for a one-to-two-page handoff document, then start a new chat in the same folder so the memory and loops files carry over.
- Memory files must stay dense and under roughly 150-200 lines; run a weekly or monthly audit where the AI flags stale, repeated, or untrue lines, moves them to an archive folder rather than deleting, and shows you changes for approval before acting.
This is the decision rule the whole video hinges on, stated compactly enough to act on immediately.
Is there a finish line? If there isn't a finish line, it might be good for this setup B. Second question, is it beneficial to understand what happened yesterday in the thread for today's work?
It names the most common structuring mistake and gives the concrete parent-folder/subfolder fix.
"Oh, I have this huge client that I want to have an ongoing thread for." That's a bad idea because for a given client, that's a topic not a task.
Its intelligence dropped like a rock after around 50% of filling up its head.
For a given client, that's a topic not a task.
Every lesson worth keeping starts as the decision your AI made without you asking.