Let a model write Shopify Liquid, never let it publish. A deterministic linter with 275 controls gates what ships. No model sits in the decision path. Constraining the input beat banning the output by 52 points.
Class 94 is live.
🤖 Made with AI
1
11
Covers the Assets API race that blinds a lost-update guard, why prompt guidance fixes a gap but never moves a prior, and the rule that reported 518 defects that were not there.
Open to anyone, no login.
ddsboston.com/pages/dva-ai-s…
5
Claude Opus 5.5 keeps thinking always on, with effort as the dial. The default dropped from high to medium. Four API changes return a 400 on Opus 5 code, and one makes agents go quiet without an error. Nine interactive drills.
Class 95 is live.
🤖 Made with AI
1
20
We cover task budgets, per-message effort that keeps the cache, thinking blocks tied to their model, progress updates, safeguard fallback, and Anthropic's prompting patterns.
Free class. No account, no email capture.
ddsboston.com/pages/dva-clau…
24
Claude Code 2.1.327 reads MCP JSON configs from stdin now. Tokens stay out of shell history. Code Mode also runs alongside agent teams, with per-teammate models and limits. Brief also covers GPT-6.2 Astra and Sol, AI patch-to-exploit tests, and Microsoft Agent 365.
1
2
65
Read the daily brief (free, no gate): ddsboston.com/blogs/vibe-cod…
5
Claude Sonnet 5.5 solved 24 of 25 bug fixes against Opus 5.5 at 42% of the list-price cost per fix. Output ran 43% faster than Sonnet 5, with no gain to first token. Six claims checked, nine drills, and the permissions lesson from the one miss.
Class 96 is live.
🤖 Made with AI
2
1
28
Measured on a subscription against Opus 5.5 and Sonnet 5: fewer steps with less batching, six claims checked, the permissions lesson from the missed run, and nine drills with the method to re-run it.
The whole class is free and there is no gate on it.
ddsboston.com/pages/dva-clau…
1
14







