links: site | repo | project page
AI Blueprint 1.5 is the release where I tried to cut down the context bloat without cutting out the parts that make the workflow useful.
The first instinct was to add a separate Lean mode. After testing it, that felt like the wrong shape. Users should not have to choose between “fast” and “safe.” The better version is just the default workflow getting smarter about what it loads, when it loads it, and how often it repeats itself.
what shipped:
- on-demand workflow context instead of loading the whole world up front
- the same Feature, Implement, Check, Try, Audit, and Complete commands
- adapter defaults focused on Claude Code and Codex, with Copilot and OpenCode still available
- smoother onboarding for repos that do not have an initial commit yet
- installer behavior that honors adapter choices already recorded during setup
- public docs that frame AI Blueprint as an AI coding workflow framework
The benchmark that convinced me was a controlled app build using the same spec against the older 1.4.1 workflow and the new 1.5.3 workflow. The new flow finished 16.5 percent faster, used 67.8 percent less cumulative input, and took 71.4 percent fewer model turns.
More important: the smaller workflow did not mean a weaker result. The 1.5.3 run preserved the requested feature, covered both screens, contained the mobile overflow, and kept the important gates intact.
That is the line I want Blueprint to hold. Less repeated context is good. Fewer wasted turns are good. But the win only counts if the code is still more grounded, more reviewable, and safer to merge.