links: site | repo | project page
AI Blueprint keeps turning into less of a prompt pack and more of a workflow system.
The new pieces are all about giving the agent a better lane before it touches code.
/debug is diagnosis-only. It reproduces a failure, narrows the cause, and
reports evidence without editing source, tests, config, Blueprint state,
branches, or commits. If the diagnosis becomes repair work, that moves to
/fix or /implement.
/audit now has focused lenses: quality, security, performance, and tests.
The old broad audit still exists, but the focused form matters when I want a
security pass to stay about security instead of turning into a grab bag of
style notes.
/discovery is the optional planning conversation I wanted without replacing
manual plans. It can help draft project-plan.md and build-plan.md, but it
stops for review before writing them. That keeps the workflow flexible: use
Discovery when a product is fuzzy, or skip straight to hand-written plans when
the shape is already clear.
The maintainer side got stricter too. Blueprint now has deterministic routing evaluations and explicit live-agent scenarios for the workflow boundaries. The live tests cost tokens, so they stay opt-in, but they catch the bugs that only show up when an agent is actually asked to behave.
The public docs are already updated at
ai-blueprint.dev. The npm package is still
create-ai-blueprint@0.6.0 as of this post, so the next release is the package
publish and clean install check.