> ## Content Index
> Fetch the complete content index at: https://www.taivo.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# What I use AI for
- URL: https://www.taivo.ai/what-i-use-ai-for/
- Published: 2026-09-21T20:40:53.000Z
- Updated: 2026-09-21T20:43:07.000Z
- Description: A growing collection of ways I use AI agents at work and in daily life, from shopping and interior design to research and meeting preparation.
- Author: Taivo Pungas
- Tags: Everything else, stream

I try to use AI for almost every virtual task, and keep discovering things they can usefully do. This is a growing collection of those examples.

As of September 2026, my daily driver is the Codex desktop app with ChatGPT Astra, though I occasionally use Claude Code as well.

# Daily life

## Shopping

When buying something, it's wise to compare alternatives to find the best option. Without AI, it was really only worth it for large purchases: real estate, car, maybe some electronics. With AI, there really is no limit to what you can comparison-shop. For a kid's birthday gift, I used it to find relevant but deeply discounted items, expanding my effective budget.

I specifically love having AI tear through second-hand marketplaces for a rough description of an item. Since I store my target wardrobe style and rough current status as a Markdown file, it's easy for AI to know which items I should be on the market for, and look for every brand and model, not just a few major ones.

On a slow Saturday I might lazily fire off a query like:

> check my wardrobe system, find if any great deals on Vinted, only new items

Which gives me back fresh recommendations, several of which I have ordered and am happy with, typically getting 50-75% off the retail price. I know some people enjoy the browsing; I do not.

## Interior design using a 3D replica

I wish I'd had this when we furnished our home. I fed in a two-minute walkthrough video (which Codex processed with [mex](https://mex.taivo.ai/?ref=taivo.ai)), floor plans, and a few furniture receipts, and got back a 3D representation of the apartment.

It is dimensionally accurate, or at least close enough that you can test the visual fit of furniture. AI image generators can do that for you too, but the dimensions are ballparked and often wrong, so a 3D model is much better.

![](https://storage.ghost.io/c/4a/28/4a28c12b-061d-4fef-9be7-3d9f25838589/content/images/2026/09/Screenshot-2026-09-20-at-23.28.04-1.png)

## Book-hunting in libraries

My wife regularly borrows new books for our children to read, and has a list of books she might want to get soon, on Notion. But actually getting them requires knowing three things: which library has them, is it currently available, and what shelf is it on.

We now have Claude search the Estonian public library system's site whenever we want to go and get a new batch, so we can make the library trip an efficient one.

## Reading public institution data

Occasionally I get curious about something that has a clear paper trail somewhere in Estonian public data. For example, a scandal in the news about purported dishonest hiring decisions in the Estonian Natural History Museum, or the decision to close down all roads near my house for a weekend because of a marathon.

I built [Kodaniku Kratt](https://kratt.taivo.ai/?ref=taivo.ai) ("citizen's agent") to help with these queries. It is essentially a knowledge base that explains how to access data provided by various Estonian public institutions, from the Prime Minister's office to the local utility.

# Work

## Copyediting

I have a setup of several CLI "prose linting" tools that together give me basic Grammarly-like functionality: [LanguageTool](https://github.com/languagetool-org/languagetool?ref=taivo.ai), [Vale](https://github.com/errata-ai/vale?ref=taivo.ai), and [textlint](https://github.com/textlint/textlint?ref=taivo.ai). When I want something checked, I have Claude run a script, look at the output, ignore the duplicates, and give me a numbered list of issues.

From there, I usually let Claude fix the simplest grammatical errors, and propose edits for the less trivial ones like too long sentences, weasel words, etc. In addition, I can of course get Claude to review anything else about the text simply by reading it. Generally, I don't use LLMs for producing text or major rewrites, because it pushes the text towards the median, which drowns out my voice.

This is not at all a novel use of AI, but I still like it over proprietary vendor tools: I can improve things myself, I can have Claude do it, and I can make granular decisions about what things to auto-fix vs where I want to give input.

## Onboarding

On an intro call with a new colleague at Pactum, she asked me to help identify relevant work already done in her area.

Since my agents have skills to access our main collaboration tools (Google Drive, Confluence, Gitbook), I quickly fed her request directly into Claude for some research. It really only took about three minutes and came back with not just links, but also a 1-sentence description of each, and a category.

I used a similar approach when onboarding to a new codebase. There was a workflow orchestration component which I did not fully understand, so I asked Claude to read all about how it works, make a summary document, and give me a tutorial with progressively more complex examples. That worked super well.

Onboarding people traditionally takes lots of work to get right because documentation is rarely kept up to date; now it can be accelerated a lot.

## Expense reports

The age-old job of parsing invoices and producing documentation for reimbursing expenses. I've used Claude to parse invoices and prepare expense reports in Excel and email for our Estonia-based employees.

## Shuffling information to and from Jira

While I was temporarily leading a small team, we experimented with keeping the team knowledge base in an Obsidian vault: Markdown files tracked in Git. We wanted all the team-relevant context in that vault.

The setup is powerful: we can have permanent documentation (e.g. team purpose and KPIs) alongside plans, notes, work-in-progress analyses, etc. Since it's Markdown, Claude can easily read, search, and edit too. But we still track progress on our plans in Jira, so we need to keep that updated. Claude handles connecting commits to ticket IDs, adding relevant information to Jira comments, and more.

## Weekly review as a chief of staff

For my weekly review, I ask Claude to look through my daily notes, calendar, emails, Slack, and meeting transcripts from the past week, and give me an update on what I did. It pulls together threads I'd forgotten about, surfaces follow-ups I missed, and reminds me what's still open.

> **Taivo:** Look through my last 2 weeks of daily notes, CTO plans, maybe weekly reviews. I am considering what I have missed or not acted on, or what topics have been in the air, to inform decision of where I should focus.

From there, I review my open projects and priorities, and ask it to help me think through where to focus this week.

I also use a variation at the end of the day:

> **Taivo:** Out of what I worked on today, consider what I should be communicating to others.

It's a good forcing function for visibility.

The pattern here is using the agent as a chief of staff who has read access to all your systems. It doesn't make decisions for you, but it gathers the context you need to make good ones faster.

## Meeting preparation

Before a day of meetings, I ask Claude to look through my calendar and consider how I might want to prepare for each conversation.

> **Taivo:** Look through my calendar for today, consider how I might want to prepare for each convo.

For a customer meeting, it'll pull recent activity from our CRM and call recordings. For a 1:1 with a direct report, it'll check our shared notes and recent Slack context.

The most useful part is when it surfaces connections I wouldn't have looked up myself: an attendee I haven't met before, a thread from two weeks ago that's relevant to today's topic, or a decision from last quarter that I'd forgotten about.

This takes about two minutes and often saves me from walking into a meeting cold.

## Writing role descriptions interactively

Most people use AI to generate a first draft. For documents where I've done lots of thinking but haven't organized it, I flip the interaction: instead of asking the agent to produce content, I ask it to interview me.

When writing a role description for a new hire, I gave the agent the high-level framing and it asked me about 20 questions over the course of an hour.

It read our Confluence docs for context, and drafted sections as we went. The result was much more specific than what I'd have produced on my own, because the Q&A format forced me to articulate assumptions I'd been carrying around.

This works for any document where the knowledge exists in your head but hasn't been structured yet.

## Building skills for your agents

I use agents to build their own capabilities. For example:

> **Taivo:** Look through emails I've sent; describe the tone / approach / communication, in roughly 5-10 bullets. Put these bullets into a new Claude skill, which is about speaking / communicating on my behalf.

I turned those patterns into a skill that helps the agent write messages in my voice. I built a skill that connects to our call recording platform, so agents can search customer conversations.

The interesting part is the feedback loop. I notice a gap ("I wish Claude could check Gong transcripts"), build the skill in an afternoon (often with Claude's help), and start using it the same day. A few iterations later it's ready to share with the team. The gap between "I wish it could do this" and "it can do this" has collapsed from months to hours.

## Market intelligence

We needed to understand which of our prospective customers use certain procurement platforms. Customer lists aren't public, so you can't just look this up. But answers can be pieced together from public sources: press releases, supplier portals, job postings, conference presentations.

I had Claude systematically research about 50 companies, checking four or five data sources for each one and compiling the results into a table. What would have been a week of manual research by an analyst took 20 minutes of Claude running in the background.

The accuracy wasn't great on the first pass. It confidently stated things that turned out to be outdated, and missed signals that a domain expert would have caught. But it gave us a starting point good enough to prioritize outreach, and the method is repeatable. Each round gets better as we learn which sources are reliable and which aren't.

## Other

There are many more small ways I use agents:

- Research + summarization as a category:
  - reading through past 1:1 notes with people (for self-reflection),
  - understanding what a team has shipped in the last quarter (through Jira and Git),
  - understanding my own Chrome history for most visited sites that I need to give agents access to.
- Reviewing a document for staleness (specifically, our [Engineering hiring page](https://github.com/pactum-ai?ref=taivo.ai)).
- Creating tickets in a particular format, e.g. for internal IT requests.

And of course, you can still use Claude Code to write code.

#