Source: raw/37_Cheat_Codes_to_Level_Up_In_Claude_Code_in_19_Minutes.md (https://www.youtube.com/watch?v=RDeofKimDxo, 20:06, uploaded 2026-08-28) · raw/19_Claude_Code_Mistakes_Pro_Users_Are_Still_Making.md (https://www.youtube.com/watch?v=icM0ewXGvAw, 20:12, uploaded 2026-08-22) Creator: Simon Scrapes (YouTube) · Platform: YouTube

Two late-August 2026 list videos from Simon Scrapes: terminal shortcuts, session and context management, CLAUDE.md structure, skill descriptions, and cost traps. The mistakes video also corrects two of his own earlier tips (removing unused MCP servers, and the 95% auto-compact threshold). Feature names follow the auto-captions and his on-screen demos; where the wiki has a first-party page, it is linked. No sponsor; the usual Agentic Academy mentions.

Key Takeaways

Prompts

  • Drop personas and “take a deep breath”. He relays a test of 162 personas across about 2,500 prompts in which a persona did no better than the plain question. Spend the words on where to look, what done looks like, and a self-check (“Before you finish, verify your answer against…”).
  • Say what to do, not what not to do. “Write it as smooth flowing paragraphs” instead of “do not return markdown”.

Context and cost

  • Connectors: load tools when needed. In chat, + → Connectors → tool access → “load tools when needed”, so tool definitions don’t load up front on every message.
  • Correction: MCP servers no longer cost much idle. Tool search is on by default, so only tool names load (about 120 tokens, his estimate) and schemas load on demand. His July advice to remove rarely used servers is out of date; check with /context and /mcp.
  • Routines get every connector by default. Remove the ones a routine doesn’t need, for safety and context.
  • Switching models mid-session can cost more. Each model has its own cache, so moving 100,000 tokens of conversation to Haiku and back re-reads it twice. “Give me a quick answer” on the current model skips deep reasoning for one turn without touching the cache.
  • Fast mode bills to usage credits, not your plan. It runs on Opus at about 2.5 times the speed, stays on across sessions until /fast off, and switching it on mid-conversation pays the uncached fast-mode price for the whole existing context. Without usage credits enabled it switched itself off in his demo.
  • Subagents are good for looking, expensive for doing. He relays the docs: agent teams use about seven times the tokens of a standard session when teammates run in plan mode. A subagent gets its own system prompt, your task text and the CLAUDE.md hierarchy, but not conversation history, output style, auto memory or files already read; the built-in Explore and Plan agents skip CLAUDE.md. Restate critical rules in the subagent prompt, or fork when the side task needs the full context.
  • Correction: auto-compact is no longer a percentage. He had said compaction happens near 95% full. Now, without an autoCompactWindow setting, it compacts at the model’s context limit (with per-model exceptions); set a fixed token count (50K, 100K) instead. He cites Anthropic’s Opus 4.6 retrieval test: 93% at about 256K tokens, 76% at 1M.
  • Partial compaction. Esc twice opens rewind; choosing a message offers “summarize from here” or “summarize up to here”.
  • Resuming a big old session. For a session over an hour old and over 100,000 tokens, Claude Code offers to resume from a summary, because the full resume re-reads everything uncached.
  • Session history is deleted after 30 days by default. Set cleanupPeriodDays (365 for a year) in settings.json; 0 deletes everything. Auto memory files are exempt.
  • Concise output style. /config → output style → Concise leads with results and drops running commentary while keeping errors and warnings. Third-party alternatives he likes: a “simple English” skill modelled on aircraft maintenance English, Stop Slop (16,000 GitHub stars per his figure; scores drafts out of 50 and revises below 35) and Caveman (100,000 stars per his figure; cuts output tokens, his example 69 to 19).
  • “Show clear context on plan accept” (captioned setting name; the exact settings.json key was not checked) set to true adds an option on plan approval to clear the exploration tokens while keeping the plan, which is saved to a file.

CLAUDE.md and rules

  • Edits mid-session are ignored. Root and user CLAUDE.md are read once at session start; restart, /clear or compact to pick up changes.
  • If a rule keeps being ignored, the file is too long. He relays Anthropic’s diagnosis, and figures that Claude Code’s own system prompt is about 50 instructions, that models follow 150 to 200 reliably, and that the best instruction files run 300 to 350 words. Ask /doctor to propose trims; his run suggested savings of 4,000 to 5,000 tokens per session.
  • Path-specific rules. Rules scoped to a path (for example src/api/**) load only when Claude works there. The rules folder follows symlinks, so one shared rules directory can serve many repos.
  • HTML comments are free. <!-- notes --> in CLAUDE.md are not sent to the model; use them for notes meant for people.
  • Hand off with a prompt, not a summary. From Paul’s Programming Notes: ask Claude to write the next session’s prompt (files, what failed, where to start) instead of a human-readable progress summary.

Skills

  • Descriptions need three parts: what it does in one line, triggers (phrases that should activate it), and anti-triggers (“does not trigger for positioning, audience research or keyword work”).
  • Skills can nest up to five levels deep in one prompt.
  • Terminal skills don’t appear in Cowork. Skills in .claude/skills/ had to be re-added through Customize (as a zip, or recreated). Cowork loads skills at session start, so a newly added skill needs a restart.
  • Design. He advises against the built-in front-end design skill (“makes every website look exactly the same”) and recommends building a design system in Claude Design instead.
  • Prompt search: Ctrl+R searches past prompts; per the captions a further key widens the search to the project and then to every project on the machine.
  • Stash a half-written prompt: Ctrl+S, run something else, and the prompt comes back.
  • Ctrl+G opens the prompt in an external editor (and the plan during plan approval, so you can edit it directly); the setting “show last response in external editor” includes Claude’s last reply.
  • Ctrl+T shows the task list and teammates; /tasks lists running tasks. Ctrl+O expands every collapsed tool call; /focus shows only prompts, a result summary and the reply. Ctrl+B sends a running task to the background. Option+T toggles extended thinking.
  • Rewind instead of fixing a broken output. After /clear, rewind can still return to the earlier context. Rewind does not restore files deleted through bash; use Git.
  • Resume across repos: in /resume, Ctrl+A shows sessions from all repositories and Ctrl+W shows worktrees. /rename sessions so you can find them. claude --teleport pulls a web session into the terminal.
  • /permissions mid-task changes allow/ask/auto rules for the running session without restarting.
  • Commands few people use: /doctor (setup check), /insights (an HTML report on up to 200 recent sessions), /btw (side question without interrupting), /branch (copy the conversation to take it another way), /recap (one-line session recap).

Verification

  • Four levels, weakest to strongest: ask for the check in the prompt; /goal; a stop hook that blocks the turn until a script passes; an adversarial review agent. Tell the reviewer to flag only what affects correctness or the stated done condition, or it keeps finding problems.

Desktop app (August 2026)

  • Memory now covers chats and Cowork together, is stored as separate topic files you can read, can be told to skip sensitive topics, and can be imported from other tools (another Claude install, Codex, ChatGPT) with a prompt Claude supplies.
  • Image annotation in the Code tab: circle something and say “remove red circle”.
  • Scheduled tasks inside a project use that project’s context directly.
  • ccusage (third-party) reads local logs from Claude Code, Codex, Gemini, Goose and OpenCode and reports tokens and cost per day and model.

Try It

  1. Add cleanupPeriodDays: 365 to ~/.claude/settings.json today, before more history is deleted.
  2. Set connector tool access to “load tools when needed” and strip connectors from each routine.
  3. Rewrite every skill description as summary + triggers + anti-triggers.
  4. Move notes-to-self in CLAUDE.md into HTML comments and scope API or front-end rules to paths.
  5. End long sessions by asking for a prompt for the next session.
  6. Check /fast is off unless you mean to pay usage-credit rates.

Open Questions

  • The persona study (162 personas) and the “300 to 350 words” figure are relayed without links.
  • The key that widens Ctrl+R search scope is unclear in the captions; check /help or the CLI reference.
  • The seven-times token figure for agent teams is attributed to Anthropic’s docs; the exact page was not checked.
  • Star counts for Stop Slop and Caveman are his on-screen figures at recording time.