Hashline editing: The failure
One engineer. The most fashionable trick in AI coding. And the eval that says it doesn't help. In Episode 02, I test the first of many harness mechanisms that are supposed to win the bet: hashline editing — the anchor-based edit trick a viral blog post claims took a cheap model from 6.7% to 68.3% on an editing benchmark. Same model, different edit tool, ten times better. So I built it. Two weeks ago. And then — this is the embarrassing part — I shipped it without ever testing whether it worked on my models. This episode is me finally testing it. On camera. With the bill shown. What's in it: The trick. What hashline / anchor editing actually is, and the failure mode it's built to prevent — the agent editing a file that drifted underneath it. The confession. I added a mechanism to the harness before I measured it. The someone who didn't test it is me. The someone is always me. The experiment. Four cheap open-source models. Three conditions — hashlines off, optional, forced. Same tasks, same grader, no cherry-picking. The result. Three completely different responses to the same toggle. One model went 1/8 → 6/8 — six times the score, and ten times the bill (3¢ → 30¢). One didn't care at all. The model I actually ship got slightly worse. The receipt. Same task, same perfect 8/8 score: 1.3¢ on DeepSeek v4, $1.37 on GLM 5.1. A 100× cost gap living inside the "cheap" tier. The verdict: the fashionable trick doesn't help. Not universally, not for free, and not for the model I run. A harness mechanism only pays off when the task's failure mode is the exact one it fixes — and most of the time, it isn't. The labs sell you the expensive model. The influencers sell you the trick. I'm selling you the eval. It's going to be Ugly. That's the whole point. ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬ 🔔 Subscribe so you don't miss the part where the trick I built doesn't work ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬ LINKS 🌐 Platform → https://ugly.bot 🛠 Studio (download) → https://ugly.bot/studio 📦 Framework → https://www.npmjs.com/package/ugly-app 💻 GitHub → https://github.com/justinmann 💼 LinkedIn → / thejustinmann ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬ #codingagent #claudecode #opensource #glm #deepseek #hashline #llm #softwareengineering #developertools #indiehackers #aicoding #buildinpublic #UglyStudio

Unfortunately, I Was Right

I Think They Are Lying To You

One man just liberated Fable... and now it’s illegal

They Lied to You About AI (This Study Proves It)

I am done with Golang

What 6 months of AI coding did to my dev team

The problem with using AI to help you learn (programming, and in general)

I Don't Think I Can Go Back To Windows...

Passkeys Explained: Are They Actually Better Than Passwords?

The Hidden Cost of AI Coding That's Destroying Engineering Teams

The Bet: One Engineer vs. Every AI Coding Agent

Linus Torvalds: AI Is Changing Linux Fast

Revenge of the Sith But the Jedi Are Smart

My Golden Retriever Heals a Terrified Rescue Kitten in Just 3 Meetings!

I Made Opus 4.8 and Fable 5 Build the Same App (RAW RESULTS)

I Hacked This Temu Router. What I Found Should Be Illegal.

Frankreich – Senegal Highlights | Gruppe I, FIFA WM 2026 | sportstudio

5 CLI Tools That Actually Changed How I Work in 2026

The Open Source community is collapsing

