Source: raw/How_the_Top_1%_Use_Claude_Code_Differently.md (https://www.youtube.com/watch?v=J9NQTYUP6k0, 18:09, uploaded 2026-10-10) · raw/Claude_Changed_Completely_-_Every_NEW_Concept_Explained.md (https://www.youtube.com/watch?v=YAuWBqZ_gjs, 19:15, uploaded 2026-10-17) · raw/5_Claude_Myths_Even_Pros_Fall_For.md (https://www.youtube.com/watch?v=v9YVQoeSp-I, 11:21, uploaded 2026-10-22) · raw/Watch_This_If_You_re_Serious_About_Claude_Code_Most_People_Skip_This_Step.md (https://www.youtube.com/watch?v=CBcyKfpldTo, 14:54, uploaded 2026-06-25) Creator: Simon Scrapes (YouTube) · Platform: YouTube

Simon Scrapes’ October 2026 trilogy argues that most of the Claude Code advice from early 2026 (start in plan mode, long CLAUDE.md, step-by-step skills, hand-built agent teams, cheap models for helpers) was written for weaker models and now costs tokens or quality on Opus 5.5. What replaces it is a short brief with a definition of done, a check Claude can run, and a loop that keeps going until the check passes. A June 2026 video from the same channel supplies the earlier “four limitations” framing. The three October videos follow the order of his new paid Agentic Academy course, which they promote.

Key Takeaways

  • Plan mode: the mode is optional, the planning is not. Boris Cherny said in January that most of his sessions started in plan mode, and later on Hacker News that “plan mode was useful and is no longer useful”. On 2026-09-23 Tarik on the Claude Code team posted that the team was thinking of removing it and using Shift+Tab for effort levels instead (about 3,000 replies). The most-liked reply: “Plan mode isn’t for the agent, it’s for me.” Simon’s position: skip the mode, keep the planning in the brief. Plan mode’s one remaining function is blocking edits until you approve.
  • The four-line brief. Say what done looks like, how it will be checked, what is out of scope, and what it must not touch. Example: add an ad-spend section after the leads section to match last month’s layout; totals must match the October sheet; don’t change the summary page; stay out of the client’s shared folder. If you can’t answer the four lines, ask Claude to interview you first, or ask for a one-line plan.
  • Check the result, not the plan. For big changes Boris asks for a diagram or an interactive demo of what changed. Simon asks for a one-page visual of what Claude did and assumed.
  • Verification is the biggest lever. Boris’s January post: a way to verify its work gives two to three times the quality. Don’t write “double-check your work” (he says the Opus 5 guide states the model already does); give a pass/fail criterion instead, ideally a script such as check_report that compares every figure to the export.
  • /goal needs evidence in the chat. After each turn a small model (Haiku, per Simon) reads the conversation and decides whether the goal is met. It does not open files or run commands, so “the report is finished” ends early. His template: /goal the check_report script passes and you show me the output; the template is unchanged; stop after 15 turns.
  • CLAUDE.md is a contents page. Keep four things in it: tools and commands Claude can’t guess, where you differ from the default, gotchas it got wrong more than once, and boundaries. Everything else is a pointer. Pointer rules: write paths in backticks with when to open them, not @ imports (an import loads the whole file every session); link every file directly from CLAUDE.md; give any file over 100 lines a contents list, because nested files may be previewed only to line 100. /doctor prompt-audit flags instructions written for older models and dead paths and changes nothing until you approve.
  • The deletion debate, settled both ways. He relays that Anthropic cut over 80% of Claude Code’s own system prompt for the Claude 5 models with no measurable loss, that the docs say to keep CLAUDE.md under 200 lines, and a study finding tool-written context files made agents slightly worse and about 20% costlier. Against deleting everything: you lose rules Claude cannot infer (a team decision, a client who hates certain phrases). Test each line with Anthropic’s question: would removing this cause mistakes? Claude Code reads AGENTS.md when there is no CLAUDE.md.
  • Skills are corrections, not manuals. Run the job without a skill first, write down only what Claude got wrong, and put that in the skill. He relays a test of nine popular skill packs, 450 runs on Opus 5.5 against a same-length dummy file: two of nine beat the dummy, and Superpowers was not one of them. Use claude plugin eval (skill vs no skill) and /skill-doctor (context cost and usage) to prune. Full rules: Skill Authoring Best Practices.
  • Rules in capitals become hooks or get deleted. Anything Claude does anyway: delete. Anything that must never happen (“never email the client”): a hook. Hooks can now be a prompt (reads the conversation, cheap) or an agent (can open files, marked experimental); a stop hook can block “done” until a second model approves.
  • Let Claude size the team. Codex’s Eric Provencher argued in September that more than two agents usually burns tokens for a “coordination tax”; three weeks later Anthropic’s hosted platform ran up to 8,000 agents on one job (one agent found 14 to 27 of 70 planted bugs; the large run found 66). Both sides say start small. Drop fixed role agents (“you’re a senior copywriter”); keep agents that carry a checklist or a restricted tool set; cap the dynamic workflow size at small in /config.
  • Subagents vs forks. A subagent starts fresh and returns only a summary; the built-in Explore and Plan agents don’t read CLAUDE.md. A fork copies the conversation and shares the cache. He says fork mode has been on by default since August and Claude picks between them.
  • Pick a model at the start and keep it. Each model has its own prompt cache, so switching mid-session re-reads the whole conversation at full price, twice if you switch back. Haiku 5.5 (released 7 October, about 40 times cheaper per token than Opus 5.5) scores 39% on Terminal Bench against Sonnet 5.5’s 71% per his figures, and one Reddit user returned to Opus after two to three fix rounds per task. Lower the effort or say “give me a quick answer” instead. /advisor opus from a Sonnet session lets the cheaper model consult the stronger one without breaking the cache (experimental).
  • Leave the defaults alone. Opus 5.5 is the default on paid plans; he says Opus 5.5 on medium matches or beats Opus 5 on high and max, and that the docs call it prone to overthinking. Auto mode (a classifier on each action) has been the default since August, and compaction points are tuned per model.

Running without you

  • Routines run in the cloud and can start from a schedule, an API call (you get a URL and a token) or a GitHub event. New routines get every connector by default; strip them to what the job needs.
  • Cloud sessions give each task its own machine on your normal plan: start from claude.ai/code, the desktop app (pick cloud), or claude --cloud; pull one back with claude --teleport.
  • Limits no longer kill a task. He says a task that hits the 5-hour limit now waits and resumes when it resets (on by default), spending your next window without asking. It does not do this for the weekly limit.
  • Projects are one coordinating conversation that starts cloud threads; see Projects.
  • Human checkpoints stay (June video). Full autonomy is not the goal. Put a checkpoint where a mistake is costly; his carousel system stops after the slide plan and again before posting (carousel system).

The June framing: four limitations of a fresh install

The June 2026 video names four limits and their fixes, which the October videos build on:

  1. It doesn’t know you. A brand context folder (voice profile, visual identity, positioning and ICP) that CLAUDE.md points to. See Brand Context Files.
  2. It stops when you leave. Auto mode (Shift+Tab), /goal with a measurable condition and an exit limit, /loop for intervals under three days, routines for longer schedules, and Channels plus tmux on a VPS for an always-on machine.
  3. It degrades on big jobs. Raise effort for hard tasks; for the largest, dynamic workflows (ultracode) write their own agent team using patterns such as adversarial verification and loop-until-done.
  4. Poor inputs scale mistakes. Human checkpoints at the costly steps.

Try It

  1. Rewrite your next task as the four-line brief before you hit enter.
  2. Turn the definition-of-done line into a script or checklist Claude can run, then use it in a /goal with a turn limit.
  3. Run /doctor prompt-audit, then cut CLAUDE.md to the four kinds of line above plus pointers.
  4. Run claude plugin eval on your three most-used skills and delete any that don’t beat no-skill.
  5. Replace each capitalised “NEVER” rule with a hook, or delete it.
  6. Stop switching models mid-session for easy questions; ask for a quick answer instead.

Open Questions

  • The nine-skill-pack test, the “context files 20% costlier” study, and the Haiku 5.5 Terminal Bench figures are relayed without links; not checked here.
  • The 8,000-agent hosted run and the 70-planted-bug result are relayed from Anthropic without a cited post.
  • “Fork mode on by default since August” and “tasks wait for the 5-hour reset” were not checked against the changelog in CLI reference.
  • Whether plan mode has actually been removed after Tarik’s post is not stated in these sources.