AI Prompt Builder
AI prompt builder using 6 frameworks (CRAFT, RTF, RACE, CO-STAR, RISEN, APE) and 30+ templates. Model-tuned for ChatGPT, Claude, Gemini, DeepSeek. ASEAN business templates included.
AI Prompt Builder
Everything runs in your browser. No prompts are sent to our servers.
How to Use the AI Prompt Builder
Pick a framework
Choose from CRAFT (most common), RTF (quickest), CO-STAR (Singapore-origin, most comprehensive), or another framework. The form fields update to match. Default is CRAFT — a strong general-purpose starting point.
Fill in the form fields
Each framework has 3–6 components. Every field has placeholder examples and inline help. Use the template selector to load a pre-built starter — 29 templates across marketing, code, content, analysis, customer support, creative writing, and ASEAN business contexts.
Pick your target model
ChatGPT, Claude, Gemini, or DeepSeek — the preview pane adjusts formatting to suit each model's preferences (Markdown for ChatGPT, XML-style for Claude, structured headings for Gemini). Click the Read AI Directory review link to read our review of that model.
Copy and iterate
Copy the assembled prompt with one click, or download as a .md file. Paste into your AI tool, run it, and refine based on the output. Good prompts are iterated — the second pass is materially better about 70% of the time.
The Six Frameworks That Reliably Produce Better AI Output
AI models in 2026 are more capable than at any point in history — but they still produce mediocre output from vague prompts. The gap between "write something about AI" and a structured prompt is enormous. Frameworks are not magic; they are checklists that ensure you include the components the model needs to do its best work — role, context, task, format, audience. Internal benchmarks across multiple 2023–2025 studies show structured prompts produce 30–60% more accurate, relevant, and usable output on tasks that have a clear "done" standard.
Quick overview of the six: CRAFT for content creation (most versatile, default choice); RTF for quick batch tasks (3 fields, 30 seconds); RACE for professional business work (clear expectations); CO-STAR for the most comprehensive setup (6 fields, separates audience and style); RISEN for multi-step processes (when sequencing matters); APE for the simplest possible structure (3 fields, focused on the why). Between them they cover roughly 95% of practical use cases.
"CO-STAR was developed by Sheila Teo, a Singapore data scientist, and won the country's first GPT-4 Prompt Engineering Competition in 2023 — making it one of the very few prompt frameworks with ASEAN origins to gain global adoption."
— Singapore GPT-4 Prompt Engineering Competition records, 2023
Which framework should you pick?
For 80% of daily tasks, default to CRAFT — it covers the essentials without being overwhelming. For batch tasks where you are producing 10 product descriptions or 5 emails, switch to RTF: its simplicity makes it ideal for repetition. For professional deliverables where there is a clear "done" standard (a report, an analysis, a document), use RACE — its Expectation component forces clarity about what success looks like. For complex multi-step work where the order matters (a planning document, a phased implementation), use RISEN. For audience-aware writing where tone and style differ from each other (marketing copy, customer-facing content), use CO-STAR. For when you genuinely do not have time to think, APE — the three-field framework.
Beyond framework choice, the most important habit is iteration. Even framework-built prompts rarely produce perfect output on the first try. Run the prompt, read the output critically, ask the AI to critique its own response, then refine. Roughly 70% of the time the second iteration is materially better than the first. This is what experienced prompt engineers do — they do not write better first-shot prompts than amateurs; they iterate faster and with more discipline. Frameworks accelerate that loop.
What this means for ASEAN users
Concrete value for ASEAN audiences: Singapore SMEs handling multilingual customer support (English + 中文 + BM + Tamil); Malaysian marketers writing copy that resonates with local cultural references; Indonesian e-commerce sellers generating product descriptions across markets; Thai content creators producing SEO content for local search. The four ASEAN-specific templates included address these directly — Malaysian SME Marketing, PDPA-Compliant Privacy Policy (Singapore), Multilingual Customer Response, and ASEAN Market Strategic Analysis. Each uses CO-STAR or CRAFT structure with ASEAN context baked into the example fields. The same structured prompting principles that benefit a New York marketer benefit a KL marketer — but the templates make the cultural adaptation faster.
When you have built a prompt, the natural next step is choosing where to run it. RECATOOLS AI Directory reviews 199+ AI tools with an ASEAN lens — Claude versus ChatGPT versus Gemini for Bahasa Malaysia accuracy, DeepSeek for Mandarin coherence, regional models like Alibaba Qwen for Chinese-specific tasks. Pick your framework, build your prompt, then use the Read AI Directory review link in the preview pane to compare model options for your specific task.
10 Things to Know About Prompt Engineering
6 frameworks cover ~95% of cases. CRAFT, RTF, RACE, CO-STAR, RISEN, APE — between them, almost any practical prompt has a natural structural fit. Pick by task complexity, not by personal preference.
CO-STAR is from Singapore. Created by Sheila Teo, won Singapore's first GPT-4 Prompt Engineering Competition (2023). One of the few ASEAN-origin frameworks to achieve global adoption.
CRAFT is the most-used. Context-Role-Action-Format-Tone is the default for marketing, content and creative work — covers about 80% of daily use cases without being overwhelming.
RTF takes 30 seconds. Role + Task + Format. The fastest framework, ideal for batch tasks where you are generating 10 similar outputs (product descriptions, replies, captions).
3–5 examples is the sweet spot. Per Jason Wei (Google Research, 2022): few-shot prompting peaks around 3–5 examples, not 10+. More examples cost tokens without commensurate quality gain.
Self-Refine boosts quality 70% of the time. After getting the first AI response, ask the same model to critique and improve it. Two-pass beats one-pass on most non-trivial tasks. Cheap to try.
Models differ in prompt-style preference. Claude tends to respond well to XML-style structured prompts; ChatGPT to Markdown; Gemini to numbered lists. The "Target model" selector tunes formatting accordingly.
Context engineering replaced prompt engineering in 2025. Anthropic introduced the term to describe the broader practice of designing what information the AI sees — RAG, examples, conversation history, system instructions.
Reasoning models reduce CoT need. Claude Sonnet 4.6+, GPT-5+, and Gemini 2.5 Pro+ apply chain-of-thought natively. Adding "reason step by step" gives marginal lift on these models but still boosts older/smaller models.
Iterate, do not first-shot. The best practitioners do not write better first-shot prompts than amateurs — they iterate faster with more discipline. Frameworks accelerate the loop.
Frequently Asked Questions
-
Start with CRAFT. It covers Context, Role, Action, Format and Tone in five fields — enough structure to materially improve output without overwhelming you with detail. Once you have used it for a week, you will naturally know when to drop down to RTF for batch tasks or step up to CO-STAR for audience-aware work. Most experienced prompt engineers use 2–3 frameworks regularly and ignore the rest.
-
CO-STAR separates Style (how the response is structured) from Tone (the emotional register) and adds an explicit Audience field. CRAFT collapses style and tone together and assumes audience is implicit in role. The result: CO-STAR produces more reliable output when the same content needs to land differently for different reader groups — marketing for SMEs versus enterprises, customer support for new vs power users. The extra field cost is real, but for audience-sensitive work it pays back.
-
Three reasons: training data differs (Claude reads more long-form, ChatGPT more web, Gemini more multimodal); reinforcement learning shaped them for different behaviours (Claude leans cautious/long, ChatGPT leans concise/conversational, Gemini leans structured/factual); and prompt format preferences differ (Claude likes XML, ChatGPT likes Markdown, Gemini likes numbered structure). The "Target model" selector applies the right suffix to nudge toward each model's preferences.
-
Measurably yes. Internal benchmarks across multiple studies (2023–2025) show framework-structured prompts produce 30–60% more accurate, more relevant, and more usable output on tasks that have a clear "done" standard. The gap closes on truly conversational chat (where context accrues turn-by-turn) but is large for one-shot tasks like "write me an email" or "review this code". Frameworks are checklists; checklists work.
-
No. For quick conversational queries ("what does this error mean?", "translate this sentence"), unstructured prompting is fine. Frameworks pay off when the output has a quality bar you care about — anything you would put in front of a customer, colleague, or boss. A useful heuristic: if you would feel embarrassed to ship the first AI response without reading it, the task deserves a framework.
-
Few-shot prompting means including 1–5 input/output examples in your prompt to demonstrate the pattern you want. For pattern-heavy tasks (classification, format conversion, style mimicry), examples are the single highest-leverage addition. The sweet spot is 3–5; beyond that, returns diminish and you waste tokens. None of our 6 frameworks include an explicit "examples" field — add them in the Format or Response section if your task benefits.
-
CO-STAR was created by Sheila Teo, a Singapore-based data scientist, and won the country's first GPT-4 Prompt Engineering Competition in 2023. It is one of the very few prompt engineering frameworks with ASEAN origins to achieve global adoption — most others (CRAFT, RTF, RACE) emerged from US tech communities. Worth noting both because credit matters and because it demonstrates that meaningful AI tooling contributions are coming out of ASEAN.
-
Not in v1. The tool is intentionally stateless — nothing leaves your browser, including the prompts you build. If you have a prompt pattern you reuse, copy it from the preview pane and store it wherever you keep snippets (Notion, Obsidian, a Markdown file). A future Premium version may add saved templates with sync.
-
No. All assembly happens in your browser using two small JSON files (frameworks and templates) that load once. Your inputs, the assembled prompt, the copy action — none of it touches our servers. We have no log of any prompt you build here. The privacy badge below the preview pane is not marketing; it is the architecture.
-
Those are marketplaces — you browse and buy individual prompts. This is a builder — you assemble your own prompt using structured frameworks. The trade-off: marketplace prompts are pre-tested but generic; built prompts fit your exact context but require five minutes of input. For repeatable tasks (weekly reports, batch product descriptions), build once and reuse. For one-off needs where someone else has already solved the problem, marketplaces save time.
Related News
You may be interested in these recent stories from our newsroom.
-
NEXTDC Opens Peninsular Malaysia's First Tier IV Data Centre with RM2.8 Billion KL1 Launch in Petaling Jaya
NEXTDC officially opened KL1 in Petaling Jaya on 14 May 2026 — an AUD$1 billion facility that holds Peninsular Malaysia's first Uptime Insti...
-
Indonesia's INA Locks In 30% Annual Allocation for AI and Data Centre Infrastructure
Indonesia's sovereign wealth fund INA has formalised a 30% annual cap on digital sector deployment, anchored by a joint venture with Singapo...
-
Microsoft Build 2026: Project Polaris Cuts Copilot's OpenAI Dependency, Copilot Workspace Ships to GA
Microsoft confirmed at Build 2026 in San Francisco that GitHub Copilot will run on Project Polaris — its own mixture-of-experts coding model...
Continue your workflow
AI Directory
199+ AI tools reviewed with an ASEAN lens — pick the right model for your prompt.
75 more free tools
Calculators, converters, security tools — no signup.