I read 31 pages of Anthropic prompting guidance so you don't have to — here's what actually changes with Claude 4.7

The short version

Claude Opus 4.7 follows prompts literally. Generic 4.6-era prompts like "review this contract" or "summarize this report" underperform now, not because the model got worse but because 4.7 stopped guessing at unstated structure.

Six shifts you need to internalize, plus a rewrite checklist you can apply to any existing prompt in under a minute.

The six shifts

1. Name every output. Name every boundary.

• 4.6-era: Review this contract.
• 4.7-ready: Review this contract. Flag risks per clause. Rate severity 1-5. Suggest one rewrite per risky clause. Return as a table with columns: Clause | Risk | Severity | Rewrite.

4.7 does exactly what the sentence says. If you don't name the columns, you get whatever columns it picks. If you don't cap severity levels, you get adjective soup.

2. Length scales with input now. Cap it explicitly.

Long input plus the word summarize used to give you a roughly fixed-length summary. Now it gives you a long summary. Because the input was long.

• Old: Summarize this report.
• New: Summarize this report in exactly 5 bullets. Each bullet under 15 words. First word of each bullet is an action verb.

3. Negative instructions don't stick. Say what TO do.

Don't use jargon is still in the context. 4.7 just doesn't reliably change behavior from it.

• Old: Don't use jargon. Don't sound like a marketer.
• New: Write in plain English a 16-year-old could read aloud. Use short concrete words. Replace "leverage" with "use". Replace "scalable" with "works at any size".

Rule of thumb: every negative instruction rewrites as a positive one plus a concrete swap example.

4. Action verbs ship specific artifacts.

Can you help me with the email? produces a helpful-but-vague paragraph. Action verbs produce a draft.

• Old: Can you help me with the email?
• New:

  Open Gmail. Find <contact> and read our last thread.
  Draft the reply email. Final draft. Send-ready.
  Goal: book a 30-min meeting by Friday.
  Length: under 90 words.
  Tone: confident, casual, specific.

Each verb at the top (Open, Find, Draft) commits 4.7 to producing a shippable artifact, not discussing one.

5. Fewer tool calls, more reasoning between. Ask for aggressive search if you need it.

4.7 calls tools less aggressively than 4.6 did. It reasons more between calls. Usually this is a quality lift. Sometimes you explicitly want broad search.

• Use web search aggressively. Verify every claim with at least 2 sources before answering.

6. Colder default tone. Name the warmth if you want it back.

4.7 dropped the "great question!" energy and most emojis. If your product brand needs warmer voice:

• Use a warm, conversational tone. Acknowledge my framing before answering.

Even better — paste 2-3 reference sentences in the voice you want. 4.7 matches rhythm well.

The one phrase that keeps delivering

Anthropic's own 4.7 doc includes this line, and it has become the single highest-leverage addition you can staple onto any creative or open-ended prompt:

Go beyond the basics. Polish like it's a real client deliverable.

Paired with a section-by-section brief, it consistently pushes 4.7 past the literal-minimum output. I've tested this on landing pages, PR documents, legal memos, and code refactors. Same pattern, same lift.

Full landing page example:

Build a landing page for my AI consultancy.

Sections (in order):
- Hero (headline + subheadline + CTA)
- Logo bar (6 client placeholders)
- 3 case-study cards (problem / what I did / result)
- Service blocks (4)
- Testimonial carousel (3 quotes)
- About (180-word bio + headshot placeholder)
- Newsletter signup
- Footer

Style: editorial, serif headlines, sans-serif body, generous whitespace.
Animations: subtle on scroll. No purple gradients.

Go beyond the basics. Polish like it's a real client deliverable.

The rewrite checklist

Run any prompt through this before sending it to 4.7:

• [ ] Every output is named (format, columns, order, length)
• [ ] Every length is capped (words, bullets, rows)
• [ ] Zero negative instructions — every "don't / no / avoid" rewritten as "do X with example"
• [ ] Action verbs first (Open, Draft, Build, Flag, Summarize — not "can you help…")
• [ ] Tool-use preference stated if it matters (Use web search aggressively or Answer from training, no tools)
• [ ] Tone named, with 2-3 reference sentences if you want warmth back
• [ ] For creative work, the quality lift phrase is appended

One place the pattern helps the most: agent behavior files

If you use Claude Code, Codex, or Cursor with a CLAUDE.md or AGENTS.md in your project root, the same rules apply to those files. Negative instructions ("don't use jargon", "don't hallucinate") age poorly. Rewriting them as positive imperatives with concrete examples measurably improves compliance.

I bundled a 12-rule CLAUDE.md template (Karpathy's original 4 + 8 more covering agent-orchestration failure modes) plus a few working skills. It's a drop-in. Free, MIT:

→ github.com/sisyphusse1-ops/claude-code-pro-pack

And a tiny Python linter that scores any existing CLAUDE.md against the 12 rules and flags leaked secrets:

→ github.com/sisyphusse1-ops/cc-audit

Both are single-commit additions to your repo. No install, no framework.

Sources

• Anthropic's 31-page Claude 4.7 prompting guide (PDF, official)
• Ruben Hassid's digest at ruben.substack.com/p/prompt-47

If this landed, share with one person who's still prompting 4.7 like it's 4.6. That's the thing that actually helps the work.