Source: raw/Anthropic_Just_Released_the_New_Rules_of_Claude_Code.md Creator: Nate Herk (channel “Nate Herk | AI Automation”, confirmed from YouTube metadata; he says he has no coding background, runs his business from one Claude Code folder, and offers a free guide in a Skool community) · URL: https://www.youtube.com/watch?v=8Pg6G5h05hk · Platform: YouTube · Duration: 9:40 · Uploaded: 2026-11-17 · Transcript fetched: 2026-11-17

A creator’s breakdown of a talk by Lydia Hallie of the Claude Code team (auto-captioned “Lydia Hi”) on how Anthropic works with Claude Code now. The talk’s title, venue and date are not in the transcript, and the wiki holds no first-party copy of it. The creator organises it as five jobs you used to do by hand in every session (steer, approve, say no, check, start) and what handles each one now. Lines the captions mark as a second speaker appear to be clips from the talk and are quoted as hers; everything else is the creator’s relay.

Key Takeaways

  • The framing, in her words: “We’ve now essentially become managers for dozens of sessions,” and “almost everything we used to do ourselves is now either the model, the harness, or the environments.”
  • Scale claim (creator’s relay, no first-party source here): Anthropic went from about 200 code changes a day in May to more than 30,000 a day in September.
  • Steer: prune CLAUDE.md. Her test for every line: “could Claude figure this out by reading the code.” All-caps IMPORTANT and NEVER lines written for older models “can actually make the model go off track,” with more tool calls and worse results. The creator says her daily driver is Opus 5.5.
  • Approve: auto mode replaces click-to-allow. A separate classifier checks commands, so tests and installs run and it stops only on what it is unsure about. You describe what to trust in plain English in settings, and that text is sent to the classifier.
  • Auto mode’s measured catch rate (talk clip): in a study “we published back in August,” about 14% of people caught a clearly dangerous command and auto mode blocked 89%, so “about 1 in 10 of those commands still went through.” This study is not in the wiki’s Auto Mode sources.
  • Say no: deny rules for what cannot be undone. Her test: “if Claude gets this wrong once, can I take it back?” Bad code can be reverted, so leave it to the model. Reading an API key or wiping unsaved work cannot, so it becomes a hard deny rule in settings, which Claude Code checks before auto mode.
  • A deny rule sees the command, not what the command does. A test run executes code you did not write. /sandbox makes the operating system limit which files a command can open and which sites it can reach.
  • Long jobs run in a cloud session. Claude gets a fresh computer with a copy of your code and none of your keys. Her caveat: “This machine is disposable, but the things that leave it aren’t,” such as a pushed commit or a message sent through a connector. Deny rules committed to the project apply there too.
  • Check: give pass/fail checks along the way. Her example is one prompt (move the login system to a new API) that can land three ways: old API left in, new logins only, or everyone logged out. A test or a screenshot catches the bad run at the first check, not at the end.
  • Hooks must exit 2. Her warning: most tools exit 1 on a failed test, and with a 1 the task “just gets marked as completed anyway.” Exit code 2 “actually sends Claude back to fix it.”
  • Claude reviews every PR first, in a fresh context, “not by the session that wrote it,” with a separate group of agents that re-check each bug it finds. People review only what is risky; docs changes merge once checks pass (creator’s relay).
  • Start: Claude Code Projects starts the sessions. One main session takes the week’s list, starts a session per task and checks what comes back. Her model advice for the main session: it “doesn’t write the code itself,” so use “a model that’s really good at deciding what to do and judging what comes back. A model like Fable.”

The five handoffs

Job you used to doWhat handles it nowWhat you still set up
Steer (watch, correct, remind)The modelA CLAUDE.md holding only what Claude cannot see in the code
Approve each commandAuto mode’s classifierPlain-English trust entries: repos, buckets, domains, always-allow, never-do
Say noDeny rules, then a sandboxA deny list for irreversible actions, committed to the project
Check the resultChecks, hooks, CI, Claude code reviewPass/fail checks per task; hooks that exit 2
Start every sessionClaude Code Projects (beta)A task list, and a coordinator model

The table is this article’s arrangement of the creator’s five sections.

Details by handoff

Steer

  • The creator’s own CLAUDE.md is a routing map: where the business wiki, brand assets and finances live. That stays because Claude cannot guess it from the files. The same role can be played by AGENTS.md (see AGENTS.md).
  • /doctor prompt-audit checks CLAUDE.md and skills against the model you are on and reports what the model no longer needs. This matches the v2.1.283 command recorded in Week 39.

Approve

  • The creator says auto mode is the default in the terminal and the Claude Code desktop app, and by default trusts only the current folder and current repo. His example: asked to ship to staging, it blocks the push to a separate deploy repo until you approve or add that repo.
  • He says a setup command (auto-captioned “/auto mode setup”) drafts the trust entries from your project and recent sessions. The exact command name is not confirmed against the docs.

Say no

  • “Never read my .env file” belongs in permissions, not in a prompt. The creator stresses the difference for client work.
  • For knowledge work the test is the same: rewriting a draft is reversible; sending an email or deleting files is not.
  • His supply-chain example: a compromised script inside your test run could open the .env file or SSH keys and send them to an unknown server. A human approving the test command would have clicked allow, because the command itself looks harmless.

Check

  • Checks do not have to be code. For a research report: every claim has a source link, every number matches the spreadsheet. A plain description of what done or good looks like also counts (creator).
  • Her habit: “I also like to ask it to show the actual output because you weren’t there to see it.”
  • The creator’s hook cadence: a format check after every edit, tests after every task, the full suite before Claude is allowed to finish. He also uses /goal, which keeps Claude working until a finish condition is true (see goal walkthrough).

Start

  • The creator says Claude Code Projects is in beta on Pro, Max and Ultra plans. The wiki’s first-party walkthrough (Claude Code Projects) says Pro and Max; Ultra is his addition.
  • He says project settings let you pick the coordinator model, and suggests Fable 5.2 as coordinator with Opus 5.5 for the threads. The Opus 5.5 pairing is his, not a quote from the talk.
  • The list does not have to be code: a client report, a research brief, or a batch of content ideas with scripts (creator).

How this compares with what the wiki already holds

  • Exit codes agree with the hooks reference. Hooks records exit 2 as the blocking signal and any other non-zero exit as non-blocking.
  • Deny before classifier agrees with the settings docs. Auto Mode records autoMode.hard_deny as blocking “unconditionally even under broader allow rules.”
  • The coordinator advice differs in emphasis from Anthropic’s Projects video. That video’s cost tip is to leave the coordinator on low effort and default threads to a smaller model. The talk clip asks for a strong judging model (Fable) as coordinator. One is about effort and the other about model, so both can hold (Fable on low effort), but they pull in different directions on cost.
  • The CLAUDE.md test is the same rule as Thariq Shihipar’s. The New Rules of Context Engineering says to avoid stating anything Claude can learn from the repo.
  • The 14% and 89% figures are new to the wiki. The earlier containment research recorded a 93% approval rate for permission prompts (see Agent Guardrails); this is a different measurement.

Try It

  1. Run /doctor prompt-audit, then delete each CLAUDE.md line that fails the test “could Claude figure this out by reading the code.”
  2. List the actions in your work that cannot be taken back (reading .env, deleting files, sending email, pushing to a deploy repo). Put each in a deny rule in the project’s settings file and commit it.
  3. In auto mode, write your trusted repos, buckets and domains in plain English in settings.
  4. Run /sandbox before any job that executes third-party code, such as installs and test suites.
  5. Check every hook script: on failure it must exit 2, not 1.
  6. Pick one task bigger than you would normally hand over, write a pass/fail check that would catch the worst outcome, start it in a cloud session, and note where you are still in the loop. That last step is Lydia Hallie’s closing advice.

Open Questions

  • The talk itself. Title, event, date and a recording or transcript are not in this source. Until a first-party copy is ingested, every number here rests on the creator’s edit.
  • The August study. Which Anthropic publication reports 14% human catch rate against 89% for auto mode, and how “clearly dangerous” was defined.
  • 200 to 30,000 code changes a day. No first-party source in the wiki, and “code changes” is not defined (commits, PRs or edits).
  • The auto mode setup command’s exact name, and whether auto mode is now the default on the desktop app for every plan.
  • Ultra availability for Claude Code Projects.