My productivity has increased by at least 300% with AI assistance. You can get amazing results nowadays. If you use the tools right. Discover 4 key ingredients that make the tools work for you in this post.
Many people have only tried free AI via ChatGPT or similar web chatbots. It’s easy to dismiss those tools, since they lack all 4 ingredients.
1. Context. If I ask you how I could improve my business, you won’t be able to provide a good answer. You don’t know all the details about my business that matter. All you can do is offer generic advice or make guesses (hallucinations). It’s the same with AI tools. Don’t rely on the knowledge baked into the LLM base models. Either provide this knowledge, or provide the tools to obtain that knowledge. You can include “all relevant knowledge” in the prompt, but this is labor-intensive. This is why you want an agentic tool.
2. Agentic tools. I’ve been using Claude Code, a CLI tool that provides agentic AI for knowledge tasks (not just coding). There is also Claude Cowork, a desktop tool, and alternatives from vendors besides Antrophic. These tools use a loop in which the AI determines whether it needs more information and then goes looking for it. You can give these tools a task or a question, and then they will, if called for, run hundreds of searches and commands. They can look at your documents, codebases, and web resources. Tell these tools “Fix GitHub issue $link”, and they’ll look at the issue, anything references on the issue, as your codebase, make changes, run tests, make more changes, check results via the browser, fix some final issues, create a draft pull request, and provide you with a summary of what was done and possible next steps.
3. Feedback harness. When writing code, you often don’t get everything correct the first time. Which is why automated tests are great. More generally, fast feedback loops are great, regardless of whether you’re doing software development. For software development, you’ll get much better results if the AI tools can actually run the code and run tests and other CI tools to verify everything is correct.
4. Model. AI capabilities are increasing at an incredible pace. If you’re using the latest models, your experience will be worlds apart from those using 2-year-old models. For maximum quality, there are 3 metrics to max out: model size/capability, model version, and effort parameters. In other words, use the latest version of the biggest model with “max effort”. At the time of writing this post, that is Claude Opus 4.7 with max-effort when using Antrophic, or GPT-5.4 Pro with heavy thinking when using OpenAI. These settings eat tokens, so you will quickly run into your subscription limits of the basic tiers. Then again, paying 200 USD a month for the higher tiers so you can 5x your productivity is quite the bargain.
Those 4 points provide a conceptual framework. There is more to learn, and the AI space is evolving quickly. Ask your favorite AI tool how you can improve your AI workflows, starting from this post, to get specifics.
Some more tips:
- Know how to use CLAUDE.md / AGENTS.md
- Create sandboxed environments so you can let agents run autonomously for longer periods of time
- Mind the current model’s tendency to sycophancy when prompting. If you go, “Here is my idea, is it good?” LLMs will often say yes even if there are issues. Adding “Be brutally honest” to your prompt or CLAUDE.md helps. It takes some practice to build up an understanding of how to prompt and in which ways responses should be distrusted. As a starting point, treat current LLMs as overeager sycophantic juniors with an inhuman, jagged skill profile, who tirelessly work at superhuman speeds.
- Claude Code (or similar) plus local files in a text format works well. I’ve been enjoying Obsidian + Claude Code for personal knowledge management.
You can stay up to speed with AI capabilities development via Don’t Worry About the Vase and Astral Codex Ten, which I both highly recommend.
Shameless plug: my company provides an AI Assistant for MediaWiki, giving you AI capabilities on top of collaborative knowledge management, ideal for organizations.