# Beyond text: images, audio, small apps and data analysis (/docs/prompting/beyond-text) --- type: Procedure title: "Beyond text: images, audio, small apps and data analysis" description: "Concepts 8 to 10 of the book's prompting guide: what AI sees well and badly in an image, how to prompt image generation and its failure modes, the diagram recipe, audio in and out, the three-slot recipe for a one-prompt app and where it runs, and how to make the model actually run code on your data instead of guessing." status: stable order: 6 generated: { by: "process:claude-code", at: 2026-09-03T08:15:00Z } verified: - { by: "human:rrizwan98", at: 2026-09-03T08:20:00Z } sources: - id: af-03 title: "AI Prompting 2026: 13 Concepts, 80% of Real Use. The AI Agent Factory, Panaversity, 2026" resource: https://agentfactory.panaversity.org/docs/ai-prompting-2026 ksor: audience: [public] owner: human:rrizwan98 approval: { by: "human:rrizwan98", at: 2026-09-03T08:20:00Z } trust_tier: human-reviewed build_id: sha256:4d4bff09bd0a1dcff0a0a50f20f7ef514fc471e2b1f9e81e0e7b6150b93c0793 dirty: true --- AI is not just a text box. It can see images, work with audio in both directions, build small working apps, and run code on your data. Most people never try any of it. [^af-03] ## Concept 8: multimodal, images and audio **Image input.** AI sees images coarsely. Strong on: overall scene and composition; distinct, large object shapes; whiteboard contents, including diagrams; handwritten and cursive text (decent; double-check for high stakes). Weak on: fine details ("what gym machines are these?" tends to fail, confidently); counting many small things in a cluttered scene; small print at the edge of an image. A teacher photographed a whiteboard where his head blocked the word "convolutional" in a diagram; the AI inferred the missing word correctly from the rest. That is what it is good at: inferring from the gist. It is not good at zooming in. For receipts, splitting a bill, or transcribing handwritten notes it works well; always double-check the totals. [^af-03] **Image output.** Two practical tips: **use a text AI to write your image prompt** ("generate me a prompt for a fantasy forest illustration in a Studio Ghibli style for a children's book cover"), then paste that into the image tool; and **build visual vocabulary**: cinematic, watercolor, cyberpunk, anime, isometric, low-poly, art-deco, claymation are levers, because image models learned those styles by name. Upload images you like and ask the AI how it would describe them. [^af-03] Image generation is a **diffusion model**, trained to remove noise from a random pixel grid step by step until an image emerges, the whole image at once; that is why you cannot stop it early the way you can interrupt a text response. Failure modes still worth watching for: [^af-03] | Failure mode | What it looks like | How to mitigate | | --- | --- | --- | | Garbled text on signs | "HAPRY BIRTDAY" | Specify the text in quotes; generate three variants; pick the right one | | Inconsistent characters across frames | Different hair colour in panels 1 and 2 | Use models with character-consistency support; pass the first image back as a reference | | Hand and finger errors | Six fingers, fused hands | Ask for compositions where hands are out of frame, in pockets, or clearly described | | Cluttered backgrounds with implausible objects | A bicycle merging into a chair | Specify a simple background, or describe it explicitly | | Wrong aspect ratio | Square when you wanted landscape | Always specify it: "1024x768 landscape" or "16:9" | A reader uploaded three handwritten recipe cards from a deceased grandmother: "Transcribe these three cards. Preserve the original wording and any abbreviations. If a word is unclear, mark it `[unclear]` and offer your two best guesses." Five minutes later all three were typed, with four `[unclear]` marks the reader checked by hand. AI did the boring 90% so the reader could focus on the careful 10%. [^af-03] ### Designer-quality diagrams without a designer A four-step recipe that produces designer-quality diagrams in about fifteen minutes, without a design tool: [^af-03] 1. **Ask Claude to visualize the concept as SVG.** Paste the underlying paragraph: "Visualize this as a diagram. Output it as SVG. Make sure every label, arrow, and relationship from the text is present." Deciding what belongs in a diagram is a reasoning task; the SVG will be structurally correct but visually plain. 2. **Convert the SVG to PNG**, at about 2x resolution (1600 to 2400 pixels wide), via the model, an online converter, or a high-zoom screenshot. 3. **Paste the PNG into ChatGPT or Gemini and ask it to redraw:** "Redraw this diagram with professional design quality. Preserve every label, every box, every arrow, and the exact structural relationships. Improve typography, spacing, color palette, and visual hierarchy. The information must remain identical; only the visual finish changes." 4. **Iterate.** Compare against the original SVG; type corrections ("the third box should be labeled 'Iterate', not 'Repeat'"). Three or four rounds. The leaders in each category will rotate, so the recipe goes stale at the tool layer. What survives: **structure first in the strongest reasoning model, polish second in the strongest text-heavy image model.** [^af-03] **Audio in, audio out.** Dictate a long prompt instead of typing it; drop in a meeting recording and ask for a summary; ask the model to read its answer aloud. The non-obvious uses: **long-form dictation** (people who hate typing produce dramatically better prompts when they speak them; the prompt grows from one line to several paragraphs, and the answer improves accordingly); **meeting transcripts as context** ("Summarize the decisions made, the open questions, and the action items by owner", one of the highest-leverage workflows for anyone with meetings); and **audio for movement**, turning a commute or a walk into thinking time. [^af-03] | Audio task | How well it works | Watch out for | | --- | --- | --- | | Transcription of clear speech | Excellent | Heavy accents, technical jargon, overlapping speakers | | Speaker identification | Decent on 2 speakers, weak on 4+ | Always check before quoting someone | | Tone, sarcasm, emotion | Improving but unreliable | Ask the AI to flag its uncertainty rather than assume | | Music or non-speech audio | Limited | Use a specialized tool | | Real-time voice conversation | Good for casual, weak for technical depth | Switch to text when precision matters | A doctor recorded a 45-minute consultation (with consent) and asked for a structured clinical note in SOAP format, flagging anything not understood confidently. Eight minutes later she had a draft that took 5 minutes to verify instead of the 25 minutes the typed version would have taken. The AI did not replace clinical judgment; it removed the typing. The future of multimodal is not "AI can do voice now": it is that the boundary between modalities disappears, and the skill becomes "what is the right combination of inputs for this job?" [^af-03] ## Concept 9: building small apps with one prompt Modern AI can build small games, websites and tools from a single prompt. Not large software; for small useful things, this is genuinely accessible to people who have never written code. All three major tools render the app in a side panel, and the thing in that panel is an **artifact**: a persistent object you can edit, iterate on, publish to a shareable link, embed, or download as code. Claude calls it Artifacts, ChatGPT and Gemini call it Canvas. You can hand the artifact to someone without sending them the chat, and when you say "make the button bigger" the tool edits it in place. For anything beyond a one-prompt build, three adjacent categories exist: dedicated AI app-builders (v0, Bolt, Lovable), command-line coding agents (Claude Code, OpenCode), and file-aware desktop apps (Cowork, OpenWork). [^af-03] The recipe is three slots: [^af-03] ```text Goal: what should this thing do? Input: what does the user provide? Output: what does the user see? ``` Examples that work today: a Pomodoro timer with a yellow theme and a click at each cycle's end; a bill splitter; an outfit picker from today's weather; a fireworks simulator on click; a place-obstacles game with a simulation. What is still hard: multiplayer over the internet (networking, accounts, matchmaking), and live AI feedback in a different language. The intuition: small things that fit on one screen, with no accounts and no external services, work. [^af-03] A parent's three-sentence prompt for a typing game: [^af-03] ```text Build a typing game for a 7-year-old. Goal: practice typing common short words. Input: words appear, the player types them before they reach the bottom of the screen. Output: a yellow theme, a cute cat mascot that cheers when the player gets a word right, increasing speed across levels. ``` It worked, iterated to "good enough for a kid" inside an hour. The skill is not coding. It is the ability to write a clear brief and iterate it. [^af-03] ## Concept 10: data analysis, the model writes and runs code When a question needs calculation or graphing, modern tools quietly write code, run it, and return the result. Code execution is just another tool the model can call, like web search, and it is far more reliable than the model doing math in its head: the calculator is precise; the model is choosing what to compute. [^af-03] **Before anything else: make sure the AI actually runs code instead of guessing.** The AI does not automatically run code on every question; it chooses to, based on how the question is phrased. On smaller questions it sometimes answers from a glance, a confident paragraph with no computation behind it that looks identical to a real analysis. Three habits: [^af-03] 1. **Ask explicitly.** "Write and run code to answer this. Show me the code you ran." 2. **Check that the code is visibly there.** No code block that ran, probably no code. 3. **Demand a verifiable specific before the analysis.** "Tell me the exact row count, the column names, and the date range of this file before you analyze anything." If it is making things up, the row count will be a suspiciously round number and the column names plausible-but-wrong. The strongest version: ask the model to declare its method up front. "Are you running code on the file, or estimating? If estimating, stop and run code instead." [^af-03] The bubble tea shop example: a year of sales data, and "Which drinks had the biggest changes in sales over the year? Graph them. Write and run code to answer this and show me the code you ran." The AI computes month-over-month changes per drink, sees most are flat and four stand out, graphs those four, and notes "Strawberry matcha rose sharply in spring; consider re-running that promotion next year." Grounded in the actual data, not generic. [^af-03] **Curate what goes in before you ask anything.** The quality ceiling of an analysis is set by what you upload. De-duplicate near-identical files; label each input when you attach more than one ("sales-2025.csv is the full year; q4.xlsx is a subset of it, ignore it for totals"); prune what the question does not need. **What to double-check even when code ran:** final totals (the AI may have summed the wrong column); labels on graphs (numbers usually right, captions sometimes confidently wrong); anything that depends on a column the AI may have misread. Treat AI data analysis like work from a sharp junior analyst: fast, almost always right, occasionally wrong in instructive ways. [^af-03] A useful pattern: your first prompt does not have to be the question. "Describe this dataset. What columns are here, what do they represent, and what 3 charts would best show what is going on?" Read the answer, pick the chart, then ask for it. This catches misinterpreted columns before they become wrong analyses. [^af-03] Continue to [safety, cost and checking](safety-cost-and-checking.md). [^af-03]: AI Prompting 2026: 13 Concepts, 80% of Real Use, Part 3 "Beyond text", Concepts 8, 9 and 10. The AI Agent Factory, Panaversity, 2026.