::which AI for which task · honest decision tree
Pick the AI
for the task you actually have.
18 task categories. For each: the right model, the alternative, the exact prompt shape, the trap. Zero affiliate revenue from any of these tools. We name them because they work, not because they pay.
::task 01 of 18
Long-form writing (article, blog, memo, essay, report)
::pick this
Claude · Sonnet 4.5+
Claude's writing voice is the most consistent across length. Long-context window holds the whole draft in working memory. Tends to ask better clarifying questions before drafting.
::or this · the alternative
ChatGPT 5+ for shorter pieces. Gemini 2.5 Pro for research-heavy long form.
::prompt shape
1. Paste your context + the audience + the constraint. 2. Ask for the outline first, NOT the draft. 3. Approve the outline. 4. Write the draft section by section, asking 'critique this section like a hostile editor' between sections.
::the trap
Asking Claude to 'just write the whole thing' is how voice dies. Iterate.
::task 02 of 18
Short copy (subject lines, headlines, taglines, microcopy, ad copy)
::pick this
ChatGPT 5+ OR Claude · either works
Both produce high-variant output fast. Ask for 12 variants across 4 registers and pick the one that fits.
::or this · the alternative
Gemini if you want fresh angle that's slightly out-of-distribution from Claude / GPT.
::prompt shape
Always ask for 12 variants across 4 explicit registers (e.g., plain / sharp / curious / specific). Then ask for the model's pick + reasoning. Then YOU pick.
::the trap
Asking for 'a few options' = 3 variants of the same idea. 'Across N registers' forces real range.
::task 03 of 18
Software architecture, system design, code review, refactoring
::pick this
Claude · Sonnet 4.5+
Best at long-context code reasoning. Most-likely to ask 'have you considered X edge case' instead of jumping to code. Strongest at structured critique.
::or this · the alternative
Cursor or Codex if you need it inside your editor with file-level autonomy.
::prompt shape
Plan FIRST. Paste target file + task. Ask: 'don't code yet. Give me 3 implementation approaches, what each breaks, the test you'd write, pick one.'
::the trap
Letting the model code first. The plan is 80% of the win.
::task 04 of 18
Quick scripts, glue code, throwaway automation, regex, shell
::pick this
ChatGPT 5+ · Claude works too
Faster turnaround on small, well-scoped tasks. No need for the long-context overhead.
::or this · the alternative
Local Ollama (DeepSeek Coder, Qwen Coder) for offline work.
::prompt shape
Be specific about input shape + output shape + edge cases. 'Take a CSV with these columns, output JSON with this shape, handle empty values like X.'
::the trap
Vague specs = wrong code. The 30 seconds spent specifying saves 30 minutes of re-prompting.
::task 05 of 18
Academic research, paper analysis, citation hunting
::pick this
Claude (long context) + manual verification
Best at multi-paper synthesis. Long context holds 20+ abstracts at once. Strong at flagging methodological issues.
::or this · the alternative
Perplexity for initial discovery (it has real web access; Claude doesn't).
::prompt shape
Paste abstracts/papers. Ask for synthesis, gap analysis, method critique. ALWAYS verify any citation Claude gives you against the actual paper — it hallucinates citations confidently.
::the trap
Trusting a citation without pulling the source. Sanctions-level mistake for legal work, reputation-killer in academic work.
::task 06 of 18
Current events, breaking news, real-time market signals
::pick this
Perplexity · or ChatGPT 5+ with web search
Perplexity grounds against the live web by default. Claude does not (it has a knowledge cutoff). ChatGPT 5+ has web search but it's slower and sometimes misses.
::or this · the alternative
Manual search + Claude for synthesis once you have the sources.
::prompt shape
Ask Perplexity for primary sources + dates + links. Then paste those into Claude for analysis.
::the trap
Asking Claude 'what happened this week?' — it doesn't know. Treat Claude as analyst, Perplexity as researcher.
::task 07 of 18
Spreadsheet analysis, CSV/Excel data, trend extraction
::pick this
Claude (with file upload) OR ChatGPT · Advanced Data Analysis
Claude's analytical reasoning is strong. ChatGPT's Code Interpreter actually runs Python on your data, which beats hand-waving.
::or this · the alternative
Wolfram Alpha for math-heavy verification. Your spreadsheet's native AI for in-app work.
::prompt shape
Always state: the data shape, the question, what you'd consider a noteworthy finding, what you'd treat as noise.
::the trap
Letting the AI 'analyze the data' without specifying what counts as a signal. You'll get patterns that aren't real.
::task 08 of 18
Image generation, illustration, design exploration
::pick this
Midjourney · or Google Imagen · or DALL-E (ChatGPT 5)
Midjourney for highest visual quality. Imagen for natural-language prompt fluency. DALL-E for in-chat workflow.
::or this · the alternative
Stable Diffusion (local) for unlimited generations and uncensored exploration.
::prompt shape
Subject + style + composition + lighting + mood + aspect ratio. The more specific, the more controllable.
::the trap
Using AI images on a face-driven brand. The uncanny shows up at scale and hurts the click.
::task 09 of 18
Audio transcription, voice cloning, voice generation
::pick this
Whisper (OpenAI) for transcription · ElevenLabs for voice generation
Whisper is the open-source standard for transcription accuracy. ElevenLabs is the highest-quality voice cloning.
::or this · the alternative
Descript wraps both with an editor UI if you don't want CLI work.
::prompt shape
For voice cloning: 1-2 minutes of clean source audio works. The 'studio' models cost more but sound noticeably better.
::the trap
Cloning someone's voice without explicit consent. Don't. The downstream is legal and ethical exposure.
::task 10 of 18
Anything with NDA, PHI, PII, financial data, customer info, source code under privilege
::pick this
Local Ollama · Llama 3.1 70B or Qwen 2.5 72B
Runs entirely on your machine. Zero data leaves. Same model family, slightly behind frontier in quality. Worth it for the data sovereignty.
::or this · the alternative
Your facility's vetted internal LLM (corporate Claude / Azure OpenAI / Google Vertex) if BAA / DPA in place.
::prompt shape
Same as Claude / GPT — these models follow instructions well. Quality is 80-90% of frontier for most tasks.
::the trap
Pasting PHI into ChatGPT.com because 'it'll be fine.' It won't be — not legally, not ethically, not when the audit happens.
::task 11 of 18
Anything with images + text (screenshot analysis, design feedback, photo Q&A)
::pick this
Claude (vision) · or Gemini 2.5 Pro
Claude's vision is sharp at structured analysis (UX heuristics, code from screenshot). Gemini's vision is sharp at natural-photo Q&A.
::or this · the alternative
GPT-5 Vision works fine; ChatGPT's interface is the smoothest for multi-turn vision conversations.
::prompt shape
Paste image + ask specific question. 'What's wrong with this design's hierarchy?' beats 'review this design.'
::the trap
Asking AI to read a screenshot of a long document. OCR-friendly source (paste the text) gives much better results.
::task 12 of 18
Math, proofs, formal verification, symbolic computation
::pick this
Wolfram Alpha + Claude · two-tool workflow
AI is bad at math. Always verify symbolic / numeric work in Wolfram or Python. Use Claude for reasoning structure and Wolfram for the actual computation.
::or this · the alternative
Lean / Coq if you're doing formal proofs. AI won't replace them.
::prompt shape
Use Claude to structure the proof. Use Wolfram to verify each step. Don't combine.
::the trap
Trusting AI math. Even high-tier models make arithmetic errors that pass the eye test.
::task 13 of 18
Multi-step agentic work (research → analyze → draft → publish)
::pick this
Claude (Sonnet 4.5+) with tool use
Currently the most reliable for multi-step planning + tool use without losing the thread. Long context holds the whole task state.
::or this · the alternative
GPT-5 with function calling is close. Use whichever has the better SDK ergonomics for your stack.
::prompt shape
State the goal, the constraints, the tools available, the success criteria, the failure-handle. Ask for the plan first, approve, then execute.
::the trap
Letting the agent loop without a budget. Cap iterations and tokens. Watch the receipts.
::task 14 of 18
Translation, localization, cross-language QA
::pick this
DeepL for accuracy · Claude for nuance + voice
DeepL is the highest-accuracy machine translation for European languages. Claude is better at preserving voice / register / cultural nuance.
::or this · the alternative
GPT for less-common language pairs. Always have a native speaker review for shipped content.
::prompt shape
Paste source + target language + the audience + the voice / register you want. 'Translate this for a 14-year-old French reader, casual register' beats 'translate to French.'
::the trap
Auto-publishing AI-translated content. Native-speaker review is mandatory for anything customer-facing.
::task 15 of 18
Summarizing long documents, books, papers, transcripts
::pick this
Claude (200K context)
Long context is Claude's killer feature. Can hold an entire book and answer specific questions across it.
::or this · the alternative
Gemini 2.5 Pro (1M context) for extreme-length material (codebases, full year of transcripts).
::prompt shape
State what you want to extract, not 'summarize.' 'Pull the 5 strongest arguments + their counter-arguments + the page numbers' is far better than 'TL;DR.'
::the trap
Summarize-then-decide is the wrong order. Extract-with-intent gives you usable output.
::task 16 of 18
Brainstorming, ideation, divergent thinking
::pick this
Any model · range > model choice here
All frontier models brainstorm well. The variable that matters is how you prompt: ask for explicit range + a wildcard + a contrarian.
::or this · the alternative
Use a different model for a second pass — different models converge on different defaults.
::prompt shape
Ask for 12 ideas across 4 different angles + 1 wildcard + 1 contrarian. Then pick the one that's furthest from where YOU were already going.
::the trap
Accepting the first 'good idea' the AI gives you. The third pass is usually the unlock.
::task 17 of 18
Personal coaching, life-decision sounding-board, therapy-adjacent reflection
::pick this
Claude · most appropriate trained behavior here
Claude's tuning leans toward asking clarifying questions and not over-prescribing. Useful for thinking-out-loud.
::or this · the alternative
ChatGPT works similarly. Avoid models tuned for entertainment / persona play for this use.
::prompt shape
State the situation, what you've already considered, what's stopping you. Ask: 'What am I not asking myself that I should be?' or 'What's the strongest argument for the option I'm not picking?'
::the trap
AI is not therapy. For depression, anxiety, suicidal thoughts — go to a real human professional. AI is the journal, not the therapist.
::task 18 of 18
I don't know what to use
::pick this
Start with Claude. Free tier. Daily limits are generous.
It's the strongest all-around model right now. Honest answers, decent voice, long context, sharp critique.
::or this · the alternative
When Claude rate-limits, hop to ChatGPT. When you need real-time web, Perplexity. When you can't share the data, Ollama.
::prompt shape
Ask Claude what model you should use for your specific task. It'll give you a straight answer.
::the trap
Subscription stacking. You do not need to pay for 4 models. Free Claude + free ChatGPT + free Perplexity + free Ollama covers 95% of work for 95% of users.
::keep going
Pick the AI. Now grab the playbook.