Claude 5 Changed Prompting Forever

Claude 5 Changed Prompting Forever — Fix Yours Now

🎬 Simon Scrapes 📅 Sep 10, 2026 ⏱ 19:11
Claude 5 prompting Opus 5 Fable 5 7 rules

🎯 Why Your Old Prompts Are Broken

Everything you were taught — be specific, spell out the steps, tell it to double-check its own work — now makes the output worse on Fable 5 and Opus 5. And that's Anthropic saying it, verbatim. 0:00

The verbatim quote from the Opus 5 guide: "It performs best when given the complete task specification up front and left to run." And the Fable 5 guide goes further: "Skills written for prior models are often too prescriptive for Claude Fable 5, and as a result can degrade output quality."
Plain English: these models are trained to do an entire job end to end. Hand them a list of steps and you're narrowing them down to your version of the job — potentially worse than what they'd have worked out from a broader goal.

➕ What to Add (Rules 1–3)

Three things your old prompts are missing. 0:35

Rule 1 — Replace step lists with a broader goal

Instead of "open the spreadsheet, work out leads/spend/cost-per-lead, compare to last week, write summary, draft email, fix wrong numbers" — just say "using this week's tracker, tell me how leads, spend, and cost per lead moved against last week, and draft the client email." Give it a metric defining success, not steps. Caveat: for genuinely small tasks, Anthropic advises dropping the effort setting to low/medium — Fable 5.1 on low is still cost-competitive with older Opus.

Rule 2 — Give the reason, not only the request

The "why" was the detail we were told to strip out. Now the Fable 5 guide has a section literally called "give the reason not only the request." On any real task there are ~50 small decisions you don't specify — is a 4% drop bad news or noise? Should the email lead with spend or cost-per-lead? If the model knows what the client is judged on, it makes those calls the way you would.
The template: "I'm working on [larger task]. It's for [who]. This output would enable them to [why]. With that in mind, [task]." — intent, then definition of done.

Rule 3 — Define "done" (the models now over-work)

Old models did too little; Claude 5 does too much. "Opus 5 by default expands scope and adds steps you didn't ask for" — a dashboard you didn't want, a chart, deep research. Fix: define what done looks like and the length explicitly. For narrow tasks, constrain scope — "a client email under 300 words, leading with X, top 3 reasons in a table."
For larger builds: don't write a giant brief you can't define — let Claude interview you ("interview me in detail using the ask user questions tool — goal, who it's for, what done looks like, edge cases, trade-offs") and write the brief itself.

➖ What to Remove (Rules 4–7)

Four things to strip from your old prompts and skills. 8:32

Rule 4 — Drop "critical / always / never / must" + CAPS

These "over-trigger" the model. Anthropic cut 80% of Claude Code's own system prompt this way — replacing hard constraints with plain instructions: use this tool when instead of CRITICAL, you must ALWAYS use this tool. "Replace the rule with the behavior you want, not the behavior you don't want — plus the reason."

Rule 5 — Remove "walk me through your reasoning" and "double-check every number"

Both now cost money for worse output. Opus 5 thinks by default — "think carefully" over-triggers a second pass. Fable 5 has a safeguard against reasoning-extraction, and "walk me through your reasoning" can trigger a refusal. "Double-check every number" buys a second pass on work already done. Simon found 11 instances of these in his own files.

Rule 6 — State the boundaries (answer vs action)

Claude 5 is much quicker to act — it can take unrequested actions (draft an email nobody asked for, create backups, make a dashboard). Fix: "when the user is describing a problem or asking a question, the deliverable is your assessment — report your findings and stop. Before running a command that changes system state, check the evidence supports it." Add this to CLAUDE.md/global instructions.
The exception: scheduled/cron tasks get the opposite instruction — operate autonomously and proceed on reversible actions, since they run while you're away.

Rule 7 — Fix the voice in one place ("remove all mannered prose")

Opus 5 is built to write long outputs; Fable 5.1's writing is denser — "mannered prose" — where "sentences run longer and there are fewer paragraph breaks." Anthropic: "This is why mannered prose irritates — it makes the reader work harder so the writer can perform." Fix: a single line Please remove all mannered prose in CLAUDE.md, instead of scattered "don't use jargon / sentence structure" rules.

🔀 Claude vs GPT-Astra

How the rules stack up against OpenAI's GPT-6 Astra guide (published the same time). 17:06

PointWhere they agree / differ
Agree: goal + intent, not stepsBoth want plain paragraphs with the why, not prescriptive steps
Agree: over-checking is wasteAstra over-checks small changes the way Opus 5 does — stating things already in the system prompt wastes tokens
Differ: bias toward actionClaude 5 over-acts (constrain it); Astra has no action bias — you must add "bias towards action" if you want it
Differ: banned wordsOpenAI ships a blocklist ("delve", "leverage"…) in system prompt 2 — humanizer skills become redundant
Anthropic's own test: "can you show your prompt to a colleague with zero context — if they'd be confused, Claude will be too." The new style reads like what you'd hand a colleague, not a spec.

💡 Key Takeaways

  1. Step lists now hurt — Claude 5 is trained for end-to-end jobs; give a broader goal + success metric, not steps.
  2. Give the reason, not only the request — intent lets the model make the 50 small decisions you didn't specify.
  3. Define "done" explicitly — Claude 5 over-works by default; constrain scope and length.
  4. Drop "critical/always/never/must" + CAPS — they over-trigger; use plain "use X when" with the reason.
  5. Remove "walk me through your reasoning" — Opus 5 already thinks; on Fable 5 it can trigger a refusal.
  6. State the boundaries — "report findings and stop" vs "take action" — Claude 5 acts unprompted.
  7. Fix voice in one place — "remove all mannered prose" beats scattered style rules.
  8. Rewrite your old skills — every skill/CLAUDE.md written before this summer needs review; Simon built a free rewrite skill for it.

📍 Timestamp Index

0:00 Why old prompts are broken
0:35 Rules 1–3 (what to add)
8:32 Rules 4–7 (what to remove)
17:06 Claude vs GPT-Astra
☰ View all