I gave the same 12-step refactor to Cursor, Claude Code, and Cline across 3 sessions, and the weird part was what each one missed. One model would nail the edits but forget the tests; another would write the plan and stall on execution. That gap is basically the whole story behind what is agentic ai. … Read more
On 14 March I gave the same brief to Claude Sonnet 4.6, GPT-5, and Cursor 0.50: pull a bug from a docs repo, summarize it, and draft a fix note. The one using an MCP server got to the right files in 18 seconds; the one without it burned 3 retries and hallucinated a path … Read more
I wasted three weeks on Cursor before realising the bottleneck wasn’t the editor. In April 2025 I switched the heavy lifting to Claude Sonnet 4.6 for spec work, kept Cursor 0.50 for code edits, and used GPT-5 only for cross-checking weird edge cases. That swap saved me about 18 hours over 12 sessions, mostly because … Read more
I ran the same 18-image brief through Midjourney v7, Midjourney v6.1, and a control set in SDXL over three sessions in April 2026. On my 50-prompt test pile, v7 finished first-pass usable on 34 prompts, versus 26 for v6.1, and it cut my average reroll count from 2.3 to 1.4. That’s the short version. The … Read more
I used to think Perplexity was just a polished search wrapper. Two months and 50 prompts later, that was only half wrong, which is annoying in the specific way these tools always are. I tested Perplexity AI across 3 sessions in March and April 2026, alongside Claude Sonnet 4.6, GPT-5, and Cursor 0.50. The short … Read more
It’s 2am. The agent loop just hallucinated my schema for the third time tonight, and the diff looked confident right up until it tried to rename a table that didn’t exist. That’s the annoying part of github copilot vs cursor: both can feel brilliant for 20 minutes, then one tiny context miss turns into a … Read more
On 14 March I gave the same brief to Claude Sonnet 4.6 and GPT-5: write a product launch angle for a boring B2B API, then rewrite it for a newsletter audience. Claude gave me a cleaner first draft in 42 seconds. GPT-5 took 58 seconds and produced the sharper hook, but it needed one extra … Read more
On 14 March I gave the same brief to Claude Sonnet 4.6 and GPT-5, then ran the draft through a Chrome-only Prompt API prototype I’d been poking at for 90 minutes. The result was blunt: the Chrome path was faster to demo, but it locked the workflow to one browser in a way I wouldn’t … Read more
The model everyone is praising is worse at shai-hulud themed malware found in the than people admit, especially when the task stops being generic summarization and starts looking like messy incident triage. I ran a small, boringly practical test to check that claim, and the results were less glamorous than the hype. This wasn’t about … Read more
I wasted three weeks on Cursor before realising it was the wrong tool for the kind of work I was doing. On April 7, 2026, I switched my spec-writing to Claude Projects, kept Cursor for edits, and moved the messy planning into Linear. That was the first week I started saying, out loud, that opus … Read more