AI Dive

I pointed REA, the viral reverse-engineering skill, at my own live site from a throwaway sandbox. Under 13 minutes and zero write requests to show that any website could fill my mailing list. 15 findings, one real leak, fixed and shipped as a PR. www.youtube.com/watch?v=63nm...

I wrote a Claude Code skill for our weekly release notes. Called by name, it ran every time. Asked in plain words in 5 fresh sessions, it loaded 0 times. I changed one line, the description, and the same request loaded it 5 of 5. www.youtube.com/watch?v=usYy...

I typed ultracode on a sixteen-line sandbox and nine agents showed up for a read-only review. On the paired test I walk through, plain and ultracode both passed 136 of 136 checks: plain used 8% of the 5-hour meter, ultracode 18% and never stopped. www.youtube.com/watch?v=f-Yg...

I built the same feature twice on one repo, with Superpowers and with Matt Pocock's Main Flow, and read the bill per stage. Superpowers driven: $3.74. Main Flow: $2.04. $2.19 of the gap is one reviewer subagent in executing-plans. www.youtube.com/watch?v=vWRc...

I rebuilt Shopify's four Helix gates as a Claude Code Stop hook on a real repo and ran it 50 times with each gate switched off in turn. Acceptance tests took the pass rate from 70% to 84%. Two reviewers after the tests: 82%, at three times the cost. www.youtube.com/watch?v=_Lrw...

I ran 182 Claude Code subagent tasks with Haiku 5.5, Sonnet 5.5 and Opus 5.5 as the worker and read the plan meter after every block. 140 Haiku runs moved it zero points. Haiku passed 70 of 70, and nothing runs on it until an agent file says so. www.youtube.com/watch?v=kfLr...

I read a month of my own Claude Code logs after the September limit cut: 455 sessions, 63,000 requests. Subagents took 48% of every token, and each launch starts at 47,000 before it does anything. The MCP server tip measured 1,350 tokens for 51 tools. www.youtube.com/watch?v=CYcl...

I ran Anthropic's AI-native SDLC playbook end to end on one repo, 16 timed sessions, $12 total. Same bug fixed directly: 2:13, 70 cents. Through all six gates: 11:31, $3.46. Three of the six pay that back. www.youtube.com/watch?v=bL08...

I ran 116 Claude Code sessions on a real repo and graded every "done" with tests the agent never saw. The built-in /verify came back PASS 24 times out of 24, cost 2.5x more on Opus, and changed no outcome. www.youtube.com/watch?v=oL2g...

I deleted the 177-line CLAUDE.md from a real repo and ran the same five tasks 44 times. One rule broke: on a new file, 3 of 4 runs hardcoded a string instead of adding the translation. Every other rule held, the code next door already taught them. www.youtube.com/watch?v=1oJS...

I installed the eight Claude Code repos that top every must-have list onto one real project and ran the same tasks with and without each. All eight cost under 7k tokens at session start. The three MCP servers they ship were called zero times in 63 runs. www.youtube.com/watch?v=73Sz...

I wrote a 42-line Claude Code Mod that redacts secrets from tool output before the model reads them. 28 ms per call, zero tokens. Then I put a 15 second sleep in it: after 10 s the engine skipped the guard and ran the command anyway. www.youtube.com/watch?v=YGxV...

okama's maintainers found eight agent-written plans rendered as public Read the Docs pages, committed exactly as Superpowers intends. Where specs go is solved. Who owns them and when one is dead is not, so I wrote down five rules for that. www.youtube.com/watch?v=IKfU...

Fresh install of everything-claude-code (250k stars), then I measured what it puts in context before you type: about 27k tokens, 14% of a 200k window, every session. Mostly descriptions of 286 skills you will never call. Five parts are worth copying out. www.youtube.com/watch?v=iu9-...

JetBrains found rtk made Claude Code sessions 7.6% more expensive. My rtk dashboard says 11.6M tokens saved. Both are right: the dashboard counts shell bytes, not the bill. I ranked the four token savers I run by what they move: www.youtube.com/watch?v=Dees...

Fable 5.1 cache reads are $0.25 per million. Opus 5's are $0.50. That is the only line in the pricing where the expensive model is the cheap one, and it only shows up if your agent keeps re-reading the same transcript. www.youtube.com/watch?v=azLx...

Measured it: same model, trimmed shell output, 41.6% less to read over 25k commands.

So this is the start of a series I've been working on to explain, in a coherent way, how we should be thinking about actually using LLMs to be useful, instead of the LinkedIn-esque "I burned $5000 on Fable AND--" The thumbnails aren't going to get any less dumb, though. That's the Ed Promise.

I read all 14 skill files of the Superpowers plugin for Claude Code. The method is good on feature work. The cost is a ceremony that never switches off: on a two-line fix, the full process is slower than typing the fix yourself. youtu.be/l7P_TQzyd7Q

Anthropic graded 400,000 Claude Code sessions. Managers beat software engineers on verified success, and novices reach 15% vs 28-33% for everyone above. I ran the same task with a novice prompt then an expert prompt to see the gap live: youtu.be/wq2QN_MvetY