How our consulting pods build software with AI agents: gated phases open and close the engagement, and a continuous per-spec build loop — Intent → Delegate → Discern — runs everything in between. A fictional engagement (Harbor Mutual, a regional insurer) runs alongside the standard as a worked example, phase by phase.
Agents write code fast now. That doesn't make software delivery faster on its own — the hard part moves to two places: saying clearly what you want, and checking that what came back is right. This standard is how our pods run a client engagement around those two jobs.
The work runs as one short loop, repeated for every piece of work, and it reuses the name's initials:
- Intent — a human writes a short, testable description of one piece of work (a spec).
- Delegate — an agent plans and builds it, inside bounds a human set.
- Discern — before anyone trusts it, automated gates and a non-author human check it's good.
Gated phases open and close the engagement; the loop fills the middle. No checking is just vibe coding — the loop closes it. That's what makes the speed safe to put in front of a client.
- Brand new to the method? Read the loop cheat-sheet (20 seconds), then the glossary for any word that trips you, then The Delivery Standard itself.
- Prefer a story to a spec? Follow The Harbor Journey — one full engagement, start to finish, as a continuous worked example.
- A pod member joining an engagement? The Delivery Standard front to back, then The Team for your role, then the build loop. Keep the glossary and anti-pattern field guide open.
- A client stakeholder or sponsor? Read In 30 seconds above, then the FAQ — it answers the questions you're about to ask (data, decisions, what you'll see, what happens when we leave).
- Leading or selling the engagement? The Delivery Standard section 12 (commercial) and section 1 (the shape), plus the FAQ for client objections.
- Migrating a team off Scrum? The FAQ ("Do we still run sprints?") and the anti-pattern field guide (especially "cutting the safety net too early").
The standard is published at https://mckruz.github.io/intent-driven-development/
| Page | What it covers |
|---|---|
| The Delivery Standard | The master document — the whole method in 14 sections |
| The Harbor Journey | The whole worked example as one continuous story — 10 stops, start to finish |
| The Team | Old roles to new roles, the concrete job of each, scaling from one pod to many |
| Phase 0: Discovery | Fixing the problem · worked example |
| Phase 1: Requirements | The signed baseline · worked example |
| Phase 2: Design | Options into signed decisions · worked example |
| Phase 3: Foundation | The factory gets built · worked example |
| The Build Loop | The heart of the method · worked example |
| A Day in the Loop | The build loop on film — four narrated clips (Orient · Delegate · Verify · Discern) and the full module |
| The Rails | The agentic CI/CD & DevOps pipeline — the gates every change rides · worked example |
| Phase 7: Documentation | Proving a stranger can run it · worked example |
| Phase 8: Deployment | The rehearsal, the ceremony, go-live · worked example |
| Phase 9: Monitoring | Alerts from real baselines, the drill, the retro · worked example |
| Phase C: Close & Transfer | The close gate, the clean exit, the harvest · worked example |
Plain-language pages for getting in the door fast. No engagement context required.
| Page | What it covers |
|---|---|
| The artifact flow | What each phase receives and produces, where each file lives, and the handoff chain end to end |
| Loop cheat-sheet | The wall card — Intent → Delegate → Discern, the checking ladder, Do/Don't |
| Glossary | Every term of art in the standard, defined in plain words |
| FAQ | Honest answers to what clients and new pod members actually ask |
| Anti-pattern field guide | The nine ways this goes wrong — symptom, cause, fix, prevent |
The standard ships as a working Claude Code harness — the engagement starter that Phase 3
installs into a client repo (GOLD-STANDARD §6 the harness standard, §10 the kit). It's the
method made executable: a governance CLAUDE.md, the spec template, permission settings, blocking
Stop/review hooks, six model-tiered subagents, six team-practice skills, the five CI rails, profile
rubrics/rulesets, an eval module, and infra starters. Teams install it via the claude-code-sdlc
plugin (/sdlc-setup); kit/ in this repo is the canonical source.
| Page | What it covers |
|---|---|
| The Kit — install & map | What's in the harness, where each file installs, the adapt order, and the shakedown drills |
| Harness research | The cited rationale behind every choice, plus the STABLE / NEWER / BLEEDING-EDGE maturity tiers |
| Call map | How the pieces wire together — hooks, agents, and the five rails, visualized |
GOLD-STANDARD.md— the master document (14 sections);GOLD-STANDARD.htmlis its render.index.html— the published landing page and site navigation.docs/— the deep-dive markdown sources of record (phase-*,build-loop,the-rails, and their*-exampleworked examples), the four plain-language reference pages (glossary, FAQ, cheat-sheet, anti-patterns — each with a paired HTML render), andjourney/team.docs/companion/— the interactive per-phase walkthroughs. These are the published HTML pages the site links for each phase (the idea, the mechanics, and the worked example on one page).docs/harness-kit/— the harness research (RESEARCH) and call-map (CALL-MAP).kit/— the installable Claude Code harness; the source of truth (theclaude-code-sdlcplugin bundles a synced copy).retros/— one file per engagement: what the harvest loop changed and why.PROGRESS.md— working notes for where the standard goes next.
The site is served with
.nojekyll, so only committed.htmlfiles render — the phase deep-dives exist as markdown sources, and their published form is the companion walkthrough. Link the companion pages (notdocs/<page>.html) when pointing at a phase online.