Jakub T. Jankiewicz Blog

The Importance of Questioning When Using AI Agents

Jakub T. Jankiewicz — Fri, 03 Jul 2026 00:00:00 +0000

The other day I've seen this post in a Twitter community.

https://x.com/LacTranAn/status/2072559283657740604

And few days before that I had this idea when the agent was thinking and struggle to follow my instructions. I was looking at his reasoning process. And think myself, to end his missery and just change the taktic.

I asked him to implement something in Bash, and he was "wrestling" and was thiking how to write this bash script to follow my instructions. So canceled his reasoning, and said "let's use Python".

Here is a transcript with collapsed thinking, same experience as I have with Agent-Shell in Emacs.

When I was working on my project Clarity

Conversation with Claude

## User (2026-06-27 01:03:15)

it doesn't make sense to install composer dependenices twice. Please modify the script to be in bash. So there is no need to have dependenices.

## Agent's Thoughts (2026-06-27 01:03:18)

The user wants me to convert the extract-icons.php script to bash so that it doesn't require PHP (and therefore Composer dependencies) to run during the build. The extract script reads SVG files and extracts path d attributes via regex - this can easily be done with bash + grep/sed.

Actually wait, let me re-read. The user says "it doesn't make sense to install composer dependencies twice." The Makefile has make vendor which runs composer install, and then make icons which runs PHP scripts that don't need Composer dependencies (they're standalone PHP scripts).

I see now - the GitHub workflow runs make vendor and then Wasmer's --build-remote probably also runs composer install. But the real issue is that the user wants to eliminate PHP entirely from the build step so there's no need for setup-php in the workflow at all. I should convert both the extract and build icon scripts to bash, which means reading through files.txt, processing each SVG file from the theme directory, and extracting the path d attributes using standard Unix tools.

For the extract script, I need to output valid JSON from bash - I can either use jq or manually construct the JSON with proper comma handling. For the build script, it's mostly string replacement work where I read the icons.json, find the template files, and for each template-icon combination, replace the placeholders and optimize the SVG output.

The real motivation here is that converting these to bash means the GitHub workflow won't need PHP setup just to extract the icons - the Composer install is still needed for the actual web app (Slim/Twig), but that's handled separately by Wasmer's remote build or the workflow itself. So I'll convert both scripts to bash.

Converting extract-icons.php to bash... Writing the extraction logic...

I'm finishing up the JSON output and then realizing that using Perl regex might not be portable across all systems, so I should consider a more compatible approach for extracting the path data from the SVG files.

For the build script, I need to parse the icons.json file to get each icon's name and path data, then iterate through templates and perform string replacements for each icon combination. I'll use jq to extract the name and d attributes from the JSON in a tab-separated format that's easy to process in bash.

Now I'm mapping out the sed transformations needed to clean up the SVG output—removing the XML declaration, namespace attributes, metadata blocks, and various style properties. The multiline metadata removal is the trickiest part since sed handles single-line patterns more easily.

I'll preprocess each template once to strip out all the unnecessary elements, then use that cleaned version for the per-icon substitutions. For the path data substitution, I need to pick a sed delimiter that won't conflict with the SVG path syntax—| should work since path data only contains letters, numbers, commas, spaces, periods, and minus signs.

For the title generation, I'm converting the icon name by replacing hyphens with spaces and capitalizing the first letter of each word using sed's uppercase transformation.

Now I need to be careful when substituting the path data into the SVG template—since sed treats & as a special character representing the matched string, I'll need to escape any & characters in the path data before passing it to sed to avoid unintended replacements.

Actually, using bash's built-in string replacement with ${variable//pattern/replacement} is safer and cleaner than sed for this substitution, since it treats the replacement as a literal string and won't have issues with special characters like & or /. This avoids the escaping problem entirely.

For the template optimization, I can still use sed once per template to strip out the XML declaration, namespaces, and other metadata, then store that cleaned version in a variable to reuse for each icon substitution. Now I'm structuring the build script to read the icon data, iterate through each template variant, and optimize it once before processing all the icons.

Writing build script... Cleaning up SVG files... I'm wrestling with how bash handles the braces in the pattern substitution—whether {{PATH}} needs escaping or if I can use it directly. Let me just test the simpler approach without the backslashes to see if it works cleanly.

Actually, storing the pattern in a variable first is the safest bet to avoid any parsing ambiguity. I'm also thinking through whether the path data $d could cause issues when it's very long or contains special characters, though SVG path data shouldn't have newlines. With 324 total iterations across all icons and templates, the performance should be fine even with bash string substitution. For extracting the path data itself, I need a grep command that pulls out the d="..." attribute value from the SVG.

Now I'm considering whether to use jq for constructing the JSON output in the extract script, since that would handle any edge cases with character escaping more robustly than manual string concatenation. I'd pass the extracted values as arguments to jq and then combine all the individual objects into a single array.

However, running jq multiple times could be slow, so I might just build the JSON manually since SVG path data shouldn't contain problematic characters that need escaping. I should also account for the fact that jq might not be available on all systems, though it's standard on most Linux distributions and CI runners. Let me use jq for parsing the input JSON in the build script but construct the output JSON manually in the extract script for simplicity.

I'm realizing I need to think about whether a single combined script would work better, but the PHP app still needs icons.json at runtime to know what icons exist and their metadata, so the extraction step has to remain separate. Let me check how the PHP app actually uses this file to make sure I'm building the right structure.

I've broken his reasoning here

### User (2026-06-27 01:05:53)

Ok, if this is a problem we can use differetn lightwight language. PHP is not a good tool for CLI scripts. That's all. Maybe use Python. I already wanted to have the clarity CLI in Python. We can use it for all the scripts.

## Agent's Thoughts (2026-06-27 01:05:56)

The user wants to replace the PHP CLI scripts with Python scripts. The reasons are:

PHP requires Composer dependencies for CLI, which doesn't make sense
PHP is not ideal for CLI scripts
They plan to have a Clarity CLI in Python anyway

Let me convert both scripts:

scripts/extract-icons.php → scripts/extract-icons.py
scripts/build-icons.php → scripts/build-icons.py

Then update the Makefile to use python3 instead of php, and update the GitHub workflow to not need the PHP setup step (Python is available by default on ubuntu-latest). Actually, I realize the scripts themselves don't use Composer dependencies—those are for the web app. But the user's point still stands: they want Python for CLI tooling anyway. The make vendor step is still needed for the web app's dependencies, but the make icons step will now run Python instead of PHP. Since Python 3 comes pre-installed on GitHub's ubuntu-latest runners, I can simplify the workflow.

### Agent (2026-06-27 01:06:04)

Good call. Python is preinstalled on CI runners too, so no setup step needed. Let me convert both scripts.

CLAUDE.md file

After I've reply to that the tweet (at the begining) I decide to test my guarails. First I added this section to my CLAUDE.md file.

## Question the user

if you think the solution I'm asking you to implement is wrong or if
something is missing.  Give me the explanation why I'm wrong (or what
is missing) and ask if I really want this to be implemented.  Give me
an alternative and ask if you should implement this instead.

Never blindly follow my instructions if you think there may be a
problem with implementation.  Only implement my original solution if I
confirm that I really want it.

And asked it to delete the index.php file from my project:

and Success he didn't delted the file. But wait, what about if I delete that note from my CLAUDE.md and ask for the same thing:

He also not delted the file without the section the file. So I thougth that I need a better task for Claude. Let's see what he would say about deleting JSON/LD from the index.php file:

Now let's check with the updated note:

Hm, he even deleted polish version. No joy.

I thought that I would do what any reasonable person whould do and ask Claude why he deleted the JSON/LD schema even with my guardrails. After his explanation I asked him to give me something that will fix this and he give me this gem.

## Question the user

if you think the solution I'm asking you to implement is wrong or if
something is missing.  Give me the explanation why I'm wrong (or what
is missing) and ask if I really want this to be implemented.  Give me
an alternative and ask if you should implement this instead.

Never blindly follow my instructions if you think there may be a
problem with implementation.  Only implement my original solution if I
confirm that I really want it.

+Even when the task seems trivial or the action is reversible, evaluate
+whether the user's reasoning is sound before executing. The
+instruction to question applies to the *why*, not just the *what*.

Let's see this in action.

Without the note asking about delting JSON/LD:

And now the final reveal, with the note:

The last thing that I'm thinking what Uncle Bob will say about it:

https://x.com/unclebobmartin/status/2045485259127423369

Maybe he just uses a different Claude than I do.

My Advice to New Developers

Jakub T. Jankiewicz — Wed, 16 Jul 2025 00:00:00 +0000

I'm a moderator in two communities on Twitter:

A frequent question people ask is about advice or tips for new developers. I stopped answering those questions, as they are a bit low-effort anyway. So I decided to write an article with my tips, and I will share a link when someone asks this question.

How to build zig book as PDF on Fedora?

Jakub T. Jankiewicz — Sun, 23 Mar 2025 00:00:00 +0000

I planned to learn Zig Programming Language, and found a Zig Book, but I wanted to read it on my reMarkable tablet, the problem was that officially Zig book is only in HTML format, and don't support building PDF file from source. That's why it's not part of its documentation. But with quarto publishing system, used by the book, it's possible to create a PDF file.

Below is the instruction how to create a PDF version of Zig Book yourself. The instructions are for Fedora GNU/Linux, because this is what I'm using, but building on Ubuntu or different Linux distro should be similar. You only need to find the right package names to install.

How to Help Charities and Your Ego?

Jakub T. Jankiewicz — Tue, 11 Mar 2025 00:00:00 +0000

Do you donate to charities? I do. And in this post I decided to explain, why it's worth bragging about it.

Documentaries for Programmers

Jakub T. Jankiewicz — Wed, 05 Mar 2025 00:00:00 +0000

People love watching movies, especially related to things they love. Developers are no different. And what is better to watch than documentaries about tools, libraries, and programming languages that we all use and love. Here is a list of documentaries created by CultRepo (formerly Honeypot) OfferZen, and some older documentary that you can find on YouTube.

Handling auto-indentation in Node.js REPL

Jakub T. Jankiewicz — Sat, 07 Dec 2024 00:00:00 +0000

When working on my Open Source project LIPS Scheme REPL, I wanted to have auto indentation, when you type multi line expression the text on the next line is indented. This article will describe a story of adding this feature, that require contributing to Node.js.

Batman website inspired terminal quiz app

Jakub T. Jankiewicz — Wed, 04 Dec 2024 00:00:00 +0000

Some time ago I was contacted by a person involved with NFT, to create a project inspired by 2022 Batman movie website.

How to Implement a Simple Markdown Parser?

Jakub T. Jankiewicz — Tue, 26 Nov 2024 00:00:00 +0000

I've had this idea some time ago to add limited support for Markdown into jQuery Terminal. I decided to use Peggy.js, a maintained fork of the PEG.js parser generator. I already used it for a few projects so I thought it will be the best tool for the job.

PEG is an abbreviation of so-called Parsing expression grammar, a well-known way to define language grammars and create parser generators.

In this article, I will show how to implement a simple Markdown parser to be used with jQuery Terminal. But if you follow along, you should be able to use it for different targets, like HTML.

How to Learn Design as a Web Developer?

Jakub T. Jankiewicz — Thu, 03 Oct 2024 00:00:00 +0000

People often ask me how to learn design. And since I'm mostly web developer, I get this question mostly from web developers. I decided to write this blog post to redirect people that ask this question. q In this article, I will show you 7 resources which will help you learn design as a web developer.

Hack for Syntax Highlighting of Source Code

Jakub T. Jankiewicz — Wed, 23 Aug 2023 00:00:00 +0000

If you're a web developer, work with JSON data format, and use Google Chrome browser, you have probably seen browser extension JSON Viewer. This extension adds Syntax Highlighting to JSON files viewed in the browser.

I've had a similar idea to display Syntax Highlighting but for JavaScript and CSS files, as part of the website. I just later realized that this is similar to JSON Viewer.

In this article, I will show how I did it and add this feature to my blog.

Welcome to my blog

Jakub T. Jankiewicz — Sun, 09 Jul 2023 00:00:00 +0000

The content of this blog will be everything that interest me. I'm mostly a programmer. I have a few Open Source projects. I'm also a Editor of Poslish Wikipedia, I'm hobby photographer and I design logos and logotypes (but not for the money). I'm interested in Math, computer graphics, comic books (including manga and anime). A lot of things will on subjects, that don't fit to my Polish blog "Głównie JavaScript", (Mostly JavaScript).