Updates — November 2026
What changed in Anthropic's own guidance for working with Claude (Claude 5 generation: Opus 5.5, Fable 5.2, Sonnet 5.5, Haiku 5.5), and what it means for researchers using this book.
Every point below was checked against Anthropic's documentation or engineering blog in November 2026 — quotes are verbatim, sources at the bottom. Popular claims that could not be verified (circulating in YouTube videos and posts) are deliberately left out.
The short version: less ceremony, more checking. Write plainly, give Claude a way to test its own work, pick the effort level per task, keep CLAUDE.md short.
1. Prompts: plain request + what Claude can't see
The elaborate prompt template — role ("You are a world-class historian with 20 years of experience…"), goal, numbered steps — is no longer what Anthropic recommends. The Claude Code best-practices page opens with: "in most cases, your first plain prompt in a fresh session already does the right thing."
What still matters is information Claude cannot get by reading your files: the purpose of the task, the audience, a convention your team agreed on, a decision that lives only in your head.
Old: You are an expert archivist. Your task is to… Step 1… Step 2… New: Turn these 12 interview summaries into a comparison table for the project meeting on Friday. The team only cares about the three villages in the pilot; ignore the rest.
On examples — a nuance. Anthropic's July 2026 post warns that for its newest models "giving examples actually constrains them to a certain exploration space." That is said about standing instructions (system prompts, skills, CLAUDE.md). For a one-off task where you need a specific citation style or tone, a sample of what you want is still the clearest specification — but don't bake examples into your permanent instructions by default.
2. Not sure what you want? Let Claude interview you
You don't have to have the task fully worked out before prompting. Anthropic suggests the reverse:
"I want to add rate limiting to the public API but haven't thought it through yet. Interview me until you could write the spec, then write it to SPEC.md."
Research translation: "I want to code these court records for kinship relations but haven't settled the categories. Interview me until you could write a codebook, then write it to codebook.md." Then start a fresh session to do the coding from that file — clean context, written definition of done.
→ A.issue.plan-mode, B.lifecycle.0.iteration
3. Give Claude a check that comes back pass or fail
This is the single most emphasised point in the current guidance:
"Give Claude something it can run that comes back pass or fail. Once it has that, it can keep going on its own and you don't have to watch."
For programmers this means tests. For researchers it means yes/no criteria stated before the work starts, not "check your work" tacked on at the end. "Is it scholarly?" is not checkable. These are:
-
Does every quotation appear verbatim in the source file?
-
Does every person in the table have a
source_refthat exists insources.csv? -
Is every abstract under 250 words?
-
Does every date fall within 1550–1650?
Then: "Before you show me the result, go through this list and report pass or fail on each line, with the evidence."
→ C.calibration, A.critical.limitations
4. Review with fresh eyes — a new session, not the same one
A reviewer that watched the work being done is swayed by the reasoning that produced it. Anthropic: "A second session is also the easiest independent reviewer you can get." Open a new session, give it the output and the original brief (or codebook, or spec), and ask what doesn't meet it.
One caution from the same page: "A reviewer that's asked to find problems will always find something." Tell it what to skip (e.g. "ignore style, report only factual and coding errors") or you will end up chasing trivia.
5. Effort level is now the main dial — choose it per task
Effort controls how long Claude thinks and how much it checks before answering. It replaces most of what people used to do with prompt wording. On the current models in Claude Code (Opus 5.5, Fable 5.2, Sonnet 5.5, Haiku 5.5) the default is medium, which Anthropic calls "plenty for everyday … work with a clear scope." Change it with /effort in the session.
| Level | Use when | Research example |
|---|---|---|
low |
You go back and forth quickly and review each result anyway | Brainstorm 15 possible titles; rough first outline |
medium (default) |
Clear, well-scoped everyday work | Summarise these five articles into the template |
high / xhigh |
Getting it right depends on easily overlooked cases; Claude "checks more edge cases and verifies more of its work" | Reconcile two spreadsheets where every number must match the source |
xhigh |
Long-running jobs (Anthropic: agentic tasks over ~30 minutes) | Process 300 letters overnight against a codebook |
max |
"A hard problem you want Claude to work through without you" | — rarely; "on ordinary tasks it tends to overthink" |
Corollary: phrases like "think hard" or "be thorough" are, per Anthropic, "ordinary prompt text … rather than keywords that raise it." Use the effort setting instead. ultrathink in a single message still works for one hard question in otherwise routine work.
→ A.issue.model-selection, A.issue.costs-quotas
6. Drop the IMPORTANT and ALL CAPS
From Anthropic's list of advice you can stop following:
"Older models sometimes needed the emphasis. Current ones follow a plainly worded line, and piling on emphasis mostly makes Claude over-apply the rule."
If Claude keeps ignoring a line in your CLAUDE.md, emphasis won't fix it. Either the file is too long and the line gets lost — shorten the file — or it's a rule that must never be broken, which belongs in a technical block (a hook or a deny rule), not in prose.
7. Anthropic cut over 80% of its own instructions — simplify yours
In July 2026 Anthropic reported: "We removed over 80% of Claude Code's system prompt for models like Claude Opus 5 and Claude Fable 5 with no measurable loss on our coding evaluations" (Thariq Shihipar, 24 July 2026). Their diagnosis: they had been over-constraining the model through the system prompt, CLAUDE.md files, and skills.
Two caveats: this was Claude Code's own system prompt, not users' files; and the post published no evaluation data, so the figure is Anthropic's self-report.
What they recommend for your CLAUDE.md:
-
"Keep your CLAUDE.md lightweight and briefly describe what your repo is for, but spend most of the tokens on gotchas" — for researchers: the non-obvious things (this archive numbers folios differently; column
dateis Julian calendar; never edit files inraw/). -
"Avoid stating 'the obvious' things Claude should know by looking at your file system."
-
Keep each CLAUDE.md under ~200 lines (memory docs).
Practical step: ask Claude to read your CLAUDE.md and propose a shorter version that keeps only what it couldn't figure out from the files themselves. Review the cut before accepting it.
→ A9.markdown-project-memory, A.markdown.token-efficient-docs
8. Memory is automatic now
Claude Code has auto memory, on by default in local sessions: notes Claude writes itself from your corrections and preferences, loaded at the start of every session alongside CLAUDE.md. Anthropic: "Claude now automatically saves memories that are relevant to the work and to you." The old # shortcut for writing to CLAUDE.md is no longer the recommended route.
Division of labour: CLAUDE.md = what you decide (project facts, rules); auto memory = what Claude learns about how you like things done. You can view and toggle it in /memory.
→ A.concept.memory, A.issue.research-data-memory
9. Context and cost: the window is big, but you still pay for it
Running out of context is rarely the problem anymore — most current models have a 1M-token window, and Claude Code compacts the conversation automatically near the top. Anthropic now lists "manage the context window aggressively" as outdated advice.
What has not changed: every request re-sends the whole conversation, so a long session costs more per turn. That is discounted while the prompt cache is warm and full price once it expires.
-
Cache lifetime: about one hour on a subscription (Pro/Max/Team), five minutes on an API key.
-
Back from lunch? On the one-hour cache, an idle session past ~200K tokens is compacted shortly before the cache expires, so you continue from a summary. If you don't need the history, a fresh session is cheaper.
-
/clearbetween unrelated tasks still the best habit — "It costs nothing." -
/compact keep the codebook decisions and file paths— compact with a focus. Note: compacting a long session is itself a big request. -
/usageshows where your quota went.
→ A.token.management, A.issue.session-endings
What this changes in the book
Most of the book already aligns (short CLAUDE.md, /clear per task, bounding). Points to read with this update in mind:
-
A.token.management — "running out of context mid-session" is now rare; the cost argument for
/clearstill holds. -
C.prompts — role-play prompts ("Play the role of a hostile reviewer…") still work as framing for a review task; just don't treat role + steps as a required template.
-
A9.markdown-project-memory — add: auto memory now handles preferences; CLAUDE.md should focus on gotchas.
Sources
-
Anthropic, Best practices for Claude Code — prompts, interviewing, pass/fail checks, fresh-session review, effort, context and cost, "What you can stop doing" (accessed 2026-11-18)
-
Anthropic, Effort — effort levels and per-model defaults (accessed 2026-11-18)
-
Anthropic, Prompting Claude Opus 5.5 — medium default, effort as main control (accessed 2026-11-18)
-
Anthropic, How Claude remembers your project — CLAUDE.md size, auto memory (accessed 2026-11-18)
-
Shihipar, T. (2026-07-24), The new rules of context engineering for Claude 5 generation models — 80% system-prompt cut, examples, CLAUDE.md, auto memory
-
Prompted by: Peverelli, C. (2026-11-06), Give Me 10 Minutes and I'll Make Your Claude 100X Better (YouTube) — only its claims confirmed by the sources above are included here
Related
-
C.rules — ten rules for AI-assisted research; this update refines Rules 1 and 3
-
C.prompts — prompt library
-
A9.markdown-project-memory — writing CLAUDE.md
-
A.concept.memory — how memory works
-
A.issue.model-selection — choosing models (and now effort)
-
A.token.management — context and quota habits
-
C.calibration — calibrating trust in outputs