Video summary

Claude Just Changed Prompting Forever (You Need To Do THIS)

Main summary

Key takeaways

Technology

Core claim of the video

The video argues that Claude 5 (Fable 5 / Opus 5) prompting guidance has changed. Older advice like:

  • “spell out steps”
  • “be overly specific”
  • “double-check everything”

can now worsen output quality and increase cost.

It also compares these changes to OpenAI’s GPT-6 Astra guidance.

What’s “better” prompting now (technological / practical rules)

1) Stop writing overly prescriptive step-by-step prompts

  • Earlier guidance: be specific, list steps, instruct careful verification.
  • New guidance (Anthropic, as cited in the video): models perform better when you provide the complete task intent/goal and let the model work end-to-end.
  • Problem with step lists: Claude 5 models may interpret step lists as narrowing the job to only the user’s steps, potentially degrading quality.

Rewrite example (conceptual)

  • Old: “Open spreadsheet → compute leads/spend → compare to last week → write summary → draft email → fix numbers if wrong.”

  • New: provide a broader goal like: “Tell me how leads, spend, and cost per lead moved vs last week, then draft the client email.” The model is expected to do the math + comparisons itself.

Effort setting caveat (cost control):

  • For simple tasks, set effort to low/medium.
  • Fable 5.1 at low effort can be competitive cost-wise with older Opus models.

2) Add the “reason/intent” (context) instead of omitting “why”

  • Older prompts were advised to be “tight” and omit background (e.g., who the client is, what they pay for, why it matters).
  • Claude 5 guidance (as quoted): performance improves when the model understands intent, with a dedicated section like “give the reason not only the request.”

The video frames this as reducing missing “small decision” ambiguity, e.g.:

  • Is a 4% drop meaningful or noise?
  • Should the email lead with spend or cost/lead?

Template structure described:

  • Larger context / what you’re working on
  • Who it’s for
  • What the output enables
  • The task / definition of done

3) Define “when to stop” via explicit scope + definition of done

  • Older models sometimes did too little, requiring “continue.”
  • Claude 5 models may do too much (over-expanding scope, adding dashboards/charts, extra research/actions).

Claude 5 guides recommend explicitly stating:

  • Definition of done (what the final deliverable must include)
  • Length/format constraints
  • Scope boundaries (what NOT to expand into)

Handling large builds (feature design / planning): Instead of guessing the definition of done upfront, the video suggests using Claude to interview you first (via the “ask user questions” tool), producing a brief that includes:

  • goal
  • audience
  • done criteria
  • edge cases
  • trade-offs

Then start implementation from that brief.

What to remove from older prompts (the “causes extra headaches” list)

4) Replace “hard” trigger words/rules with direct behavior

The video says Anthropic recommends changing patterns like:

  • “critical / must / always / never”

into more natural instructions like:

  • “use this tool when …”

Rationale:

  • over-triggering system behavior
  • reduced token waste by avoiding negation-heavy constraints

5) Review & rewrite existing “skills” and prompt templates written for older models

Takeaway: skills written before Claude 3 (or before the Claude 5 guidance shift) should be reviewed and rewritten.

The video creator claims to provide a “rewrite skill” that:

  • preserves goal/audience/done
  • removes outdated constraints
  • converts negations (“never/always/must”) into positive instructions with intent/reason
  • removes step lists in favor of goal + intent formatting
  • shows changes for approval

6) Remove “think step-by-step / show reasoning / double-check” rituals

Cost + refusal risk on Claude 5:

  • Asking for “think carefully” can over-trigger reasoning since the model already thinks by default.
  • “Walk me through your reasoning…” may cause refusal because extracting reasoning is discouraged.

Verification passes are redundant:

  • Claude 5 already checks its own work; telling it to “double-check every number” can waste tokens/time.
  • Similar issue with “use a sub-agent to verify…”

7) Clarify whether you want an answer or an action (prevent unrequested actions)

Claude Fable 5 may take actions not asked for (e.g., editing sheets/drafting emails/backups).

Recommended boundary rule:

  • If the user is asking/thinking, deliver assessment/findings, then stop.
  • If changes are not requested, do not execute them.
  • Before commands that change system state, require evidence supports the action.
  • Scheduled tasks are treated differently: they can proceed autonomously.

Example boundary idea:

  • “When someone asks a question about a draft, answer the question. Rewrite only when asked.”

8) Fix output “voice” centrally by removing “managed prose”

The video targets overly verbose AI-sounding writing.

  • Claude 5.1 is described as having “managed prose” behavior (dense, fewer breaks, more stock phrasing changes).
  • Recommendation: add one instruction in a global place (e.g., claude.md / desktop instructions) like: “Please remove all managed prose.”

Then remove redundant prompt lines about jargon avoidance/structure—keep it simple: clear, concise outputs.

Cross-model comparison: Claude 5 vs OpenAI GPT-6 Astra (Astra 6)

The video claims Anthropic and OpenAI guidance largely agree on prompt structure:

  • include goal + intent/justification
  • avoid rigid step-by-step prescriptive instructions

Key difference:

  • Astra 6 needs explicit instruction if you want it to act quickly (Claude is more action-biased by default).

Also:

  • OpenAI reportedly has a blocklist for certain humanizer/marketing terms (e.g., “delve,” “leverage,” and related phrases).

Final principle offered:

  • prompts should read like something you’d give a colleague: intent + definition of done > “no/never” constraints.

Explicit “actionable task” for viewers

Run the creator’s rewrite skill (or apply the same concept) on:

  • your Claude.md
  • your three most used skills

Target outcome: improved performance specifically on Claude 5 models.

Main speakers / sources

  • Speaker (primary): the YouTube presenter/creator (not named in the subtitles).
  • Referenced sources:
    • Anthropic prompting guides for Claude 5, including Fable 5 / Opus 5 / Fable 5.1 and a Claude Code best practices guide
    • OpenAI GPT-6 Astra guidance (“Astra 6” prompt guidance)

Original video