HuggingChat Prompts

I want to benchmark open models on my real task: [DESCRIBE TASK, e.g. summarizing customer tickets, writing SQL]. My evaluation criteria: [CRITERION 1], [CRITERION 2], [CRITERION 3]. Design a test: 5 representative inputs ([INPUT 1]...), the exact prompt to use for each model, and a scoring sheet 1 to 5 per criterion. Recommend which HuggingChat models to include for this task ([MODEL FAMILY HINTS, e.g. a Llama, a Mistral, a Qwen, a DeepSeek variant]) and why each is a good candidate. Include the Omni router as a baseline entry.

Help me design a HuggingChat custom Assistant for [USE CASE, e.g. onboarding new hires, answering product FAQs]. Define: the system prompt (role, tone [TONE], scope boundaries, what it must never do: [RESTRICTION]), the knowledge base documents to attach ([DOC 1], [DOC 2]), which tools it needs ([WEB SEARCH, CODE EXECUTION, IMAGE GEN]), 5 test questions with ideal answers, and a shareable description for my team at [TEAM OR COMPANY]. Make the system prompt robust against prompt injection attempts like [EXAMPLE ATTACK].

I want to build a RAG powered Assistant over these documents: [LIST DOCUMENTS, e.g. 40 help center articles, the employee handbook]. Walk me through: how to chunk and upload them, what chunk size and overlap to use for [DOCUMENT TYPE], 10 test questions covering edge cases ([EDGE CASE 1], [EDGE CASE 2]), how to evaluate answer grounding (quote vs paraphrase), and a maintenance plan for when documents update [FREQUENCY]. Warn me about the failure modes of RAG on [MY DOCUMENT TYPE] and how to detect them.

I have uploaded a CSV about [DATASET TOPIC] with columns [COLUMN 1], [COLUMN 2], [COLUMN 3]. Using the code execution tool, analyze it for me: load and describe the data (rows, missing values, types), compute summary statistics for [METRIC], plot [CHART TYPE] showing [RELATIONSHIP], test the hypothesis that [HYPOTHESIS], and report the 3 most surprising findings. Show the Python code you ran for each step so I can reproduce it locally. If the file is too large, sample it with [SAMPLING STRATEGY].

Using the image generation tool, create [NUMBER] concept images for [PROJECT, e.g. a landing page hero, a book cover]. Brief: subject [SUBJECT], style [STYLE, e.g. cinematic, minimal line art], palette [COLORS], composition [NOTES], aspect [RATIO]. For each concept give the exact prompt you used so I can regenerate variations. Then critique the set: which best serves [GOAL], what to fix in round 2, and 3 negative prompt additions to avoid [ARTIFACT, e.g. warped text, extra fingers].

I uploaded [FILE TYPE: PDF or spreadsheet] containing [DESCRIPTION, e.g. 200 invoices, a research dataset]. Parse it and extract into a clean table with columns: [COLUMN 1], [COLUMN 2], [COLUMN 3]. Rules: normalize dates to [FORMAT], flag rows with missing or ambiguous values instead of guessing, and give me a data quality report (completeness per column, duplicates found, outliers in [NUMERIC COLUMN]). Then answer: [ANALYSIS QUESTION 1], [ANALYSIS QUESTION 2] based on the extracted data.

Using web browsing, brief me on the latest developments in [TOPIC] over the past [TIMEFRAME]. I need: the 5 most important stories with source links and dates, what changed versus [EARLIER PERIOD], 2 expert opinions that disagree (with attribution), and what to watch next ([UPCOMING EVENT]). My background: [YOUR ROLE], so focus on implications for [YOUR DOMAIN]. Cite every factual claim with its source; if browsing returns thin results, say so instead of filling gaps from training data.

I want maximum privacy for discussing [TOPIC]. Explain HuggingChat's local browser inference option (WebLLM): which smaller models can run fully in my browser, what the quality and speed tradeoffs are versus server models for [MY TASK], hardware requirements ([RAM, GPU] guidance), and exactly what data leaves my machine (nothing, if local). Then help me pick between local inference and the privacy mode server option for my threat model: [DESCRIBE WHAT YOU ARE PROTECTING].

I will dictate my meeting notes by voice about [MEETING TOPIC] with [ATTENDEES]. After I finish dictating, do the following: clean up the transcription (fix obvious speech to text errors, keep [JARGON TERMS] intact), structure it into decisions made, action items with owners ([OWNER 1], [OWNER 2]) and deadlines, open questions, and a 3 sentence summary for [STAKEHOLDER]. Ask me clarifying questions about anything ambiguous before finalizing. Tell me how to use Whisper voice input for the dictation step.

I am uploading an image of [IMAGE DESCRIPTION, e.g. a chart, a product photo, a UI screenshot]. Analyze it thoroughly: describe what you see objectively first, then interpret it for my goal ([GOAL, e.g. checking chart accuracy, reviewing design]). Point out: [ASPECT 1 to check], [ASPECT 2 to check], anything anomalous or suspicious, and what additional image ([ANGLE, CLOSE UP]) would help you analyze better. If it is a chart, extract the data values you can read and note where the image is too low resolution to be certain.