Agentic SDLC
I could describe how to apply AI across the development lifecycle. I hadn't actually run a product through it end to end — ideation to release — with AI doing the execution and me directing it. That's a credibility gap, and no amount of reading closes it.
So I built the practical experience by building the thing itself: a shared set of agents, skills, rubrics and templates that runs a product initiative through six gates, plus a cross-cutting retrospective. This is what's actually in it.
Market research, then a 4-lens idea council — feasibility, market, user-value, risk — scores the idea independently before a synthesis step reconciles the views into a single recommendation.
A PRD gets authored and put through single-critic review, then the MVP scope is extracted from what survives — the smallest slice that tests the idea, not the whole backlog.
Tech options, tech design, UX design and test strategy are authored, then reviewed by a 5-lens design council — architecture, security, accessibility, test strategy, UX. An optional design-council-challenger brings an unscored outside view from a non-Claude model, so the review isn't grading its own homework.
Two phases. Phase 1 turns the accepted design into Linear issues — a Project per scope area, plus stories, bugs and spikes. Phase 2 runs a dev/review/SDET agent loop that takes a named issue end to end: branch, code, PR, review, test, through to Ready to Merge. Linear's GitHub integration drives the transitions git can observe — a human's merge is what actually marks an issue Done.
Checks what the pipeline actually did — integration/e2e, SAST, dependency scan, accessibility scan — against what the test strategy promised it would do, folding the result into a release readiness check.
Drafts release notes, confirms rollback, monitoring and feature-flag state are in place, and records the human go/no-go for production. The system prepares the decision; it doesn't make it.
Mines a corrections log for repeated patterns across every gate and proposes rubric, prompt and template diffs — for human approval, not automatic adoption. This is how the toolkit gets better instead of just getting used.
Why It Holds Together
Every gate ends the same way: a rubric-scored, evidence-quoted verdict, with an explicit human decision point before anything moves forward. Critics never author and authors never critique— the two roles don't blur, even under pressure to move fast.
And revision loops are capped, not run to convergence: if a critic and author can't agree within a set number of rounds, it escalates to a human rather than looping forever chasing an automatically-approved answer.
See the projects →