Maxx your prompt.
Twelve checks, zero servers.
Paste any Claude, GPT or Gemini prompt below. We score it on twelve documented best-practice checks — role, format, examples, constraints, chain-of-thought, vague qualifiers, unfilled placeholders — and append the template snippets you are missing.
100% in-browser. Your prompt never leaves the page. Soft cap: 20 analyses/day, 8 000 characters per prompt.
The 12 checks, decoded
Why each rule moves the output.
- 01 Establish a role.
"You are a senior security engineer reviewing pull requests." Setting an explicit persona consistently outperforms zero-role prompts on reasoning, tone and domain accuracy. Anthropic and OpenAI both highlight this as the single highest-leverage prompt move.
- 02 Use a specific action verb.
Summarise, analyse, list, generate, refactor, classify — clear verbs reduce ambiguity. "Help me with X" and "Can you do something about Y" force the model to guess the deliverable.
- 03 Define the output format.
JSON with these keys. Markdown table. Three paragraphs. Bullet list. Without a format spec the model defaults to prose, which is usually the wrong shape for downstream use.
- 04 Include examples (few-shot).
Even one input/output example dramatically improves consistency. "Few-shot" prompting consistently beats "zero-shot" on classification, extraction, formatting and tone-matching tasks.
- 05 State the constraints.
"Do NOT mention X. Avoid Y. Only return Z." Negative instructions are as important as positive ones — they prevent drift, off-topic prose and unwanted disclaimers.
- 06 Specify tone and style.
Without a tone target, the model defaults to friendly-assistant prose. "Terse." "Academic." "Matter-of-fact, no greeting, no sign-off." Add it explicitly.
- 07 Bound the length.
"Under 150 words." "Three bullets max." "Five sentences." Length caps force the model to compress, which usually produces better writing than the default rambling baseline.
- 08 Kill vague qualifiers.
Good, nice, best, great, interesting — these force the model to invent a quality bar. Replace with measurable criteria: "easy enough for a junior engineer", "shorter than three paragraphs", "no jargon a non-technical reader would not know".
- 09 Fill every placeholder.
[BRACKETS], {CURLY_VARS}, TODO, FIXME and XXX survive into production prompts more often than you think. The model treats them literally — and you get back a response addressed to "[USER_NAME]".
- 10 One focused task at a time.
Asking three things in one prompt halves quality on each. "Summarise this AND write a tweet thread AND extract action items" — chain three prompts instead. The model is parallel-bad and sequential-good.
- 11 Specify the audience.
"For a non-technical founder." "For a Senate staffer." "For a 9-year-old." Audience changes vocabulary, structure and what counts as a finished answer. Specify it.
- 12 Ask for reasoning (chain-of-thought).
For non-trivial tasks, "Think step by step before producing the final answer" measurably improves accuracy. Wei et al. 2022 popularised this; it remains one of the cheapest accuracy boosts in the LLM toolkit.
Before / after
Real examples.
Two common asks, rewritten with the 12 checks applied. The before reads like a real prompt people send. The after is what scores 90+.
Can you help me write a good email to my boss about the project being late?
You are a senior product manager writing a status update to a director-level boss who hates surprises. Write an email that: - Opens with the one-line status (red / yellow / green) - States the new ETA with one sentence of justification - Proposes two mitigation options the recipient can pick from - Closes with a single clear ask (decision needed by Friday EOD) Tone: terse, no apology theatre, no filler. Under 120 words. Plain text. Think through the framing before drafting.
Make my landing page copy better.
You are a B2B SaaS copywriter who has written for Linear, Vercel, and Stripe.
Rewrite the landing page hero copy below for stronger conversion. Constraints:
- Headline: max 8 words
- Subhead: max 24 words, must promise a specific outcome (not a feature)
- CTA: 2-3 words, verb-first
- Tone: confident, no exclamation marks, no superlatives ("best", "amazing")
Format: return as JSON with keys "headline", "subhead", "cta".
Example:
Input: "Run faster. Smarter. Together."
Output: {"headline": "Ship 4x faster", "subhead": "...", "cta": "..."}
CURRENT_COPY: [paste here] FAQ
How does Promptmaxxing work? +
It runs 12 regex- and heuristic-based checks against your prompt in the browser. Each check maps to a documented prompt engineering best practice (role, format, examples, constraints, etc.). Your prompt never leaves the page — no API calls, no logging.
Why 12 checks and not more? +
These are the rules with consistent, replicated evidence across model providers (Anthropic, OpenAI, Google) and the prompt-engineering research literature. We deliberately do not check stylistic preferences ("use bullet points!", "always say please!") that vary by task and model.
Is my prompt sent anywhere? +
No. The analysis runs entirely in your browser using local regex and rule logic. You can disconnect from the internet after loading the page and the tool still works. We do not log prompts.
Why is there a daily limit if everything is client-side? +
Two reasons. First, a soft cap (20/day per browser) prevents accidental flooding of the UI from automated tests or scripts. Second, when we add an optional "Rewrite with AI" feature in a future version, the same limit framework will apply at the server level — so we are not introducing it as a surprise then.
Can I bypass the daily limit? +
Yes — clearing localStorage resets it. We intentionally designed it as a soft limit, not an adversarial one. If you genuinely use prompt engineering 50 times a day, the local rules pattern is the same as ours and you can just run the checks mentally.
Will you add AI-powered rewrite? +
Probably yes, as a V2. A real LLM rewrite would catch nuances regex cannot (tone matching, domain-specific jargon, structural rebalance). It would have a per-day quota and a paid tier for power users. The current rule-based optimiser gets you about 80% of the value at zero cost.
Does this work for image / audio / video prompts? +
Partially. The role, format, length, constraints and audience checks apply equally to image and video generation prompts. Examples and chain-of-thought are less applicable for image prompts. We are considering a dedicated image-prompt mode in a future version.
What does NOT show up in the check list? +
Stylistic preferences (Markdown vs plain), model-specific quirks (Claude likes XML tags, GPT likes JSON, Gemini likes structured output), and domain-specific patterns (code generation prompts have their own ladder). The 12 checks are the universal core. The other 30% lives in the per-task playbook.