← Oleh Shevtsov

CASE STUDY · HOW I SHIP

Agentic SDLC

An AI-driven delivery process for small product teams, packaged as code: Claude Code plugins take an idea through specification, architecture and tasks; an Archon workflow — “the factory” — turns each task into a tested, reviewed, merged and deployed pull request.

1_mvp_sdlc on GitHub ↗Reference workflow ↗Built with it: Pilot →
Spec firstIdea brief → PRD → C4 architecture + ADRs → data model → task briefs.
Tests firstTDD commit order is checked in CI, not just requested.
GatedA phase gate refuses to skip steps; hooks protect main, secrets and tags.
Honest reportsA rolled-back deploy is never reported as a success.

The problem

Coding agents are fast, and that speed hides predictable failures: stale context, changes outside the requested scope, “done” claims without evidence, and deploys nobody checked. At ManageBac I worked AI-first in a large Rails monolith with Claude Code, Copilot and custom MCP servers, and saw that the prompt was never the bottleneck — the process around the agent was.

So I engineered the workflow instead of polishing prompts, and used it to build Pilot end to end.

How it works

  1. Specify. Claude Code plugins (sdlc + my Rails adaptation mvp-sdlc) produce the artefacts in the project repo: idea brief, PRD, C4 architecture and ADRs, data model, task briefs. A feature starts from a Trello card.
  2. Plan. Each task goes to the Archon factory workflow. It shows a plan and waits for a person's “ok”.
  3. Build. Tests first, then implementation until green. The draft PR is reviewed through several lenses — code, tests, docs, seams — and P0/P1 findings block the merge.
  4. Merge & deploy. CI green and conventions checked → squash merge into develop → automatic dev deploy → health check and error-tracker check → automatic rollback when either is red.
  5. Release. A person merges develop into main and tags vX.Y.Z. Production deploys only from that tag and is watched for ten minutes.

Guards, not good intentions

  • Phase gate (sdlc-gate): a step can't start until the previous artefact exists.
  • Hooks that block pushes to main, reading secrets and moving published tags.
  • CI checks for TDD commit order, branch names, PR titles and a non-empty changelog on release PRs.
  • Agent tool contract (contracts.yaml) linted in CI together with skills, hooks and plugin versions.
  • Supervisor that answers only routine gates from a policy file, escalates everything else and writes one ledger line per decision.

What it delivered

  • Pilot — Rails API + React SPA, live in production, built and operated with this toolkit.
  • Project templates for a new Rails API and React app, and bin/project-add: one command from a name to repos, environments, secrets, a Trello board and a report.
  • A public, provider-agnostic reference workflow distilled from it.

Learning behind it

Model Context Protocol: Introduction and Advanced Topics (Anthropic Education), and SULICOM AI Academy workshops on Agentic Engineering Workflows and Harness Engineering. See the certificates →

“The work is complete when the evidence matches the claim.”