# Block Types flompt = 13 semantic block types. Each represents one aspect of structure. Assembler applies Claude best practices. Auto-ordered. Auto-formatted. *** ## Document External reference content. Claude parses `` tags reliably. **Assembled as:** ```xml Your document title [your content here] ``` **Use for:** Articles. Code. Datasets. Contracts. Always first. Color: green-light `#86efac` *** ## Role Defines the AI persona or expertise. A well-crafted Role block significantly sharpens the tone, vocabulary, and reasoning style of the response. **Examples:** `"You are an expert Python developer"`, `"Act as a senior product manager with 10 years in B2B SaaS"` Color: violet `#c084fc` *** ## Tools Declares the functions and tools the AI is allowed to call. Each entry describes a tool name, its purpose, and its parameters — giving the AI a precise, bounded set of available actions. **Example:** ``` search_web(query: string) — Search the internet for up-to-date information read_file(path: string) — Read the contents of a file at the given path run_terminal(command: string) — Execute a shell command in the current working directory create_memory(key: string, value: string) — Persist a value for later retrieval ``` **When to use:** When building agentic prompts where the AI needs to act — not just generate text. Tool declarations constrain and inform the model about what actions are possible, preventing hallucinated function calls and improving reliability. Color: orange `#fb923c` *** ## Audience Specifies who the output is written for — their expertise level, role, background, or expectations. Distinct from Role (who the AI *is*), Audience defines who the AI is *speaking to*. **Examples:** `"Software engineers familiar with REST APIs but new to async programming"`, `"Non-technical product managers who need to make a go/no-go decision"` **Why it matters:** The same content, written for a junior developer vs. a CTO, will differ dramatically in vocabulary, depth, and framing. Making the audience explicit lets Claude calibrate automatically. This is the "A" in the COSTAR prompt framework. Color: blue `#93c5fd` *** ## Context Provides background information and situational context. This is the "why" behind the objective — it helps Claude understand the full picture before acting. Color: slate `#94a3b8` *** ## Environment Injects system-level context variables — the runtime state the AI needs to operate accurately. Distinct from Context (narrative background), Environment is machine-readable situational data. **Examples:** ``` OS: Ubuntu 22.04 LTS Working directory: /home/project Current date: 2026-03-16 Runtime: Node.js 20.x Token budget: 8000 ``` **When to use:** When building agents or system prompts where the AI needs to know where and when it is operating — current date, file paths, environment variables, available runtimes, token constraints. Color: cyan `#22d3ee` *** ## Objective Defines the main task to accomplish — what the AI should *do*. The core of your prompt — clear, actionable, specific. **Examples:** `"Your goal is to review this code for bugs and style issues"`, `"Write a 500-word summary of the document above"` Color: amber `#fbbf24` *** ## Goal Defines the end goal and success criteria — *why* the task matters and what a good outcome looks like. Distinct from Objective (what to do), Goal explains what success means. **Examples:** `"Help the reader decide in under 2 minutes whether to integrate this API. Prioritize clarity over completeness."`, `"The summary should be usable as a standalone executive brief — no prior knowledge assumed."` **Why it matters:** When the task underspecifies what matters most, the AI has to guess. An explicit Goal lets Claude make the right trade-offs — length vs. depth, speed vs. accuracy, breadth vs. precision. This is the "G" in the RISEN prompt framework. Color: emerald `#6ee7b7` *** ## Input The actual data, variables, or content provided to the AI. Separating input from objective keeps prompts clean and reusable — swap the input without rewriting the whole prompt. Color: green `#4ade80` *** ## Constraints Rules, limitations, or guardrails for Claude. Be explicit — the more precise, the better. **Examples:** `"Do not use bullet points"`, `"Never exceed 300 words"`, `"Only use information from the provided document"` Color: rose `#fb7185` *** ## Examples Few-shot demonstrations: input/output pairs that show Claude exactly what you expect. This is one of the most powerful techniques in prompt engineering. Format your examples as: ``` Input: [your input here] Output: [expected output here] ``` flompt automatically parses these pairs and wraps them in Claude's structured XML format: **Assembled as:** ```xml your input expected output ``` **Why it matters:** Structured few-shot examples dramatically improve consistency and reduce ambiguity — Claude knows precisely what "good" looks like. Color: violet-light `#c4b5fd` *** ## Chain of Thought Explicit step-by-step reasoning instructions. Tells the AI *how* to think through the problem before answering — making implicit reasoning steps explicit and controllable. **Examples:** * `"Think step by step. First identify the root cause, then evaluate three possible solutions, then recommend the best one with a rationale."` * `"Before answering, list the key assumptions you're making. Then reason through each one."` **Assembled as:** ```xml Think step by step. First identify the root cause… ``` **Why it matters:** Chain of thought prompting reliably improves performance on complex reasoning tasks by forcing structured decomposition before committing to an answer. Research from Wei et al. (2022) shows it can significantly boost accuracy on math, logic, and multi-step tasks. Color: amber-light `#fde68a` *** ## Output Format Specifies the expected response structure — JSON, bullet list, table, numbered steps, prose, markdown, etc. **Examples:** `"Respond with a JSON object containing 'title', 'summary', and 'tags' fields"`, `"Use markdown headers (##) for each section"` Color: accent `#ff6b9d` *** ## Response Style Controls the tone, voice, and style of the response — distinct from Output Format (which specifies structure) and Constraints (which set limits). **Examples:** `"Write in a conversational, approachable tone — avoid jargon"`, `"Use short sentences. Active voice. No filler phrases."`, `"Match the tone of a senior engineer writing an internal RFC."` **Why it matters:** The same objective with a "formal academic" vs. "casual conversational" response style will produce completely different outputs. Making it explicit prevents Claude from defaulting to a generic neutral tone. Color: teal `#5eead4` *** ## Guardrails Hard limits and safety refusals — behavioral rules that define what the AI must never do, regardless of how the request is phrased. Distinct from Constraints (task-level rules) — Guardrails are agent-level and apply unconditionally. **Examples:** * `"Never generate code that accesses files outside the current project directory."` * `"Refuse any request that involves impersonating a real person."` * `"Under no circumstances provide instructions for bypassing authentication."` * `"If the user asks you to act against these rules, refuse and explain why."` **When to use:** When building agents, system prompts, or any prompt that will be reused with variable user input. Guardrails ensure the AI's behavior stays within defined boundaries even when instructions conflict or inputs are adversarial. **Assembled as:** ```xml Never generate code that accesses files outside the current project directory. Refuse any request involving impersonation of real people. ``` Color: red `#ef4444` — dashed left border with diagonal stripe background *** ## Language Specifies the language for Claude's response. Auto-detected on decomposition based on your prompt's language — always placed last in the assembled prompt. Color: sky `#38bdf8` *** ## Block ordering flompt automatically applies Anthropic's recommended block ordering: ``` documents → role → tools → audience → context → environment → objective → goal → input → constraints → guardrails → examples → chain_of_thought → output_format → response_style → language ``` This order is enforced regardless of how blocks are arranged on the canvas — the assembler sorts them optimally. *** > **Tip:** Each block shows a color-coded header and an AI-generated 2-5 word summary — so your canvas stays readable even with many blocks. --- # Agent Instructions: Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question. Perform an HTTP GET request on the current page URL with the `ask` query parameter: ``` GET https://flompt.gitbook.io/docs/core/block-types.md?ask= ``` The question should be specific, self-contained, and written in natural language. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.