claude-sonnet-5-5¶
Every rule as validia lint -m anthropic:claude-sonnet-5-5 reads it, along default → anthropic/default → anthropic/claude-sonnet-5-5. On this model is the severity lint gives it, in bold beside the default where the two differ.
wording¶
Released 2026-10-07; a pin reads it as wording = "1.0.0" under [lint.rules].
Writing for anthropic/default
Current Claude models overtrigger on loud emphasis.
- Drop capitals and exclamation marks: Opus 4.5 to 5.5 and Sonnet 4.6 overtrigger on them.
Sources: Anthropic, Prompting best practices; Anthropic, Prompting Claude Fable 5, Sonnet 5, Opus 4.8 and Opus 5; Anthropic skills repository, prompt-audit guide.
From anthropic/default, 1.0.0.
| Rule | Finds | On this model | From |
|---|---|---|---|
wording/capitals |
Emphasis in capitals | info | anthropic/default |
wording/capitals-in-tool |
Steering in capitals in a tool description | warn | — |
wording/credential |
A credential in the prompt | error | — |
wording/exclamation-marks |
Repeated exclamation marks | info | anthropic/default |
wording/expired-date |
A date-limited instruction | warn | — |
wording/length-cap |
A hard length cap in the prompt | warn | — |
wording/old-model-name |
Names an old model | warn | — |
wording/unwrapped-input |
Untrusted input without delimiters | warn | — |
instructions¶
Released 2026-10-07; a pin reads it as instructions = "1.0.0" under [lint.rules].
Writing for anthropic/default
Current Claude models follow a plain instruction with its reason better than a bare one.
- Say what to do instead of what not to do; keep a prohibition for a real business rule, with its reason.
- Give the reason with the rule: Claude generalises from the reason to cases the rule never names.
Sources: Anthropic, Prompting best practices; Anthropic, Prompting Claude Fable 5, Sonnet 5, Opus 4.8 and Opus 5; Anthropic skills repository, prompt-audit guide.
From anthropic/default, 1.0.0.
| Rule | Finds | On this model | From |
|---|---|---|---|
instructions/completion-booster |
A completion booster | warn | — |
instructions/default-when-unsure |
Makes a tool the default when in doubt | info | anthropic/default |
instructions/habit-to-break |
Names a habit to break | warn | — |
instructions/hedged-requirement |
A hedged requirement | warn | — |
instructions/prohibition-list |
Three or more prohibitions in a row | warn | — |
instructions/rule-without-reason |
A rule without its reason | warn | — |
context¶
Released 2026-10-07; a pin reads it as context = "1.0.0" under [lint.rules].
Writing for anthropic/claude-sonnet-5-5
Sonnet 5.5 holds its role against a pasted persona, and is less accurate when it skips thinking on totals and rules.
- Spend the role line on audience, product and quality bar; the role itself holds.
- Move point systems and totals into code or a tool, or leave thinking on where the prompt asks for arithmetic.
Sources: Anthropic, Prompting best practices; Anthropic, Prompting Claude Sonnet 5.5; Anthropic, Skills best practices; Anthropic skills repository, prompt-audit guide.
From anthropic/claude-sonnet-5-5, 1.0.0.
| Rule | Finds | On this model | From |
|---|---|---|---|
context/arithmetic-in-prompt |
Arithmetic left to the prompt | warn | — |
context/bullet-wall |
Behaviour written as a wall of bullets | info | — |
context/do-not-hallucinate |
Forbids hallucinating without saying what to do | warn | — |
context/identity-stub |
An identity stub for a role | warn | — |
context/reminder-line |
A reminder line | warn | — |
context/repeated-instruction |
Repeats an instruction | warn | — |
context/restated-default |
Restates what the model already does | warn | — |
context/special-cases |
A list of special cases | warn | — |
context/unchecked-quotes |
Asks for quotes without checking them | info | — |
reasoning¶
Released 2026-10-07; a pin reads it as reasoning = "1.0.0" under [lint.rules].
Writing for anthropic/default
Thinking is a setting on current Claude models: effort and the thinking mode, not prose.
- Set effort in the request instead of writing 'think step by step' or 'don't overthink'.
- Do not ask Fable 5 or 5.1, Opus 5 or 5.5, or Sonnet 5.5 to show their reasoning: their safeguards refuse it. Ask for a short justification after the answer instead.
Sources: Claude API docs: Extended thinking, Effort, Refusals; Anthropic, Prompting best practices; Anthropic skills repository, prompt-audit guide.
From anthropic/default, 1.0.0.
| Rule | Finds | On this model | From |
|---|---|---|---|
reasoning/numbered-script |
A numbered script for a judgment task | warn | — |
reasoning/plan-first |
Tells the model to plan | warn | — |
reasoning/self-verify |
A self-verification line | info | — |
reasoning/show-reasoning |
Asks to see the model's reasoning | error (default warn) | anthropic/default |
reasoning/strategy-coaching |
Coaches a strategy | warn | — |
reasoning/think-in-prose |
Controls thinking in prose | warn | — |
output¶
Released 2026-10-07; a pin reads it as output = "1.0.0" under [lint.rules].
Writing for anthropic/default
Fable 5.1 and Sonnet 5.5 narrate progress between tool calls; the request settings decide where updates go.
- Ask for progress updates in the request (display 'updates') rather than writing 'hold all findings' or 'don't narrate'.
- Ask for JSON with structured outputs or a strict tool, not with 'output ONLY valid JSON'.
Sources: Claude API docs: Structured outputs, Errors, thinking display; Anthropic, Prompting Claude Sonnet 5.5; Anthropic skills repository, prompt-audit guide.
From anthropic/default, 1.0.0.
| Rule | Finds | On this model | From |
|---|---|---|---|
output/format-ban |
Bans formatting | info | — |
output/hold-progress |
Holds back progress updates | warn (default info) | anthropic/default |
output/json-in-prose |
Asks for JSON in prose | warn | — |
tools¶
Released 2026-10-07; a pin reads it as tools = "1.0.0" under [lint.rules].
| Rule | Finds | On this model | From |
|---|---|---|---|
tools/discourages-tools |
Discourages tool use across the board | warn | — |
tools/examples-in-tool |
Examples inside a tool description | warn | — |
tools/missing-when-not |
Does not say when not to use the tool | warn | — |
tools/rules-in-tool |
Behaviour rules inside a tool description | warn | — |
tools/sales-pitch |
Sells the tool instead of describing it | warn | — |
security¶
Released 2026-10-07; a pin reads it as security = "1.0.0" under [lint.rules].
| Rule | Finds | On this model | From |
|---|---|---|---|
security/hide-instructions |
Asks the model to keep its instructions secret | info | — |
security/obey-input |
Tells the model to obey the input | warn | — |
maintenance¶
Released 2026-10-07; a pin reads it as maintenance = "1.0.0" under [lint.rules].
| Rule | Finds | On this model | From |
|---|---|---|---|
maintenance/change-history |
Describes a change instead of how it works now | warn | — |
maintenance/grading-talk |
Talks about grading | warn | — |