Prompt engineering for AI and LLM models

This skill covers practical prompt engineering techniques and patterns for AI and LLM models, including named frameworks (RAIN, CLAR, FLOW, PIVO, SEED, RTF, RISE, RISEN, RHODES, etc.), chain-of-thought, few-shot prompting, role prompting, system prompt design, structured output, token optimization, and prompt injection mitigations. It includes model-specific guidance for Claude, GPT, Gemini, and open-source models, as well as tool-specific prompt shortcuts for Claude Code (slash commands, custom skills, plugins, and session management). The skill also covers writing system prompts for LLM-powered agents in automation platforms like n8n, including backend selection, tool enumeration, guardrails, memory configuration, and production reliability patterns. Broader context on the AI ecosystem — Anthropic's infrastructure, agentic email pipelines, Anthropic Academy, and the AI coding tools market — is included as supporting background. Always distinguish between model-level prompting (applicable in any interface) and tool-specific commands (applicable only in a particular IDE or CLI).

When to use this skill

• Apply this skill when the user asks about prompt engineering techniques (chain-of-thought, few-shot, system prompts, structured output, named prompt frameworks), model-specific prompting quirks across Claude/GPT/Gemini/open-source, token optimization strategies, prompt injection mitigations, community-validated prompt templates, tool-specific prompt shortcuts such as Claude Code slash commands and custom skills, or writing system prompts for LLM-powered agents in automation platforms (e.g., n8n AI Agent nodes). (source, source, source)

Critical rules

• DO NOT conflate Claude Code slash commands with prompting techniques for the Claude.ai web chat or API. Slash commands like /compact, /clear, /cost, /model, /plan are ONLY available in Claude Code (Anthropic's CLI coding agent). If a user references these in the context of claude.ai web chat, correct them immediately. (source, source, source)
• Not all built-in slash commands are visible to every user — some depend on platform or subscription plan. Before recommending a specific slash command, advise the user to type / in their Claude Code session to see available commands for their plan. (source, source)
• When recommending ANY named prompt framework (RAIN, CLAR, FLOW, PIVO, SEED, RTF, RISE, RISEN, RHODES, etc.), ALWAYS note that different models (GPT, Claude, Gemini) may respond better to different structural formats. Instruct the user to A/B test outputs and iterate — no single framework is universally optimal across all LLMs. (source, source, source)
• Place the task instruction or role assignment at the BEGINNING of the prompt, before contextual information or data. This aligns with Microsoft's prompt engineering research and is the structural pattern used by RAIN (Role first) and CLAR (Context first). DO NOT bury the core instruction deep in a wall of context. (source, source)

Structured prompt frameworks — selection guide

• Structured prompt frameworks work because they force the user to specify role, context, constraints, and success criteria BEFORE sending the prompt, giving the model all the information needed to deliver higher-quality results. The more detailed the instruction, the more likely the output meets expectations. When a user's prompt is vague or underperforming, recommend one of the frameworks below based on the task type rather than offering generic "be more specific" advice. (source, source, source, source)
• Use R-A-I-N (Role, Aim, Input, Numeric target, Format) for KPI-driven, measurable tasks — dashboards, conversion analyses, data summaries with specific targets. Template:

You are a [ROLE]. Aim: [AIM] Input: [DATA or reference to data] Numeric target: [TARGET, e.g., "increase CTR by 15%"] Format: [OUTPUT FORMAT, e.g., "CSV table + bullet recommendations"]

Tradeoff: highly effective for quantifiable deliverables but too rigid for exploratory or creative work. Pair with few-shot examples or a JSON schema to reduce hallucination on numeric claims. (source, source, source, source)

• Use C-L-A-R (Context, Limits, Action, Result) for constrained diagnostics, incident analysis, or business communication where hard boundaries matter (word count, audience, tone). Template:

Context: [SITUATION and background] Limits: [CONSTRAINTS — word count, audience, time, scope] Action: [WHAT the model should do] Result: [DESIRED output format and emotional/informational outcome]

Example: framing a company-wide announcement with a 200-word limit and a reassuring tone. (source, source, source)

• Use F-L-O-W (Function, Level, Output, Win metric) for audience-tailored content generation — blog posts, tutorials, marketing copy. The "Level" component targets the audience tier (beginner / expert / executive), and "Win metric" defines a measurable success signal (keyword density, CTA inclusion, readability score). Template:

Function: [WHAT the model should produce] Level: [TARGET audience tier] Output: [FORMAT — blog post, slide deck, email sequence] Win metric: [MEASURABLE success signal] (source, source, source)

• Use P-I-V-O (Problem, Insights, Voice, Outcome) for strategic planning, persuasive writing, or tasks where tone control is critical. The "Voice" component explicitly controls style — e.g., "confident, solution-focused" or "empathetic, data-driven". Template:

Problem: [PROBLEM statement] Insights: [KEY data, research, or context the model should incorporate] Voice: [TONE and style descriptors] Outcome: [DESIRED deliverable and its purpose] (source, source, source)

• Use S-E-E-D (Situation, End goal, Examples, Deliverables) for curriculum design, roadmaps, onboarding programs, or any multi-layered structured output that benefits from worked examples. Template:

Situation: [CURRENT state and audience] End goal: [WHAT success looks like] Examples: [SAMPLE outputs, reference formats, or few-shot examples] Deliverables: [EXACT artifacts to produce — e.g., "5-module syllabus + quiz bank"] (source, source, source)

• These five frameworks (RAIN, CLAR, FLOW, PIVO, SEED) are NOT the only named prompt frameworks. They belong to a broader landscape that includes RTF (Role, Task, Format), RISE (Role, Input, Steps, Expectation), RISEN (Role, Input, Steps, Expectation, Novelty), RHODES (Role, Objective, Details, Examples, Sense Check), and others. If the user's task does not map cleanly to one of the five above, check whether RTF, RISE, or a simpler role+task+format pattern fits better. (source, source)

Applying prompt frameworks effectively

• When filling any framework template, follow these steps: (1) Copy the framework skeleton into the system or user message. (2) Fill EVERY slot with concrete, specific values — DO NOT leave placeholders or vague descriptions. (3) Include at least one few-shot example or a desired output schema (JSON, CSV, markdown table) to anchor the model's format. (4) Send the prompt, evaluate the output, then iterate: adjust constraints, add examples, or switch frameworks if the output type does not match. (5) For production prompts, version them and run A/B tests or evals to measure quality changes. (source, source, source)
• Role prompting — assigning the AI a specific persona or expert role — is the single most common element across RAIN, FLOW, PIVO, SEED, RTF, RISE, and RISEN. When a user is new to prompt engineering, start with role prompting before introducing a full framework. Place the role in the system message when using an API, or lead with "You are a [ROLE]" in a chat interface. (source, source, source)
• For API-based usage (OpenAI, Anthropic, Gemini APIs): place role assignments and persistent instructions in the system message; place framework-filled templates and few-shot examples in the user message. Request structured output (JSON mode or a JSON schema) to reduce hallucination and enforce format compliance. Version your prompt strings for reproducibility. (source, source)

Claude Code slash commands reference

• Use /compact [focus instructions] to compress session history and free context tokens. Pass optional instructions to preserve specific context, e.g. /compact keep the database schema and unit tests. Tradeoff: compaction extends session length but MAY drop older context and reduce fidelity of very old details or previously invoked skills. Use this proactively before hitting token limits on long sessions. (source, source, source)
• Use /clear to wipe the conversation history and context window, effectively starting a fresh session without restarting the CLI. (source, source)
• Use /cost to check token usage and spend for the current session. Output detail varies by subscription plan. (source, source, source)
• Use /model to switch the active Claude model (e.g., Sonnet, Haiku, Opus) mid-session without restarting. (source, source)
• Use /plan [description] to enter plan mode, which instructs Claude to outline steps before executing. Plan mode can also be activated with --plan at startup or toggled mid-session with Shift+Tab. (source, source)
• Use /memory to open CLAUDE.md memory files for editing. Use /init to initialize a project with a CLAUDE.md guide if one does not exist. (source, source, source)
• Use /mcp to manage MCP server connections and OAuth authentication. Use /permissions to manage allow, ask, and deny rules for tool access. Use /agents to manage AI sub-agent configurations. (source, source, source)
• Use /usage to check plan usage limits and current subscription status (distinct from /cost which shows session-level token counts). (source, source)

Claude Code session management

• Run /rename my-session-name to set a human-readable session title (omit the name to auto-generate one). Resume later with /resume <name> in-session or claude --resume <name> from the CLI. Best practice: ALWAYS rename important sessions before running /clear or heavy compaction so you can reliably resume them. (source, source)
• Use --add-dir <path> at startup or /add-dir <path> during a session to grant Claude file access in additional folders. Note: added directories grant file access but do NOT become full configuration roots — however, skills in .claude/skills/ ARE loaded from added directories. (source, source, source)

Claude Code custom skills and plugins

• Create reusable prompt shortcuts (custom slash commands) by adding a file at ~/.claude/skills/<skill-name>/SKILL.md with YAML frontmatter (name, description, allowed-tools, etc.). Use $ARGUMENTS or $ARGUMENTS[N] for positional parameters. Example skeleton:

name: explain-code description: Explain this file with diagrams.

Explain $ARGUMENTS.

This makes /explain-code <args> available in all sessions. Use /skills to list available skills. (source, source)

• Use /plugin to open the interactive plugin manager. Add marketplaces with /plugin marketplace add anthropics/claude-code, install plugins with claude plugin install <name>@<marketplace>, then run /reload-plugins to apply changes without restarting the session. Plugins can extend Claude Code with custom commands, agents, hooks, skills, and MCP servers. (source, source, source)
• Slash commands and skills accept inline arguments and can be invoked mid-message — the / autocomplete can trigger at the cursor position, not strictly at the beginning of a message. As of 2026-04-13, verify this behavior against the latest Claude Code changelog as autocomplete mechanics evolve between releases. (source, source) (medium confidence)

If the user asks about prompt frameworks, do this

• If the user asks "which prompt framework should I use?", ask what type of task they are doing, then recommend: • Measurable/KPI-driven tasks → RAIN • Constrained diagnostics or business comms → CLAR • Audience-tailored content creation → FLOW • Persuasion, strategy, or tone-sensitive writing → PIVO • Curriculum, roadmaps, or multi-layered programs → SEED • Simple role+task+format needs → RTF (lighter-weight alternative) Provide the matching template from the "Structured prompt frameworks" section and walk the user through filling each slot. (source, source, source)
• If the user's task is exploratory, creative brainstorming, or open-ended ideation, DO NOT recommend RAIN — its rigid numeric-target structure constrains creative output. Suggest PIVO (for voice/tone control) or a simple role+task prompt instead. For open-ended research, chain-of-thought prompting without a framework often outperforms template-based approaches. (source, source)

Writing system prompts for n8n AI agents

• n8n's AI Agent node accepts a system prompt that defines the agent's behavior, persona, and tool-usage rules — the same prompt engineering principles (role assignment, structured constraints, few-shot examples) apply here as in any LLM system message. When a user asks how to write a prompt for an n8n agent, treat it as a system prompt design task: place role + behavioral rules in the system prompt field of the AI Agent node, and use any of the named frameworks (RAIN, CLAR, PIVO, etc.) to structure the instructions. Using ChatGPT or Claude to draft the system prompt before pasting it into n8n is a documented, effective workflow. (source, source, source)
• n8n's AI Agent node supports multiple LLM backends: OpenAI, Anthropic Claude, Google Gemini, Groq, DeepSeek, Mistral, and Azure OpenAI. When writing a system prompt for an n8n agent, tailor the prompt style to the selected backend — Claude responds well to XML-structured instructions, GPT to numbered lists, Gemini to concise directives. The same A/B-test-across-models advice from the frameworks section applies here. (source, source)
• n8n agents are action-taking, not just text-generating — they can query APIs, update CRMs, send emails, and file reports autonomously. When crafting the system prompt, EXPLICITLY enumerate which tools the agent may use and under what conditions. Include guardrails in the prompt such as "DO NOT send emails without user confirmation" to prevent unintended side effects. n8n also supports human-in-the-loop approval (via Chat, Slack, or Telegram) before executing sensitive tool actions — instruct the user to enable this for any agent that modifies external systems. (source, source, source)

Warnings — n8n AI agent prompting pitfalls

• DO NOT tell users that n8n agents have persistent memory out of the box. n8n's memory sub-node does NOT persist between sessions by default — seamless stateful agents require explicit configuration with a Redis or Postgres backend. If a user's system prompt assumes the agent remembers prior conversations, warn them to configure a persistent memory store or design the prompt to be self-contained per invocation. (source, source)
• DO NOT claim that n8n AI agents bypass LangChain. Under the hood, n8n implements its AI Agent functionality through a hierarchical node system built on the LangChain JavaScript framework. If a user asks whether n8n uses LangChain, confirm that it does — despite marketing claims to the contrary. This matters for prompt engineering because LangChain's agent executor has its own prompt template wrapping the user's system prompt, and debugging may require understanding that layer. (source, source)
• DO NOT imply that a single ChatGPT-generated prompt will produce a fully functional n8n agent with zero manual configuration. Credential setup (API keys for OpenAI, Claude, Gmail, Airtable, etc.) MUST be configured manually per service in n8n's credential store. Workflow wiring — connecting trigger nodes, tool nodes, and the AI Agent node — requires understanding n8n's visual builder, which has a meaningful learning curve. Advise users that the prompt is only ONE piece; they also need to configure triggers (Webhook, Chat, schedule), connect tool nodes, set up credentials, and configure error handling. (source, source, source)
• For production n8n agents, configure Error Workflows and enable 'Retry on Fail' on LLM-calling nodes. LLM API calls are inherently unreliable (rate limits, timeouts, model errors). Include in the system prompt instructions for graceful degradation — e.g., "If you cannot complete the task, respond with a structured error message rather than guessing." This pairs the prompt-level safeguard with n8n's infrastructure-level retry logic. (source, source)

If the user asks about n8n AI agents, do this

• If the user asks "how do I build an AI agent in n8n?", guide them through these steps: (1) Choose a trigger node (Webhook, 'On Chat Message', schedule, or platform-specific trigger). (2) Add an AI Agent node and select the LLM backend (OpenAI, Claude, Gemini, etc.). (3) Write a system prompt using any of the frameworks in this skill — CLAR works well for constrained agent tasks, PIVO for tone-sensitive agents. (4) Connect tool nodes for the actions the agent should perform (send email, query database, update CRM). (5) Configure credentials for each connected service. (6) Enable human-in-the-loop approval for any destructive or external-facing tool. (7) Set up Error Workflows and Retry on Fail. (8) For persistent memory, add a Redis or Postgres memory sub-node. Remind the user that n8n self-hosted is free with unlimited workflows; n8n Cloud starts at ~$20/month as of 2026-04-17. (source, source, source, source, source)

Gemini Deep Research Agent

• Google's Gemini Deep Research agent is available exclusively via the Interactions API (not through generate_content), accessible using a Gemini API key from Google AI Studio. You start a research task by calling the Interactions endpoint with agent set to a Deep Research preview model (e.g. 'deep-research-preview-04-2026' or 'deep-research-pro-preview-12-2025'), set background=true to run asynchronously (max 60 minutes, most tasks complete within 20 minutes), and optionally set stream=true for real-time updates including intermediate image output. The agent is powered by Gemini 3.1 Pro and achieves benchmark scores of 66.1% on DeepSearchQA (strict success rate), 46.4% on Humanity's Last Exam (HLE), and 59.2% on BrowseComp. It supports deep web navigation, analysis of user-uploaded documents (PDFs, CSVs, docs) via File Upload and File Search Tool, granular citations, and structured JSON schema outputs. Native chart/infographic generation can be enabled by setting agent_config.visualization to 'auto' in the Interactions request. Remote MCP server integration is also supported via a 'mcp_server' tool entry (type: 'mcp_server' with name, url, headers) in the tools list of the interactions.create payload. Google also open-sourced the DeepSearchQA benchmark with 900 hand-crafted tasks across 17 fields to evaluate agents on multi-step web research comprehensiveness. (source, source, source, source, source, source, source, source, source, source, source)
• As of April 21, 2026, Gemini for Google Home began rolling out 'Continued Conversation' to early access users, allowing natural back-and-forth exchanges without repeating 'Hey Google' for each follow-up. Activate it via 'Hey Google, let's chat' (Gemini Live mode); the mic stays open ~15 seconds after each reply (device lights pulse to signal the listening window), Gemini retains context across the session, and smarter side-talk detection reduces accidental activations. The feature must be manually enabled in the Google Home app under 'Gemini for Home voice assistant' settings, is rolling out across all supported languages and regions, and some extended Live capabilities require a Google Home Premium subscription. (source, source, source, source, source, source)

Game Generation Prompting

• OpenGame is an open-source agentic framework from CUHK Multimedia Lab (released April 21, 2026) that generates end-to-end playable web games from a single natural-language prompt, outputting a playable index.html via a CLI-driven pipeline. It is powered by GameCoder-27B — a 27B code LLM trained via continual pre-training, supervised fine-tuning, and execution-grounded reinforcement learning — and uses a 'Game Skill' module comprising a Template Skill (project skeleton library) and a Debug Skill (verified-fix protocol) to manage cross-file consistency. Games are evaluated by OpenGame-Bench, which scores Build Health, Visual Usability, and Intent Alignment using headless browser execution and VLM judging across 150 diverse prompts. (source, source, source, source, source)

GPT Image 2 Prompting

• OpenAI released ChatGPT Images 2.0 on April 21, 2026, powered by the new gpt-image-2 API model, available to all ChatGPT and Codex users with the underlying model accessible via the API for developers. It introduces a 'thinking mode' (reserved for Plus, Pro, Business, and Enterprise subscribers) that combines reasoning, web search, and iterative prompt refinement to generate up to eight consistent images from a single prompt, at resolutions up to 2K and flexible aspect ratios from 3:1 to 1:3. The base Images 2.0 is available to all ChatGPT tiers including Free. Key capabilities include improved rendering of small text, iconography, UI elements, dense compositions, and multilingual non-Latin text (Japanese, Korean, Chinese, Hindi, Bengali), with a reported 99% typography accuracy rate and generation speed roughly twice that of its predecessor. The model has a knowledge cutoff of December 2025 and can use the web to research up-to-date information before generating images. Use cases highlighted by OpenAI and reviewers include scientific diagrams, infographics (e.g., Cantor diagonalization), lecture-hall math proofs, menus, multi-panel comics, and posters. Independent testing notes that non-Latin text rendering, while improved, can still produce malformed or semi-garbled output in some languages. OpenAI describes the architecture as 'revamped from scratch' — a 'generalist model' or 'GPT for images' capable of 3D-style perspective shifts and complex spatial reasoning — and positions it as a 'step change' for following instructions, rendering dense text, and placing/relating objects in a scene. No dedicated technical paper has been published as of the announcement date. The model includes next-generation watermarking and provenance classifiers for enterprise use; OpenAI notes that higher realism increases misuse risk (e.g., more convincing deepfakes). API pricing slightly undercuts gpt-image-1.5 on the output side, with tiered pricing for different quality/resolution outputs. OpenAI is deprecating GPT-Image-1.5 as the default model, though it remains accessible via the API for legacy support. (source, source, source, source, source, source, source, source, source, source, source, source, source, source, source, source)

OpenClaw Gateway Prompting

• OpenClaw v2026.4.20 strengthens system prompt adherence and introduces several actionable prompt-engineering behaviors: (1) set agent system prompts via agent config or per-group systemPrompt so the upgraded SOUL/identity overlay applies to new turns; (2) prefix assistant output with the exact token NO_REPLY (case-insensitive) to suppress delivery on silent/background turns — v2026.4.20 also fixes silent-turn handling so DM chats can receive rewritten replies while groups stay quiet; (3) run /new or /reset (optionally /new <model>) to mint a fresh session ID and clear stale auto-pinned model/provider overrides before each turn; (4) after upgrading, review session.maintenance settings (pruneAfter, maxEntries) and run openclaw sessions to inspect retention and prevent unbounded session growth in long-running gateways. (source, source, source, source, source, source, source) (medium confidence)
• Hermes Agent (Nous Research) is an open-source MIT-licensed AI agent (100k+ GitHub stars) with a full Terminal UI supporting slash-command autocomplete, conversation history, and streaming tool output. It connects to a wide range of providers — OpenRouter, OpenAI, Anthropic, Gemini, Ollama, vLLM, LM Studio, llama.cpp, and more — with no code changes needed to switch, and supports local model endpoints (e.g., Ollama at http://127.0.0.1:11434/v1) so no external API key is required. Atomic Bot provides a native macOS front-end (open-source, one-click, local or cloud) for running Hermes alongside OpenClaw, featuring a unified workspace with chat, terminal, memory/skills, logs, and file explorer. (source, source, source, source, source, source, source)

AI Coding Tool Landscape

• SpaceX (via SpaceXAI, following Elon Musk's February 2026 merger of SpaceX and xAI valued at $1.25 trillion) and Cursor announced a partnership to build next-generation AI models focused on coding and knowledge work, with SpaceX securing the contractual right to acquire Cursor for $60 billion later in 2026 or alternatively pay $10 billion for their collaborative work. The partnership leverages xAI's Colossus supercomputer (equivalent to one million Nvidia H100 chips) to accelerate Cursor's model training; prior to this announcement, xAI had already agreed to rent tens of thousands of GPUs to Cursor to train its latest model, Composer 2.5. Cursor CEO Michael Truell confirmed the deal on X, describing it as 'a meaningful step on our path to build the best place to code with AI.' Two senior Cursor engineering leaders, Andrew Milich and Jason Ginsberg, left Cursor to join xAI reporting directly to Musk ahead of the announcement. Neither Cursor nor xAI currently has proprietary models matching leading offerings from Anthropic and OpenAI — Cursor still uses and sells access to Claude and GPT models even as those firms roll out competing coding tools — and this deal is seen as an effort to close that gap. Separately, Cursor was reported to be in talks to raise a $2 billion funding round at over a $50 billion valuation, co-led by Andreessen Horowitz. (source, source, source, source, source, source, source, source, source, source, source)

Claude Automation Runtimes

• Anthropic offers three complementary automation runtimes with distinct tradeoffs: (1) Claude Cowork scheduled tasks (launched Feb 25, 2026) run locally in the Claude Desktop app, can access local files (xlsx, pptx, docx, pdf), but only fire while the computer is awake and the app is open. (2) Claude Code Routines (launched April 14, 2026 as a research preview) run on Anthropic's cloud infrastructure so tasks survive a closed laptop, support scheduled, API/webhook, and GitHub-event triggers, and are capped at 5/day (Pro), 15/day (Max), or 25/day (Team/Enterprise). (3) Claude Managed Agents (public beta April 8, 2026) is a fully managed cloud runtime providing sandboxed execution, session tracing, credential management, and scoped permissions — early adopters include Notion, Rakuten, Asana, and Sentry, with Rakuten reporting a shift from quarterly to biweekly releases and a 97% drop in critical errors. (source, source, source, source, source, source, source, source, source)

Last updated: 2026-04-21T07:52:47.468Z