Prompt ChatGPT (R-T-C-F)

The RTCF framework for LLM prompts

RTCF stands for Role · Task · Context · Format and is one of the most quoted templates for structuring prompts sent to ChatGPT, Claude and Gemini. It became popular in 2023 inside the developer community on Twitter/X — including posts from Greg Brockman (OpenAI) — as a quick checklist that turns a vague request into a deterministic instruction. The premise is simple: large language models need constraints, and ambiguous prompts produce vague outputs while structured prompts produce specific ones.

Each letter answers a different question. Role defines the persona the model should impersonate (“You are a senior copywriter specialised in B2B SaaS”) — this anchors vocabulary and depth. Task is the explicit verb (“Write an outreach email to a CTO”). Context is the background that the model could not guess (“Product: monitoring tool; ICP: Series A startups; goal: book a 15-min demo”). Format is the output shape (“Email under 100 words, subject line below 6 words, conversational tone, single CTA”).

Anatomy of an RTCF prompt

Role: You are a senior B2B SaaS copywriter.
Task: Write a cold outreach email to a CTO.
Context: Product = a monitoring tool for Node.js APIs.
         ICP = Series A startups (50-200 engineers).
         Goal = book a 15-minute demo.
Format: 80-100 words, subject < 6 words,
        conversational, single CTA, no emojis.

Notice how every line removes ambiguity. Without Role the model defaults to a generic assistant tone; without Format it tends to over-explain; without Context it hallucinates product features.

Variants: CRAFT, CO-STAR, RISEN, CARE

RTCF is one of a family of complementary acronyms. CRAFT — Context, Role, Action, Format, Tone — explicitly separates tone. CO-STAR — Context, Objective, Style, Tone, Audience, Response — was popularised by GovTech Singapore for enterprise use and adds Audience. RISEN — Role, Input, Steps, Expectations, Narrowing — suits multi-step pipelines. CARE — Context, Action, Result, Example — favours few-shot prompting. STAR (Situation, Task, Action, Result) comes from interview frameworks and is great when you need the model to reason rather than write. 5W1H (Who, What, When, Where, Why, How) is the journalistic baseline. Pick one and stick to it — consistency matters more than the specific letters.

Prompt-engineering principles behind RTCF

Anthropic’s and OpenAI’s official prompt-engineering guides converge on four rules. Be specific: replace adjectives with numbers (“short” → “under 100 words”). Give examples (few-shot): 2-3 worked samples beat any adjective. Break complex tasks: chain-of-thought (“think step by step”) and explicit sub-steps. Specify the format: JSON, Markdown, bullet list, table — otherwise you get prose. RTCF bakes all four into a memorable template.

Anti-patterns to avoid: vague openers (“Write a blog post about X”), conflicting constraints (formal AND funny), missing context (model invents facts), and bloated system prompts — remember every token costs money and inflates latency, so balance verbosity against cost.

Tools that store RTCF prompts

• ChatGPT Custom GPTs — system prompt persisted at the GPT level.
• Claude Projects — persistent context window across conversations.
• LangSmith — versioning and A/B testing for prompts in production.
• Promptly, PromptLayer, Helicone — observability and prompt registries.

FAQ

Is RTCF better than a free-form prompt? Yes whenever you need consistent, repeatable outputs — documentation, automations, customer-facing emails. For exploratory chat (brainstorming, learning) the structure is overhead.

When should I NOT use RTCF? Pure brainstorming, casual Q&A, code-completion inside an IDE — in these cases the model already has the context (open files) and the structure slows you down.

Can I save my RTCF as a reusable template? Yes — Custom GPTs (OpenAI) and Projects (Claude) both let you persist the system prompt. For dev workflows, store the template in a .prompt file inside your repo and version it with Git.

Does RTCF work in Portuguese too? Yes — modern LLMs handle both languages, but pin the output language explicitly in Format (“Responda em português brasileiro”) so the model does not switch mid-answer.