Search for a command to run...
Score your LLM prompt against a checklist of clarity, structure, and format rules — and get a templated rewrite. Everything runs locally with no model calls.
Suggestions will appear here as you type.
Best for: Anything that doesn't fit a more specific shape. The default fallback.
You are a helpful expert assistant. # Task [State the task in one or two sentences.] # Context [Add background, audience, and any data the model needs.] # Constraints - Keep the response under 300 words. - Avoid speculation; ask for missing information if needed. # Output format Respond in clean Markdown with short headings and bullet points. # Examples [Optional: paste one or two input/output examples here.]
Drop in the prompt you'd send to an LLM — a question, a task, a long instruction block. Nothing leaves your browser.
A 0-100 grade plus five category bars (clarity, specificity, structure, format, examples & constraints) update on every keystroke.
Each detected issue is labelled critical, warning, or info — with a one-line explanation of how to address it.
Pick a template (structured, code, writing, analysis), copy or download the rewrite, or click "Use rewrite" to replace your prompt above.
The optimizer runs a set of pure-JavaScript heuristics — regexes for role / audience / tone / format / examples, word counts, vague-pronoun density, hedge-word density, action-verb detection, etc. Each detected issue subtracts points from a category cap of 20. There is no model call; the same prompt always gets the same score, and your text never leaves the page.
Strong action verb, no hedge words like "maybe" or "stuff", no all-caps shouting, no chained tasks.
Long enough to carry signal, short enough to be focused, with concrete nouns instead of vague pronouns.
Names a role ("You are…"), an audience ("for a senior engineer"), and a tone — so the model knows who to be.
Explicit output shape: JSON, Markdown, a bullet list, a table — anything but "a paragraph or two".
At least one worked example and a couple of constraints (length, what to avoid). Examples are the single biggest signal.
Role · Task · Context · Constraints · Output format · Examples. The safe default when you don't have a more specific shape in mind.
Senior-engineer persona, language/framework constraints, single fenced code block plus a short "what changed" rationale.
Methodical partner: ranks hypotheses, asks for a verification step, then proposes a fix grounded in the evidence you provided.
Blockers / Suggestions / Nits / Good catches, each referencing exact lines or symbols — no vague feedback.
Staff-level architect: explicit assumptions, two options compared with diagrams, recommendation, risks and mitigations.
Diverge before you converge: 8+ ranked ideas from safe to wild, ending with a curated top 3 to explore.
Careful-analyst persona, evidence-grounded findings, TL;DR + bullets + open questions output structure.
Strict JSON only, no invented fields, `null` for unknowns. Includes a worked input/output example for the schema.
Editor persona, audience & word budget, active-voice constraints, return the draft only — no preamble.
Patient teacher: what it is, why it matters, an analogy, a minimal example, the common misconception, a next step.
LLMs charge and remember by tokens, not characters. The counter above uses a BPE-style heuristic that lands within ~10% of the real tokenizer for English and code — same one the JSON Formatter uses. The chips show which standard context windows your prompt fits in, leaving ~25% headroom for the model's reply.
"Write a function and also translate it to Spanish and also explain it" yields shallow results on each. Split into separate prompts.
"You're the best AI ever, please please…" doesn't improve quality. It eats tokens and dilutes the actual instruction.
API keys, tokens, customer PII — even into a local optimizer. Once you copy the rewrite into a real model, that data leaves your device.
Without a target (audience, format, constraints), "creative" defaults to clichés. Pair it with a concrete success criterion.
Put the task in the first sentence. Background, audience, and examples come after. Models attend more to the start.
Use exact numbers: "3 bullet points", "under 80 words", "5 ideas". Fuzzy quantifiers produce fuzzy output.
Same promise as the rest of tool127: no model calls, no analytics, no network. The prompt you type, the score, and the rewrite all live in your browser — and disappear when you close the tab.
Format megabytes of JSON in milliseconds.
Data never leaves your browser for maximum privacy.
Works perfectly even without an internet connection.