These 18 Claude AI tips are the ones we apply every day at Tandem. I have already written several editions about the tool. Mainly because Anthropic cracked a few concepts before everyone else. MCP first, the connection between the model and your applications. Skills next, which document your processes for more precise results. Since then, every player has adopted the same logic, and sometimes the same words.
To look back at the earlier editions about Claude, you have:
- Why everyone is switching to Claude (and what you are missing)
- Claude Cowork, the tool you are not using (and that changes everything)
- A practical guide to Claude Code (including for non-developers)
The problem is that the tool moves fast, and part of what I wrote is already out of date.
Why I am coming back to Claude AI tips
Whether you use it on your own or inside a company, we all face the same thing. We use an internal ChatGPT type tool (Claude, ChatGPT, MS Copilot) at 20 % of its potential.
And behind the tool, there is an adoption issue. The idea in this edition is to share our tips to maximise your usage. It is centred on Claude here, but it hardly matters, the logic is the same. (we support companies on Microsoft Copilot, ChatGPT Enterprise, and it is very comparable)
1. Start with the most powerful model, then switch
Start on Fable 5 for the part that needs real thinking when you have a problem, then switch to a lighter model for the back and forth. You can change model in the middle of a chat without losing the context. Very handy to manage your tokens too.
For example:
2. Generate images with Claude
Claude does not generate images the way ChatGPT does. It does write HTML, which is more than enough for structured visuals through artifacts. But going through artifacts, the export is not clean. Claude Design can also be an alternative (I covered it in this edition on rebuilding our website with AI).
On my side, I go through Pillow, a Python image manipulation library that Claude Code drives to output clean, reproducible .jpg files. For more inspiring visuals, I go through the API of Fal, an aggregator that gives access to most image generation models in one place.
Here for example, Pillow generated these images in jpg while respecting the Tandem brand guidelines, inside the thread, to illustrate the carousel mentioned below.
Or to generate the thumbnail of my newsletter:
3. Edit your message instead of correcting it
When you got your prompt wrong or the result is bad, do not follow up with another corrective prompt. Edit your message directly in the chat.
You start again on a clean branch, without dragging the bad answer into the context. It changes the quality of the rest of the conversation. (and your token usage too)
4. Use projects when it makes sense
Many users get projects wrong. They are not made for creative tasks, they are made for tasks where you need to pull in internal context.
You drop your documents, your reports, your contracts. Every conversation in the project inherits them. This is the right place to stop repeating what is already documented.
5. Manage permissions and accept authorisations
On connectors and MCP, depending on your authorisations, you can frame action by action what Claude is allowed to do. Worth doing as soon as a sensitive tool enters the loop.
On Claude Code, the topic is different. I used to switch to a mode that ignores permissions, otherwise the tool asked a question every 30 seconds and that is unbearable...
This is exactly what auto mode solves, and it has become the default mode. Every action is checked before it runs, and it deliberately ignores your own allow list rather than trusting it. (Shift + Tab switches mode easily)
6. Add the connectors
Many tools have shipped their MCP in recent months. Take 10 min to check whether yours did, it changes your usage overnight: Qonto for invoicing, Pennylane on the accounting side, your CRM, your enrichment tools.
One tip, do not leave all your connectors switched on permanently. You clutter the context and you open access for nothing. Enable when needed, disable afterwards (and remember to switch them back on next time).
Honestly, having your everyday tools reachable from a chat in plain language changes your life.
7. Claude has no access to social networks
Contrary to what many people imagine. No reading of LinkedIn, no retrieval of someone's posts, no monitoring on X. You have to go through third party APIs.
You can use ScrapeCreators for that, or RapidAPI. Useful for social media monitoring, niche research, content ideation, ad analysis, influencer outreach.
An interesting combo = retrieval through the API, then updating your CRM through MCP in the same conversation for example.
8. Manage the context window
This is the most underestimated topic. On every message, Claude re-reads the whole conversation from the start. On my carousel, the first message costs 500 tokens and the thirtieth costs 15,500, which is 31 times more. A long conversation gets expensive and less precise with every turn. A glimpse of my LinkedIn carousel on the cost of tokens.
On Claude Code, you get visibility with /context, and /compact to restart from a summarised base.
Simple rule, when you change subject, change conversation.
9. Ask Claude to ask you questions
Asking Claude what it is missing is the reflex that makes the biggest difference for the least effort. Before executing, put the question to it. Rather than heading in the wrong direction on rough assumptions, you correct before producing.
10. Always check the content produced
Models are more powerful, but they also take more initiative. That is the case with Opus 5 compared to 4.8. The model is more powerful, but hallucinations became more significant because it tries to be as autonomous as possible. On the AA-Omniscience benchmark from Artificial Analysis, read in August 2026, Opus 5 at maximum effort shows a 50 % hallucination rate, against 36 % for Opus 4.8.
Despite how models evolve, remember to check the content.
11. Claude Code is the mode to favour
This is my real underlying advice. Claude Code is more powerful than Cowork in every situation. The only problem is the barrier to entry of moving to a slightly more technical environment (which requires no skill at all).
The mistake is to believe it is reserved for developers. Even for data analysis, document generation or contracts, Claude Code gives the best result. You can go further and rebuild your website with it, build knowledge bases, complete automations. I covered it on the rebuild of our website and in the last edition on the video editing done on Claude Code.
I really need to write an edition dedicated to the most powerful use cases of Claude Code.
12. Use routines
For everything that comes back every week in the same format, such as your monitoring review, your reporting or your meeting preparation, think about creating routines.
Routines let you schedule tasks (locally or not) using your skills, your MCP connections, and so on.
You can even set up a notification on Slack at the end of the run:
We do not think about it enough, but there is a lot of value in this feature.
13. Claude Design for your presentations and your charts
And above all, share screenshots of what you want to obtain. A screenshot guides the model better than paragraphs of description on a visual subject. The tool keeps being improved.
Remember to create your design.md (your brand guidelines, shareable with the model)
14. Stay on a fixed plan
You can enable usage based cost, but it runs up fast. Stay on the fixed plan, it is the only way to get a predictable bill. And it is not with the 250 euros offered by Anthropic that I am going to have fun switching credits on.
Beyond 150 users, moving to a per token model becomes mandatory at Anthropic, and in those cases cost management becomes a topic of its own.
15. Create your skills
This is the key point to stop sharing the same instruction twice. A skill is a file that describes how to carry out a precise task, properly. Claude loads it on its own when it is relevant.
Two things to get right, the content and the description. The description is what lets the model know when to activate it.
16. Record yourself to create a skill
Still in beta, but I put it here because it is going to become the norm. You do the task once while recording yourself, and the skill writes itself from there. Documentation by demonstration rather than by writing.
17. Build a skill on your writing style
Otherwise you produce AI slop, and it shows immediately. Give it your existing texts, let it extract patterns from them, then keep optimising those skills.
18. Do not keep changing tools
Make your choice and stop chasing models. ChatGPT, Claude, MS Copilot, you will find the same features everywhere with a few months of lag. What counts is compounding.
Stop asking which model is best. It does not matter. What makes the difference is the configuration of the tool (skills, MCP) and mastering the different modes.
The real shift, configuring rather than prompting
Before, we compensated for the lack of context with endless prompts. Staying that way, the cost explodes and the quality stalls.
The approach has changed. Invest time in configuring the tool and the context. The prompting effort collapses, the quality of the answers rises.
Which Claude AI tips should you keep first?
What makes the difference today is no longer the model, it is your configuration.
Three things to remember:
- Configure rather than prompt: skills, connectors and projects used well beat ten endless prompts.
- Protect your context: short conversations, messages edited rather than corrected, connectors enabled only when needed.
- Compound on one ecosystem instead of hopping from one tool to another.
And the next step is really to switch to Claude Code. It takes a small click in your head, but the difference is huge.
