The quality of the answer depends largely on how you formulate your request. A vague prompt like "Tell me about AI" can lead to a generic or superficial response. A specific prompt like "Explain the main differences between supervised and unsupervised learning in 3–4 paragraphs for a beginner" gives the model a clear goal and format.

Always add context: who you are, what you need the answer for, and what level of detail you expect. If you're a developer, say so. If you need a quick summary for a presentation, say so. Context helps the model tailor the response.

Specify the desired format when it matters: a list, a table, step-by-step instructions, or a short paragraph. This reduces the need for follow-up questions and saves time.

Avoid overly broad or multi-part questions in a single message. If you have several questions, it's better to ask them one by one or explicitly list them.

Use action verbs: "Explain", "Compare", "List", "Summarize", "Write", "Analyze". They help the model understand the type of response you expect. "Explain why X happens" is clearer than "I want to know about X."

For technical or specialized topics, mention your level: beginner, intermediate, or expert. This affects the depth and terminology of the answer. A beginner explanation avoids jargon; an expert one can go into implementation details.

You don't have to get the perfect answer on the first try. Neural networks are designed for dialogue: you can ask follow-up questions, refine the answer, or ask for corrections.

Use the conversation history: the model remembers the context. If the answer was too long, ask for a shorter version. If something was unclear, ask for clarification. If you need more detail on a specific point, ask about it.

• Break complex tasks into smaller steps — ask for one thing at a time.
• Use Edit and Rewrite to continue or correct the conversation if something went wrong.
• Don't hesitate to say "I meant something else" or "Focus on X instead of Y" — the model will adjust.
• For long answers, ask to expand a specific section: "Go deeper into point 2" or "Explain the third paragraph in more detail."

If the model misunderstood or gave an irrelevant answer, reformulate your request more clearly rather than repeating the same question.

The model has a context window: it "sees" the last several messages. Very long conversations may lose early context. For a completely new topic, consider starting a fresh chat.

Iskra offers several modes that determine whether the model uses web search and how deeply it researches. Choosing the right mode saves time and improves results.

In Search and Deep modes you can choose sources: Web, Academic papers, Discussions. Combine them depending on the topic — for scientific questions add Academic, for opinions use Discussions.

Deep mode consumes more tokens and takes longer — use it when you really need a thorough analysis. For quick facts or simple comparisons, Search or Auto is enough. Chat mode is fastest and doesn't use search at all.

You can attach PDF, DOCX, XLSX, and images to your messages. The AI reads the content and answers questions about it. Specify which file or section you're asking about if you have several attachments.

For PDF and DOCX, ask for summaries, key points, or specific information. For XLSX, you can ask to edit cells, add rows, create tables, or perform calculations. The AI will use special tools to work with spreadsheets.

When asking about a file, give context: "In the attached report, what are the main conclusions?" or "Add a row to the table with the following data." The clearer the request, the more accurate the result.

Large files may take longer to process. For very long documents, ask about specific sections or chapters. "Summarize chapter 3" works better than "Summarize the entire 200-page document" in one go.

For XLSX, the AI can read, edit cells, add rows, and finalize changes. Describe the edits clearly: "Change cell B5 to 42" or "Add a row at the end with Name: John, Age: 30." After edits, the model will offer to save the file.

Iskra lets you choose a role before sending a message: Teacher, Programmer, Writer, Marketer, SMM specialist, Content Creator, Data Analyst, and others. The role sets the expertise and style the model adopts — it will answer as that type of specialist.

Teacher: step-by-step explanations, examples, analogies. Programmer: technical accuracy, code examples, best practices. Writer: attention to style and structure. Marketer and SMM: audience, value proposition, channels, and social content; for ads and integrations, answers lean on retrieved sources for rules and verifiable facts (not a substitute for legal advice). Content Creator: scripts, hooks, native integrations. Data Analyst: facts, metrics, structured analysis.

The model also remembers your name and preferences if you share them. "My name is Alex, I prefer short answers" — such context improves responses over time. Use the same chat for related topics to keep this context.

In Search and Deep modes, the AI finds real sources and cites them. Click the source links to verify. For facts, dates, statistics, or news, always prefer modes with search — the model's training data has a cutoff date.

Even with search, the model can make mistakes: misread a source, mix up similar concepts, or overgeneralize. For critical decisions (health, law, finance), cross-check important claims with primary sources.

Academic mode searches scientific papers and publications. Use it for research, literature reviews, or when you need peer-reviewed information. Web mode is better for news, tutorials, and general articles.

If the answer cites a source you can't access or that seems wrong, ask: "What exactly does source X say?" or "Find another source that confirms this." The model can refine its research.

For creative writing — stories, emails, posts — describe the tone (formal, casual, humorous), length, and key points. "Write a 3-paragraph email declining an invitation politely" works better than "Write an email."

For code, specify language, framework, and constraints. "Python function to parse JSON, handle errors" or "React component with useState for a counter." Include relevant code if you're modifying something: paste it and say what to change.

For brainstorming, start broad and narrow down. "Give me 10 startup ideas in edtech" → pick one → "Expand idea 3 into a one-pager." Iterative refinement works well for open-ended tasks.

Chat mode (no search) is often best for creative and coding tasks — you don't need external sources, and the model can focus on generation. Use Search or Deep when you need up-to-date APIs, libraries, or references.

Even experienced users sometimes make these mistakes. Avoiding them will significantly improve the quality of responses.

• Too general: "Tell me about AI" — add what exactly you need: applications, history, risks, or something else.
• Missing context: "Fix this" — explain what and why. The model doesn't see your screen or know what "this" refers to.
• No format: "Summarize" — specify length (3 sentences, 1 paragraph), or structure (bullets, numbered list).
• Ignoring history: for related questions, use follow-ups instead of starting a new chat. The model uses the conversation context.
• Asking for everything at once: "Write a business plan, marketing strategy, and financial forecast" — split into separate requests or ask for one first, then build on it.

If the answer is wrong or irrelevant, don't assume the model "doesn't understand." Try reformulating, adding context, or breaking the question into smaller parts.

Neural networks can hallucinate — invent facts or sources. For critical decisions, verify important information. In Search and Deep modes, check the cited sources.