19 Claude Code mistakes

19 Claude Code Mistakes "Pro" Users Are Still Making

🎬 Simon Scrapes 📅 Aug 22, 2026 ⏱ 20:12
Claude Code prompting context compaction verification

✍️ Prompting (Mistakes 1–2)

Mistake 1 — personas are dead weight. Someone tested 162 personas across 2,500 prompts: "prompts with the persona did no better than just asking the question directly." Anthropic confirmed from the inside — they cut their own system prompt ~80% and now say stuffing it with examples is no longer best practice. 0:00

The replacement — three things instead: (1) where to look, (2) the definition of done, (3) a self-check ("before you finish, verify your answer against…"). "Same word count, but now every word is actually used by Claude." 0:53

Mistake 2 — stop writing "do NOT do X." Negative instructions confuse the model when they conflict with the actual task. "Instead of 'do not return markdown,' say 'write it as smooth flowing text paragraphs' — tell it what to do, not what not to do." 2:01

🔧 Tools & Connectors (Mistakes 3–6)

#MistakeThe fix
3Routines run with full tool access by default (unsafe + costly)Remove/redact connector access per routine
4Connectors load tool definitions up-front on every message (thousands of tokens)Set tool access to "load tools when needed"
5Removing MCP servers to save context (outdated advice)Tool search is on by default — only names load (~120 tokens), schemas on demand. Check with /context
6Skills don't coexist between terminal & co-workRe-add as zip in the customize window; co-work loads skills at session start — restart after enabling
MCPs do NOT eat context anymore: "There used to be advice about removing infrequent MCP servers because schemas loaded every session. That's out of date — only tool names load in (~120 tokens), schemas pull on demand. Adding more MCP servers now has minimal impact." 3:27

👥 Subagents (Mistakes 7, 11)

Mistake 7 — subagents cost ~7× more than you think. Anthropic's own docs: agent teams use ~7× more tokens than standard sessions in plan mode, because each teammate maintains its own context window as a separate Claude instance. 5:17

The routing rule: "Sub-agents are brilliant for looking at things — broad searches, parallel investigation where little context is required. They're expensive for doing things. If a task needs high context, keep it in the main session — you'd be passing uncached context into a separate window on the same model." 6:03

Mistake 11 — subagents skip your Claude.md. A subagent gets its own system prompt, your task text, and "kind of" the Claude.md hierarchy — but no conversation history, no output style, no auto-memory, no already-read files. And the built-in explore/plan agents skip Claude.md entirely. "If a subagent needs certain rules, restate them in the subagent's instructions directly. Or fork the conversation instead — a fork inherits the full context." 10:37

⚡ Fast Mode (Mistake 8)

Fast mode is billed to API credits, not your subscription. If usage credits are on, fast mode "will cost you a lot of money and stay on by default until you type fast off — even across sessions." 6:39

The dangerous part: "The first time you enable fast mode in a conversation, you pay the full uncached fast-mode price for the entire existing context. You switch to fast mode and suddenly all that cached data becomes uncached — you pay the full fast-mode uncached input price for everything." Flipping it mid-conversation can be "a really costly move." 7:16

📄 Claude.md (Mistakes 9–10)

Mistake 9 — don't edit Claude.md mid-session. "The project-root and user-level Claude.md files are read once at session start and held in memory. Until you clear or start a new session, it ignores any edits." New rules need a restart, clear, or compact. 8:51

Mistake 10 — the real impact of a long Claude.md: "If Claude keeps doing something you don't want despite a rule against it, the file is probably too long and the rule is getting lost." The numbers: models reliably follow ~150–200 instructions; the instruction files that perform best sit around 300–350 words (words, not instructions). Run /doctor to propose trims — "it saved 4,000–5,000 tokens every session." 9:32

🧠 Context & Compaction (Mistakes 12–14)

Mistake 12 — stop compacting at 95%. The old "compacts at 95% full" behavior is gone. Now: no autocompact limit → compacts at the model's context limit, with per-model exceptions. But you can set a fixed autocompact window (e.g. 100k tokens) — "quality degrades long before the window is full." 11:57

Mistake 13 — bigger context windows are worse. Anthropic's long-context retrieval test: Opus 4.6 scored 93% at 256k tokens, but dropped to 76% at 1M tokens — "one in four rules not retrievable. A bigger context window isn't more memory; it's harder to retrieve from." 13:08

Mistake 14 — compaction isn't all-or-nothing. Press escape twice for "rewind mode" — you can restore code/conversation, or summarize up to a specific point (autocompact everything before it) and continue from there. 14:05

📋 Plan Editing (Mistake 15)

Ctrl+G edits the plan directly. In plan mode, instead of asking the model to rework the plan with more back-and-forth tokens, hit Ctrl+G to open the plan in the editor, edit it yourself, then tell it "updated the plan." 14:48

💰 Cheap Models (Mistake 16)

Cheaper models can be more expensive. "Every model has its own cache. If you're 100k tokens into a conversation with Opus and switch to Haiku for an easy question, you rebuild Haiku's cache with 100k tokens — and switching back rebuilds Opus's cache again. The moment you switch, it gets uncached and you pay for it all again." 15:42

This is the exact same principle as the GLM 5.3 "late switch costs more" video — a model switch mid-conversation invalidates the prompt cache. 15:42

✅ Verification (Mistake 17)

"Anthropic's exact words: the single most impactful tip is verification. If there's no check Claude can run itself, then you're acting as the verification loop for Claude — every error sits there waiting for a human to notice." 16:28

LevelVerification
1Ask for the check in the same prompt
2/goal — a separate evaluator rechecks a condition every turn
3Stop hooks — physically block the turn from ending until the script passes
4Adversarial review agent — finds gaps (but tell it to flag only correctness/requirements, else it never stops)

🗂️ History & Commands (Mistakes 18–19)

Mistake 18 — history deletes itself after 30 days. By default Claude Code only keeps 30 days of conversations. Fix in settings.json: cleanupPeriodDays: 365. (Warning: 0 does NOT mean unlimited — it wipes everything as it arrives. Auto-memory files are exempt from the sweep.) 17:56

Mistake 19 — five commands nobody types: 18:52

CommandWhat it does
/doctorFull setup checkup — finds unused skills/MCPs, dedupes Claude.md, proposes trims
/insightsHTML report analyzing up to 200 recent sessions
/btwQuick side question without interrupting the main conversation
/branchCopy the conversation at this point to a new direction (original stays intact)

✅ Key Takeaways

  1. Personas are dead. Tested across 2,500 prompts, they add nothing. Replace with where-to-look + definition-of-done + self-check.
  2. Say what to do, not what not to do. Negative instructions confuse the model.
  3. MCP servers barely cost context anymore (tool search loads schemas on demand) — but connectors in "auto" mode still bleed tokens up front.
  4. Subagents are for looking, not doing. ~7× token cost for context-heavy delegation; keep high-context work in the main thread.
  5. Fast mode is an API-credit trap — flipping it mid-conversation uncaches everything and costs a fortune.
  6. Keep Claude.md ~300–350 words. Rules in bloated files get lost; models follow ~150–200 instructions reliably.
  7. Verification beats vibes. Stop being Claude's verification loop — use prompt checks, /goal, stop hooks, or an adversarial reviewer.

🔗 Resources & Links

📍 Timestamp Index

0:00 Mistake 1 — personas
2:01 Mistake 2 — "do not"
2:20 Mistakes 3–4 — routines/connectors
3:27 Mistake 5 — MCPs
4:18 Mistake 6 — skills
5:17 Mistake 7 — subagent cost
6:39 Mistake 8 — fast mode
8:51 Mistakes 9–10 — Claude.md
10:37 Mistake 11 — subagents skip Claude.md
11:57 Mistakes 12–13 — compaction/context
14:05 Mistake 14 — partial compaction
14:48 Mistake 15 — Ctrl+G plan edit
15:42 Mistake 16 — cheap models
16:28 Mistake 17 — verification
17:56 Mistake 18 — history
18:52 Mistake 19 — hidden commands
☰ View all