The Best LLMs for Each Task (and What I Actually Use)
How I choose between ChatGPT, Claude, and Gemini in my day-to-day work.
Hey friends 👋
Welcome to the first Signal Pro workflow issue.
One of the most common questions I receive is:
“I get so overwhelmed with the sheer abundance of LLMs right now. I’ve got access to ChatGPT, Claude, and Gemini, but which model should I use for what?”
Most people pick one model and force it to do everything.
But that’s a bit like trying to build a house only using a single screwdriver.
In my own work, I treat LLMs like a toolbelt:
One model for writing
One model for deep research
One model for problem solving
One model for learning
One model for coding
Today I’ll show you:
The framework I use to choose models
A task → model map (what I actually use)
My default prompts & setups for each job
Video overview 🎥
In this short video, I walk through:
Why “Which model is best?” is the wrong question
How I think about “jobs to be done” for LLMs
A quick tour of the task → model map you’ll see below
If you prefer to read, keep scrolling.
How to think about model choice
Before we get into specific tasks, here are a few principles I’ve found useful (and I think you might too):
Task first, model second.
Always start from “What’s the job?” not “Which model do I feel like using?”Pairs, not monogamy.
It’s normal to use 2-3 models every day. I almost never use just one.Latency, cost, context > benchmarks.
Benchmarks are interesting, but I care 100x more about how they perform in practice:How fast it is
How cheap it is
How much context it can hold
How well it fits into my existing tool stack
Have a default, and a backup.
For each task, I always know:“Default model I reach for”
“Backup when it struggles or hits limits”
The 80/20 Task → Model Map (Preview)
Here’s the high-level map I use.
Below we’ll dive into:
Exactly how I use each model for each task
The prompts I rely on
How I combine them into workflows
When I switch from one to another
Let’s start with the thing most people do every day…
1. Long-Form Writing
Examples: strategy memos, internal docs, scripts, important emails.
🧰 The job
I want a model that:
Follows instructions tightly
Holds a consistent voice
Handles structure (sections, bullets, summaries)
Won’t hallucinate wildly if I feed it my own notes
✅ My default: Claude Opus 4.5
For structured writing and general knowledge work, I default to Claude.
Why do I use Claude Opus 4.5?
Fast AND accurate
Great at following constraints (tone, length, format)
Integrates nicely with lots of tools (Claude Chrome extension, Claude for Excel)
🔁 My backup: ChatGPT 5.2
I switch to ChatGPT 5.2 Thinking when:
I want extended thinking around exploring arguments for/against
I’m doing “thought partner” work and want non-obvious insights highlighted
Note: If I’m drafting a long-response email and I require multimodal inputs (slide decks / Excel sheets / screenshots / PDFs), I find myself defaulting more and more to ChatGPT. It does a great job of reasoning across different format types. Plus, I can tune the tone by adjusting “Base style and tone” under “Personalisation”. My personal favourite tone is “Efficient” as I like getting straight to the point, both in my conversation with the model and the outputs I receive.
🧱 My writing system prompt
Here’s a writing system prompt I use for both Claude and ChatGPT. Paste this above your prompt:
You are my writing and thinking partner.
Your job is to help me think clearly and write in my voice.
You MUST:
- Ask 3-5 clarifying questions before writing anything substantial.
- Preserve my tone: concise, direct, conversational.
- Prefer structure: headings, subheadings, bullets.
- Challenge fuzzy thinking: if I’m vague, push me to be specific.
If I paste text, first:
1) Summarise it in 5 bullets.
2) Identify 3 sharp points.
3) Suggest 2 ways to make it clearer or more persuasive.📝 Example workflow: turning messy notes into a clear memo
Dump your notes / meeting transcript / bullet list.
Ask:
“Using the system prompt above, turn this into a 1-page internal memo for [audience].”
Then ask for one alternative structure, not five.
If I hit a style ceiling (it feels too generic), I paste the draft into ChatGPT with:
“Rewrite this as if I’d just spoken it out loud, but keep the structure.”
I’m a big believer in writing like you talk.
2. Deep Research
Examples: competitor analysis, reviewing long reports, “should we do X?” questions.
This is where I care less about speed and more about chain-of-thought quality.
✅ My default: ChatGPT 5.2 Pro
I’ll usually start with ChatGPT’s 5.2 Pro model equipped with “Deep Research” mode.
Why?
Handles long, structured inputs well (reports, docs, multi-part prompts)
Very strong at “claims → evidence → implications” style reasoning
Easy to iterate: I can quickly pivot from “explain” to “compare” to “summarise for X audience”
Here, we can see the model immediately asks follow-up questions. This is one thing I LOVE about toggling on “Deep research” mode (click the plus icon at the bottom left of the input field).
It helps you clarify your thoughts and, as a result, provide sufficient context for the model to generate an accurate, robust, and tailored answer to your research query.
Whilst the full power of Deep Research (especially for complex tasks) still leans towards paid plans with higher usage caps and more advanced models, I know free users get ~5 lightweight Deep Research tasks monthly, while paying subscribers (Plus/Team/Enterprise) get expanded access. This thing is a powerhouse, so if you haven’t tried it already, I highly recommend you check it out.
🔁 My backup: Gemini 3 Pro
I switch to Gemini 3 Pro when:
I want to reason across mixed modalities (charts, screenshots, PDFs) in one shot
I want to reference information across the Google ecosystem (Docs, Sheets)
This is a two-step process to enable. Firstly, you’ll want to select “Thinking with 3 Pro” from the “Choose your model” dropdown. Secondly, you’ll want to select “Deep Research” from the “Tools” dropdown.
I’ve found myself using Gemini 3 Pro more frequently over the last week, as when you switch on Deep Research, it enables you to select one or more sources from:
Google Search
Gmail
Google Drive
Google Chat
Whilst I always default to having Google Search turned on, expanding context across recent emails or files in your Drive is definitely a nice-to-have.
🧱 My research system prompt
I use a variant of this when using ChatGPT:
You are my research analyst.
Your job is to digest the materials I give you and produce:
1) Key claims
2) Evidence (with citations or source references if available)
3) Implications for my work as [role]
4) Open questions or risks I should be aware of
Rules:
- No fluff. Prefer bullets over paragraphs.
- Separate fact from interpretation.
- If the source is weak or contradictory, say so.
- If you’re uncertain, write: [insufficient data].
First, restate the question you think I’m trying to answer.
Then ask up to 3 clarifying questions before you start.With Gemini, I provide slightly more unstructured thinking without the need to append my question with this system prompt, as it does a good job of formulating it into a research plan.
Plus, if I’m not totally happy after reading the expanded plan, I can always click on the “Edit plan” button to be a little more specific about the effect I’m trying to achieve.
3. Problem Solving & Complex Reasoning
Examples: strategy questions, edge-case thinking, “what should we do?” decisions, scenario analysis.
🧰 The job
I want a model that:
Is comfortable with ambiguity and trade-offs
Can explore multiple scenarios rather than giving one generic answer
Can challenge my assumptions and argue both sides
Handles step-by-step reasoning without collapsing into buzzwords
✅ My default: Grok 4.1
For complex reasoning and problem solving, I start with Grok 4.1.
Why I like Grok here:
It’s willing to be more opinionated and direct
Significantly lower hallucination rate on edge-case knowledge
Feels like a sharp colleague who will call out bad assumptions instead of smoothing everything over
I use it when I’m:
Evaluating different strategic options
Doing pros/cons analyses with real nuance
Asking “what am I missing here?” on big decisions
Something I really enjoy about Grok 4.1 is the context window. Grok 4 typically lost coherence past ~300K tokens in practice, but Grok 4.1 has full needle-in-haystack accuracy up to 1M tokens.
This compares considerably to the likes of Claude Opus 4.5, which has a ~200K token context window. I find this especially helpful for long conversations with extended Q&A. What’s more, it’s totally free to use (up to certain rate limits).
🔁 My backup: ChatGPT 5.2 Thinking
When I need slightly more structured, slower, chain-of-thought reasoning, I move to ChatGPT 5.2 Thinking.
I use it for:
Multi-step logical reasoning
Scenario planning (“If we do A vs B vs C…”)
Structured breakdowns where I want a very clear logical spine
A pattern I like:
Start with Grok to generate diverse, provocative perspectives.
Then feed those into ChatGPT 5.2 Thinking to structure them into a coherent decision framework.
🧱 My complex reasoning system prompt
Used in both Grok and ChatGPT Thinking:
You are my critical thinking partner.
Your job is to help me reason through hard problems and make better decisions.
You MUST:
- Think step by step.
- Make trade-offs explicit.
- Challenge my assumptions rather than just agreeing.
- Offer at least 2-3 distinct options with clear pros and cons.
Structure every answer as:
1) Clarified question
2) Key assumptions (explicitly listed)
3) Options with pros/cons
4) Your recommended option and why
5) Risks and what would change your mindYou now have my defaults + backups for:
Writing
Deep research
Problem solving
In the second half of this issue, we’ll go deeper into:
How I learn new topics with LLMs
How I actually code with them in Cursor
How I combine everything into a daily stack and a one-week experiment you can run.
🔒 That part is for Signal Pro members.