One Markdown File Made My AI Agent 23 Points Smarter

16 June 202616 June 2026 by susiloharjo

Last week I read a paper that made me re-evaluate everything I have written about AI agent optimization. Microsoft and three Chinese universities published a method called SkillOpt. The result: a single Markdown file, between 300 and 2,000 tokens, lifted GPT-5.5 by an average of 23 points across six procedural benchmarks. No fine-tuning. No new model. No extra tools. Just a Markdown file that gets fed to the agent as context at inference time.

The skill beats handwritten instructions, one-shot LLM-generated instructions, and four specialized training methods (Trace2Skill, TextGrad, GEPA, EvoSkill). It works in Codex. It works in Claude Code. It transfers across model sizes. A spreadsheet skill trained in the Codex loop lifts Claude Code to the same level as a skill trained directly in Claude Code.

After reading the paper, I stopped adding features to my AI agent for a week. I started writing skill files instead. Five of them. All under 1,000 tokens. All producing measurable improvements in my daily work. This post is the five skills and the pattern I now use by default for any procedural task.

I Tried GROW Coaching in My 1:1s. It Cut Them in Half.

17 June 202615 June 2026 by susiloharjo

I Tried GROW Coaching in My 1:1s. It Cut Them in Half.

Last week I ran a 1:1 that lasted 12 minutes read more about leadership and delegation. The engineer walked out unblocked, with a clear next step, and didn’t ping me for the rest of the day. A month ago, the same engineer would have walked out of a 45-minute 1:1 with a vague explore one-page procedure for team clarity “I’ll think about it” and pinged me twice before lunch.

The only thing I changed was the questions I asked. I stopped solving problems in the meeting. I started running them through a 30-year-old coaching framework called GROW — Goal, Reality, Options, Way Forward.

How to Fix Linux Menu Icons for Portable Electron Apps

14 June 202613 June 2026 by susiloharjo

Linux desktop menu showing an Electron application shortcut with its icon properly rendered beside the system tray

How to Fix Linux Menu Icons for Portable Electron Apps

You downloaded a Linux app as a portable folder or a raw binary — Antigravity IDE, a portable VS Code, a side-loaded Electron tool. You wrote a .desktop file, dropped it in ~/.local/share/applications/, and the app is gone from the menu. The launcher grid shows nothing. The search box finds nothing. You open the .desktop file and the Icon= line points at a perfectly valid PNG discover how 3 bash aliases saved me 10 hours last week, yet the system acts like the icon does not exist learn how to resize your Ubuntu VM disk step by step.

The problem is almost never the icon file. The problem is whitespace in the path.

GNOME, KDE, and most freedesktop-compliant desktop environments parse .desktop files with a config parser that is unusually strict about spaces. A path like /home/ubuntu/Downloads/Antigravity IDE/resources/app/resources/linux/code.png will silently fail to load for two reasons: the unquoted space inside Icon= breaks the parser, and even when the parser survives, the desktop file validator (which runs when update-desktop-database regenerates the cache) treats the entry as malformed and drops it.

I burned an hour on this the first time. The second time I wrote the fix down. This is the four-step version that survives every Electron portable app I have thrown at it.

I Set 2 Playwright Goals at 4 AM and Both Beat Fajr

14 June 202613 June 2026 by susiloharjo

Browser window with Python Playwright code on a desk in the dim light of pre-dawn, a closed Quran beside the keyboard

I Set 2 Playwright Goals at 4 AM and Both Beat Fajr

This Saturday morning I woke up at 4 AM with the kind of half-baked idea that only survives a brain that hasn’t fully booted. I wanted to see if I could give an LLM a real browser three AI code review tools I run before every PR, a goal written in two sentences, and walk away. Not a toy. Not a sandbox. LinkedIn. Tokopedia. The real web, with logins and JavaScript and the kind of DOM that makes scrapers cry learn how rogue AI processes can eat your server.

By the time I finished praying Fajr and closed my Quran, both jobs were done. One returned 23 LinkedIn profiles as structured JSON. The other returned 14 G-Shock listings under Rp 1.000.000 from Jakarta sellers offering instant payment. I had not touched the keyboard.

This is the prompt pattern that made it work — and the three things I almost got wrong.

5 Things I Stopped Doing as a Lead

12 June 202612 June 2026 by susiloharjo

I led a small engineering team for two years. We shipped on time, the code was clean, and the standups were short. By every metric I cared about, the team looked fine. Then three of my best engineers quit within a month. None of them had a single conversation with me before they handed in their notice.

The exit interviews told me what I should have seen earlier. They were not burned out by the work. They were burned out by the way I gave it to them.

I spent the next six months rebuilding how I delegated. I read four books on engineering management. I ran a one-on-one feedback survey every Friday for twelve weeks. I shipped a working team again, and the difference was not subtle.

Here are the five delegation habits I had to kill. If you are a tech lead and any of these feel familiar, the exit interview pattern will too.

Why a Simple If-Else Can Beat an LLM

12 June 2026 by susiloharjo

Code logic: simple if-else flow chart contrasting with LLM API endpoint

Why a Simple If-Else Can Beat an LLM

TL;DR: When you can describe the inputs and the expected outputs in advance, you don’t need a model — you need a function. Here’s the principle, the proof, and the one case where the principle breaks.

A teammate burned $47 of API credits last quarter on a “smart” classifier. The job: sort incoming support emails into four buckets (billing, technical, account, other) and route them to the right Slack channel. The model nailed it about 91% of the time. The remaining 9% it was confidently, hilariously wrong — sending a billing dispute to the technical channel, an outage report to “other.”

I replaced it with a 40-line Python script using if and a handful of keyword checks. It runs in 12 milliseconds per email, costs $0, and gets the same 91% — except the 9% it gets wrong are predictably wrong, so we know to watch them. The classifier used to hallucinate categories that didn’t exist. The script never invents a fifth bucket.

That’s not an edge case. That’s the principle: when the parameters are already known, deterministic code is the right answer. The question is why this works, and when it stops working.