Skip to main content
Explainers · Prompting Opus 4.8
A Visual Primer

Prompting Opus 4.8
for legal work.

4.8 takes you literally. Brief it like a junior. Nine stages, end-to-end.

Scroll
What changed in 4.8
Even more literal.Anthropic’s own guidance says 4.8 “does not silently generalise an instruction from one item to another.” State scope explicitly.
Default effort is now “high”.On both the Claude API and Claude Code, high is the new baseline; xhigh is Anthropic’s pick for the hardest analytical work.
Abstains on uncertainty.Lowest incorrect-rate of six models on every benchmark. Still verify every citation — but 4.8 is materially less likely to bluff than 4.6 or 4.7.
Flags ambiguity by default.If you want it to make assumptions and proceed without asking follow-ups, say so. Otherwise it will pause and ask.
01

4.8 takes you literally.

“Clean up this skeleton argument before I file it.”
4.8 RESPONSEFixes typos, grammar, punctuation, paragraph numbering. Returns the document otherwise unchanged. If your authorities section is thin or your relief sought is muddled, that stays muddled.

It did exactly what you asked.
On 4.8, “clean up” means clean up. If you want it restructured, say so.
02

Brief it like a junior.

XML tags separate the brief from the materials. Click any line for what that part does.
03

Direct verbs, not soft language.

Lawyer’s instinctWhat 4.8 needs
Consider tightening the relief soughtRewrite the relief sought as three numbered orders
You might want to add some authorityAdd two binding Court of Appeal authorities post-2015
Feel free to restructure if helpfulRestructure as: issue, law, application, relief
Perhaps a summary at the endAppend a three-sentence executive summary
If appropriate, address the limitation pointAddress the limitation point in a separate section
04

Use XML tags to separate parts.

FLATHere’s my draft. The relevant authority is in this judgment I’m pasting next. Then the witness statement. Then my opponent’s skeleton. Tell me what I’m missing.

[3,000 words of mixed content]

Where does the brief end and the material begin?
STRUCTURED<authority>
[judgment text]
</authority>

<witness_statement>
[statement text]
</witness_statement>

<opposing_skeleton>
[opponent’s skeleton]
</opposing_skeleton>

<task>
Identify the three weakest points in my position and the most likely line of attack.
</task>
05

Show, don’t tell.

✕ DON’T-DO-THIS
“Don’t be verbose.
Don’t use legalese.
Don’t hedge.
Don’t include disclaimers.
Don’t repeat the question.
Don’t add caveats unless necessary.”

4.8 burns tokens parsing rules and often violates them anyway.
✓ SHOW-WHAT-YOU-WANT
“Match this style:

‘The defendant is not entitled to rely on clause 14. The notice was given out of time. The right is lost.’

Three short sentences. No hedging. No preamble.”

Positive examples lock in the pattern. Two beats five rules.
06

Bound the scope.

1
State what to INCLUDE and what to OMIT.
4.8 will not assume exclusions. Spell them out.
Include: chronology, key documents, relief sought. Omit: costs, interest, procedural background.
2
Make rules generalise explicitly.
"Apply this to the first paragraph" does not extend to the rest.
Apply the same citation format — neutral citation and paragraph reference — to every authority cited throughout the document, not just the first.
3
Number multi-part tasks.
Prevents 4.8 from stopping after step one.
Complete all four: (1) identify the weakest argument, (2) rewrite it, (3) check pleadings consistency, (4) list any new authority needed.
4
Batch your questions in one turn.
Each new turn adds reasoning overhead and can drift from earlier literal interpretations.
Three points to address in one message beats three sequential follow-ups.
07

Verify every citation.

!

Opus 4.8 hallucinates less. It still hallucinates.

4.8 is materially better than earlier models at abstaining when uncertain rather than bluffing — Anthropic reports the lowest incorrect-rate of six models across every benchmark. But fabricated case names, wrong neutral citations, and overruled authorities still happen — and they are professional misconduct waiting to occur. Ayinde and the cascade of Hamid hearings make the point. Your verification duty is unchanged.

PATTERN: ground-then-verify
For every authority you cite, quote the specific passage you are relying on verbatim. If you cannot quote it, do not cite it. List each authority with neutral citation in a separate section at the end so I can verify against the National Archives or BAILII before filing.
08

Set the right effort level.

low
Triage. Tagging documents, simple extractions.
On 4.8, low scopes work tightly to exactly what was asked.
medium
●●
Drafting. Letters, short summaries, routine emails.
Default effort on 4.8 across the Claude API and Claude Code.
xhigh
●●●●
Hard problems. Multi-document synthesis, complex appeals, statutory interpretation.
Anthropic recommends xhigh for the most intelligence-demanding work.
max
●●●●●
Final pass. Correctness-critical review. Can overthink.
09

Build a prompt. Live.

Task
Jurisdiction
Audience
Output format
<role>
You are pupil counsel to a senior junior preparing a skeleton argument.
</role>

<context>
Jurisdiction: England and Wales. All authorities must be binding or persuasive in this jurisdiction.
Audience: judicial reader. Tone: precise, restrained, no rhetorical flourish.
</context>

<documents>
PASTE: witness statements, key documents, opponent's case
</documents>

<task>
Read the materials in <documents> and identify the three strongest grounds. For each ground, state: (1) the issue, (2) the binding authority, (3) the application to our facts. Quote the passage you rely on verbatim. If a ground depends on facts not in the documents, flag it rather than infer.
</task>

<format>
Format: numbered paragraphs. Each paragraph under 80 words. Cite authorities by neutral citation with the paragraph relied on, e.g. [2024] EWCA Civ 123 at [45].
</format>

<verification>
For every authority cited, quote the passage you rely on verbatim. List every neutral citation at the end. If you cannot quote it, do not cite it.
</verification>

<failure_mode>
If any required fact is missing from the documents, state the assumption and flag it. Do not infer. Do not fabricate citations.
</failure_mode>
Compliance note. For client work, run Opus 4.8 through an enterprise tool with a UK or EU endpoint and contractual zero data retention. Do not paste privileged or client-identifying material into consumer-tier interfaces. BSB Core Duties CD6 and UK GDPR remain your obligations regardless of model capability.

4.8 doesn’t read your mind.

It does exactly what you ask. The skill is asking precisely. Brief, don’t chat. Show, don’t tell. Verify every citation. Then let it work. Next in the series: where the text goes when you press send.

Next: Where Your Data Goes →