Paste a prompt, choose where you will run it, and get deterministic suggestions grounded in official Claude prompting guidance. Your prompt text never leaves this browser.
Your inputs never leave this page — no sign-up, no upload, no model call.
Run the linter to see severity-tiered findings.
Heuristic, not a quality score. Rules cite official guidance; your judgment decides.
System prompt
Discourage unnecessary planning loops.
Prefer the smallest plan that lets you act safely; update it only when new facts change the path.Citations
System prompt
Keep implementations simple.
Choose the simplest implementation that satisfies the stated constraints and tests.Citations
User message
Request concise output.
Answer in the fewest words that preserve the key decision, evidence, and next step.Citations
Harness
Force progress checks in agent loops.
After each major step, summarize what changed, what was verified, and the next action.Citations
Harness
Report only verified progress.
Do not claim success until a command, test, or observable artifact verifies it.Citations
System prompt
Permit safe independent action.
When the next step is reversible and within scope, take it without asking.Citations
System prompt
Clarify what not to modify.
Stay within the requested scope; if required information is missing, state the blocker instead of guessing.Citations
multishot-examples
Examples are the most reliable way to steer exact output shape.
Format-sensitive task without examples. Add 3–5 diverse examples in <example> tags.
Citations
long-context-ordering
Long-context prompts are easier to ground when evidence is loaded before the ask.
Large pasted context appears after the task. Put documents first, then the query at the end.
Citations
reasoning-echo
Echoing hidden reasoning can trigger reasoning-extraction safeguards.
Prompt asks to reveal/transcribe hidden reasoning. Ask for a concise rationale or summarized thinking instead.
Citations
structure-sections
Structure helps the model separate source material, task, and expected output.
Long prompt with no section markers. Use <context>, <task>, and <output_format> or clear headings.
Citations
negative-instructions
Positive behavioral targets are easier to follow than dense prohibitions.
High density of do-not/never instructions. Say what to do instead of only what to avoid.
Citations
constraint-without-rationale
Rationale improves compliance with unusual constraints.
ALWAYS/NEVER/MUST without nearby rationale. Add because/so that language explaining why the rule matters.
Citations
missing-output-format
Explicit output contracts reduce unusable responses.
Prompt does not state the desired output format. Name the format: JSON, table, markdown, bullets, or exact fields.
Citations
vague-qualifiers
Measurable targets make subjective editing requests actionable.
Vague words such as better/professional without criteria. Use measurable criteria, rubrics, examples, or constraints.
Citations
missing-context-intent
Audience and purpose help choose tone, depth, and trade-offs.
Task-only prompt lacks audience, goal, or purpose. Use “I am working on X for Y because Z.”
Citations
caps-emphasis
Aggressive emphasis can overtrigger modern Claude models.
Excessive CRITICAL/MUST/NEVER style emphasis. Use calm wording and one explicit priority statement.
Citations
fable-overprescription
Fable is more reliable with concise intent than enumerated behavior laws.
Repeated always/never/must behavior steering. Prefer a short instruction describing the desired behavior.
Citations
fable-effort-guidance
Effort levels differ by surface: API supports low|medium|high|xhigh|max; Claude Code settings.json effortLevel caps at xhigh. Irrelevant in claude.ai.
Complex work on API/Claude Code can benefit from effort guidance. Use high as a floor; xhigh for hard agentic/coding work. max is API-only — Claude Code tops out at xhigh. Never set effort for Haiku.
Citations
prompt-bloat
The concept asks for local token-range feedback; bloat detection turns that estimate into an actionable rule.
Very long prompt appears repetitive or boilerplate-heavy. Keep durable constraints, remove repeated boilerplate, and move reusable policy to system or harness context.
Citations