Everything you need
to get the most from CiarAI

1. Enter your OpenRouter API key in the header field.
2. Type your request in the main input field — describe what you want to achieve, not how to do it.
3. Select the AI Tool you intend to use the prompt with (e.g. Claude, Midjourney, Cursor).
4. Select a Use Case that best matches your task (e.g. Coding, Marketing, Research).
5. Optionally add context in the three context fields, or use a Negative Prompt to exclude things.
6. Click ⚡ Generate Prompt. The result streams in live and is ready to copy, save, or test.

No. That's the point. CiarAI applies prompt engineering for you — including the tool-specific conventions, correct structure, optimal length, and language dialect. You just describe what you want in plain English.

If you do know prompt engineering, CiarAI still adds value: it handles tool-specific conventions you might not know off the top of your head, produces consistent structure, and applies research-backed length constraints for each target tool.

OpenRouter is a unified API gateway that provides access to dozens of AI models — including Gemini, Claude, GPT, Mistral, Llama, Qwen, and DeepSeek — through a single API key and a standardised interface. This means you don't need separate accounts and keys for each AI provider.

CiarAI uses OpenRouter so you can switch models freely without any reconfiguration.

1. Go to openrouter.ai and create a free account.
2. Navigate to Keys in your dashboard and create a new key.
3. Copy the key (it starts with sk-or-v1-…).
4. Paste it into the API Key field in the CiarAI header.

Your API key is stored in your browser's localStorage only — it never leaves your device via CiarAI. When you click Generate or Test, your browser sends the request directly to OpenRouter. CiarAI has no server that could ever see your key.

You can clear all locally stored data at any time using the 🗑 Clear all local data button in the privacy panel (the ⓘ icon next to the API Key field).

★ Auto (Free, most reliable) is the default. It routes to whatever free model is currently live on OpenRouter, so it keeps working even when individual free models are withdrawn. Choose a named paid model when you want a specific one. Your selection is remembered.

Yes. The selected model is used to generate your prompt, not to run it. Larger, more capable models generally produce better-structured, more tool-aware prompts. Claude Opus 4.8 or Gemini 3.7 Flash tend to produce the highest quality outputs. The Auto (Free) option is the best balance of quality and cost for everyday use.

Quick Prompt is the default for new users — it gets you generating in seconds. Switch to Advanced when you need to add context, constraints, or use a specialist input mode.

The three context fields (labelled with placeholder hints based on the active use case) let you add constraints and guidance that shape the generated prompt. Common uses:

• Audience — e.g. "Senior engineers with no AI background"
• Tone — e.g. "Formal, concise, no bullet points"
• Output format — e.g. "Return as a table with columns: Task, Owner, Date"
• Constraints — e.g. "Max 500 words. UK English only."

These are added to the meta-prompt as "ADDITIONAL CONSTRAINTS" and are reflected in the generated prompt's own constraints section.

PPTD stands for Persona / Problem / Task / Deliverable. It replaces the main text field with a structured four-field input designed to capture exactly the context a well-engineered prompt needs.

• Persona — the role you want the AI to inhabit (e.g. "Senior product manager with fintech experience")
• Problem — the specific challenge or pain point (e.g. "40% drop-off in onboarding step 3")
• Task — what you want done (e.g. "Write a phased plan to simplify the flow")
• Deliverable — the exact output format (e.g. "5-section PRD, max 1,500 words")

PPTD is ideal for complex professional prompts where a single text field doesn't capture enough structure. It produces noticeably better outputs for strategy, legal, HR, and product use cases.

Claude Code Mode is a specialist input optimised for agentic coding tasks. It adds three fields specific to what Claude Code (Anthropic's terminal coding agent) needs to operate safely and precisely:

• Goal & Acceptance Criteria — what done looks like, as a numbered checklist
• Files & Scope — @-references for files in scope, and explicit "do NOT touch" items
• Tech Stack & CLAUDE.md Rules — language, framework, and any project-level conventions

Selecting Claude Code in the tool dropdown automatically activates this mode and sets the use case to Claude Code Task.

A negative prompt tells the AI what to exclude from the generated prompt. Examples:

• "Do not use bullet points"
• "Avoid jargon and acronyms"
• "Exclude competitor names"
• "No emojis or informal language"

For image generation tools (Midjourney, Stable Diffusion), the negative prompt maps directly to the tool's native negative prompt field — CiarAI labels this as a separate [NEGATIVE PROMPT]: section in the output.

CiarAI supports 36 tools across six categories:

Microsoft CoPilot

Text LLMs

Coding

Research

Image Generation

Video Generation

Presentations

Open Source / Local

Each of the 36 tools has a detailed profile in CiarAI's TOOL_PROFILES data structure, covering:

• Strengths — what the tool excels at
• Conventions — how to structure prompts for that specific tool
• System note — what to include in the system prompt vs user prompt
• Style — tone, format, and structural preferences

These profiles are maintained with current model versions and are updated as tools evolve. Additionally, certain tool + use case combinations trigger cross-matrix rules — for example, Claude + Legal adds XML tag guidance specific to that pairing.

CiarAI offers 21 use cases. Each one applies a profile that adjusts the emphasis, suggested sections, and formatting conventions in the generated prompt:

When a tool is selected, incompatible use cases are greyed out automatically (e.g. Coding is hidden for Midjourney; WAN / ComfyUI is only available with the WAN tool).

Yes. Your dialect preference is saved in localStorage and automatically restored the next time you open CiarAI. You only need to set it once.

CiarAI auto-sets Creativity based on the selected tool using a research-backed mapping. When auto-set, the selector turns violet and shows a tooltip explaining why it was set.

Examples of auto-selection logic:

• Precise (0.2) — Claude Code, Cursor, Windsurf, Perplexity, Elicit, CoPilot Excel, all coding use cases
• Balanced (0.5) — Gemini, CoPilot Word, CoPilot Studio suite, strategy and HR use cases
• Creative (0.7) — Claude, ChatGPT, Gamma, CoPilot PowerPoint, marketing and social use cases
• Exploratory (1.0) — All image generation tools, all video generation tools

You can override the auto-selection at any time by clicking a different option. The violet highlight clears, and your manual selection takes over. Both auto-set and manual selections are remembered in localStorage.

This is intentional and based on how Midjourney processes prompts. Midjourney has an effective influence limit of approximately 40–60 words. Beyond this, words lose their influence on the generated image — the model cannot attend to everything equally, so longer prompts actually produce worse results by diluting the most important descriptors.

CiarAI constrains Midjourney prompts to 30–60 words with the most important visual keywords ordered first. This is not a bug — it is the correct behaviour for this tool.

HeyGen and Synthesia are avatar-based video platforms where the generated prompt is the avatar's speaking script — the AI reads it word for word. A 50-word prompt would result in a 15-second video. A useful avatar video typically requires a complete script of 200–1,500 words.

CiarAI generates full speaking scripts for these tools, written in natural spoken-word English with short sentences and appropriate rhythm for avatar delivery.

It affects the generated prompt that CiarAI produces — not the final response from the target AI tool directly. However, because the generated prompt carries tone instructions, the target tool will typically reflect them in its output too. A Formal prompt tends to produce a formal response; a Direct prompt tends to produce a tighter, more action-oriented one.

Loop Prompting tells the AI to generate its response, score it against defined criteria, then revise and repeat until the quality is high enough. It is a single quality control that replaces the older CoVe and No Fluff toggles.

The instruction is deliberately specific, so the loop is reliable rather than vague self-editing:

• Score the draft against each criterion out of 10
• Revise any criterion scoring below 8
• Run a maximum of three cycles
• Stop early once every criterion reaches 8 or higher
• Show the improvements in the revised output rather than just claiming them

Use it for strategy, product, legal, HR and writing tasks where a first draft is rarely the best draft. It is automatically disabled for image and video generation, where iterative self-scoring does not apply.

They were replaced by Loop Prompting in v1.8.2. Loop Prompting covers the same ground more effectively, because a scored, multi-cycle loop improves both accuracy and concision in one pass rather than needing two separate toggles.

Fact-Check stays as its own button. It asks the AI to verify claims against outside knowledge, which is a different task from self-scoring, so it is kept alongside Loop Prompting. The two are independent and can be combined freely.

Every time you click Generate, CiarAI saves a numbered version (v1, v2, v3…) of the output in the version bar above the prompt area. You can switch between versions by clicking the tabs — the output area updates immediately.

Versions are stored in the current session only. They are not persisted between page refreshes, unlike saved prompts and history entries.

After each generation, a word and character count appears in the Card 04 header next to the "Generated Prompt" title — for example, 143 words · 891 chars. This gives you an immediate sense of prompt length without having to count or copy the text. The count clears when you reset the form.

From v2 onwards, each version tab shows a small word-count delta compared to the previous version — for example, +12w, −5w, or ~same. This lets you see at a glance whether a regeneration made the prompt longer, shorter, or broadly unchanged, without having to switch tabs and read.

The ⊕ Diff vs prev button shows a word-level diff between the currently selected version and the previous one:

• ■ Green highlight — text added in this version
• ■ Red strikethrough — text removed from the previous version

This is useful for quickly seeing what changed between two generations — for example, if you adjusted the context fields or switched use case between runs.

The ▷ Test Prompt button sends the generated prompt directly to the selected AI model via OpenRouter and streams the response into a separate results panel below the prompt output.

This lets you see how the target AI responds to your generated prompt — without leaving CiarAI or copying it anywhere. It's a quick quality check before you use the prompt in production.

Yes. The 👍 / 👎 rating buttons below the output let you rate the quality of a generated prompt. Ratings are stored in your history against each prompt entry, making it easy to find and reuse your best-performing prompts later.

Two export options are available in the output toolbar:

• Export .txt — plain text file named [tool]-[usecase].txt, ready to paste anywhere.
• Export .json — structured JSON including the prompt text, tool, use case, input, and timestamp. Useful for integrating with other tools or keeping records.

CiarAI has no backend server. There is no CiarAI infrastructure that your prompts pass through. When you click Generate or Test:

Your browser sends requests directly to OpenRouter. Your prompts are subject to OpenRouter's privacy policy. CiarAI has no visibility into that transaction.

All other data — history, saved prompts, API key, settings — is stored in your browser's localStorage and never transmitted anywhere.

Click the ⓘ button next to the API Key field to open the privacy panel, then click 🗑 Clear all local data. This removes your API key, history, saved prompts, templates, settings, and all preferences from localStorage.

Alternatively, clearing your browser's site data for the CiarAI domain will achieve the same result.

• Be specific in your input. "Write a product requirements document for a mobile banking app targeting Gen Z users in Ireland" is far better than "write a PRD".
• Match the tool to the task. Use CoPilot Word for document creation, Claude Code for agentic coding, Perplexity for research — don't default to "General" when a more specific tool is appropriate.
• Use PPTD mode for complex requests. The structured four-field input reliably produces more targeted, higher-quality prompts for professional use cases.
• Iterate with versions. Generate, adjust a field or context, generate again, and compare with Diff. Two or three iterations typically produces a significantly better prompt than the first attempt.
• Set your dialect once. If you're in Ireland or the UK, switch to UK/IE and leave it. The setting persists across sessions.
• Trust the auto-Creativity setting. CiarAI's temperature auto-selection is based on published best practices for each tool. Override it only when you have a specific reason.
• Use the context fields for format constraints. Adding "Output format: numbered list, max 10 items" in a context field ensures the generated prompt explicitly requests that format from the target AI.

Almost certainly not. CiarAI applies research-backed length constraints per tool. If a prompt looks short, check which tool you've selected — image generation tools (especially Midjourney) intentionally produce very short outputs because longer prompts actively degrade image quality.

If you genuinely believe a length is wrong for your specific use case, you can add a length instruction in one of the context fields: "Output should be approximately 400 words".

Yes. CiarAI ships with 18 built-in templates covering common professional scenarios — product requirements, research briefs, code reviews, image prompts, WAN/ComfyUI, and more. Each template pre-fills the tool, use case, and a sample input so you can generate immediately or adapt it to your needs.

Access templates from the Templates section of the sidebar. You can also save any of your own generated prompts as custom templates for future reuse.

v1.8.1 was a quality update focused on the generated prompts themselves.

• Fuller prompts. The generation budget was raised so long structured prompts are produced in full rather than being cut short.
• Closer adherence to your request. The generation instructions are now sent as a system-level instruction, so the model follows them more closely and the output matches what you asked for.
• Better image prompts. Midjourney guidance moved to V8.1 as the current default, and Stable Diffusion guidance was rewritten to lead with Flux-style natural language.

v1.8.0 (Lavender) is a major update that merges the Cloud and Local editions into a single file with a Cloud/Local toggle, updates 12 tool profiles, refreshes the OpenRouter model list, and adds three new use cases.

• Cloud/Local toggle — switch between OpenRouter (cloud) and Ollama (local) from the header. Both configurations persist independently.
• Claude — updated to Opus 4.8 as flagship, with Opus 4.6 and Sonnet 4.6.
• Midjourney — updated from V7 to V8.1 (native 2K HD default, Raw mode for prompt adherence).
• Cursor — updated from 2.0 to 3.5 (Cloud Agents, Composer 2.5, Build-in-Parallel).
• Windsurf — renamed to Windsurf 2.0 / Devin Desktop (Devin Cloud integration, SWE-1.6 model).
• GitHub Copilot — Opus 4.8, Cloud Agent, Plan agent, usage-based billing (June 2026).
• Gemini — noted Gemini 3.5 Flash and Gemini Omni availability.
• Qwen — updated from 3.5 to 3.7 Max.
• DeepSeek — updated from V3.2 to V4 (V4 Pro and V4 Flash).
• Mistral — added Medium 3.5 and Small 4.
• Perplexity — added Sonar Pro Search and Model Council.
• Legal & Compliance — split into three standalone use cases: Legal, Compliance, and Vendor Management & Contracts.
• OpenRouter models — retired models removed (Gemini 2.0 Flash, Claude 3.5 Haiku), Opus 4.6 added, list reordered by usage.

• Generate button moved above the output window — The Generate Prompt button and the enhancement toggles now sit directly above the output area. They control what gets generated, so they belong before the result, not after it.
• Action buttons moved below the output window — Copy, Save, Template, Export, and Reset now sit directly below the output. They act on the result, so they belong after it.
• Card header simplified — The header now shows only the card title and word count. The button row has moved to its logical position below the output.

The flow now reads: describe your request → trigger generation → read the result → act on it.

v1.7.7 restructured the Card 04 layout and updated ChatGPT, Claude, and Gemini prompt profiles based on official guidance. v1.7.6 added the No Fluff Response button. v1.7.5 added the No Fluff Response toggle for copy-editing instructions.

v1.7.4 added three new open source and local model tools and updated existing open source profiles: Gemma 4, Kimi K2.6, and DeepSeek were added as dedicated tools; Qwen was updated to 3.5; Mistral updated to include Devstral 2.

v1.7.3 updated all AI tool profiles: Claude to Opus 4.7, Gemini to 3.1 Pro, Runway to Gen-4.5, Kling to 3.0, Pika to 2.5, and Midjourney with V8 Alpha guidance. Sora was removed following its shutdown. A prompt output formatting conflict between XML tags and bracket-style labels was also fixed.

v1.7.2 added seven improvements: Tone Selector, Word Count Display, improved Copy confirmation, Reset to Defaults, Version Diff Summary, Session Export, and the Fact-Check and CoVe Verification buttons.

v1.7 was a major feature release introducing three new header controls:

• Dialect Selector — Controls language conventions (US / UK/IE / AU) applied when writing generated prompts. Selection persists between sessions.
• Creativity Selector — Exposes the temperature used during prompt generation (Precise 0.2 / Balanced 0.5 / Creative 0.7 / Exploratory 1.0). Auto-sets based on the selected tool and highlights in violet when auto-set. Manual overrides are remembered.
• Prompt Length Optimisation — Research-backed per-tool length constraints silently applied to every generated prompt. Midjourney prompts are constrained to 30–60 words; HeyGen/Synthesia generate full speaking scripts; Claude Code prompts include full agentic context. No UI change required — the correct length is applied automatically.

v1.7 also introduced ⓘ tooltip icons on both Dialect and Creativity selectors, explaining each control inline without requiring a visit to this FAQ.