CAMII CMS — Agentic Engineering Playbook

A curated, real-world harness for building software with AI agents — the skills, subagents, slash-commands, hooks, and rules that were authored and proven while shipping CAMII CMS, a production .NET 8 / Blazor Server maritime crew-management system (~15 EF Core DbContexts, dozens of KPI/workflow modules).

This repository is not the application. The app has been removed. What remains is the agentic engineering layer: the reusable machinery and the write-ups that explain why each piece exists and what problem in the build it solved. It is meant to be readable by a newcomer to AI-assisted engineering and useful to an experienced SWE looking for battle-tested patterns to steal.

Sanitized public copy. All credentials, keys, internal documents, and real company/crew data were removed; any names, vessels, or emails that remain are fictional sample data. © All rights reserved — provided for review and learning, no license granted to reuse.

---

The five pillars

#	Pillar	The idea in one line	Start here
1	Context engineering	Externalize what the model must know; treat disk as the contract, not the context window.	docs/01-context-engineering.md
2	Harness engineering	Shape the agent's action space — skills, commands, agents, hooks, rules — so the right move is the easy move.	docs/02-harness-engineering.md
3	The Butler (orchestration)	A stateful lead-architect agent that remembers work across sessions and dispatches subagents on the right model.	docs/03-orchestration-the-butler.md
4	DB-guard & migration safety	Encode your #1 footgun into the harness; prove schema will actually deploy — evidence-first.	docs/04-db-guard-and-migration-safety.md
5	Verification, gates & learning	Never trust "it builds"; gate every commit, verify with evidence, and let the harness learn.	docs/05-verification-gates-and-learning.md

New to all of this? Read docs/00-the-camii-case-study.md first — it sets the scene and the philosophy.

---

Repository map

.claude/                     ← the harness (this is the substance)
  rules/                     ← the project "constitution" the agent always obeys
  skills/                    ← on-demand playbooks (one folder each, SKILL.md)
  commands/                  ← slash-commands: entrypoints & pipelines
  agents/                    ← specialized read-only / scoped subagents
  hooks/                     ← shell hooks that automate the loop (session feed, security audit)
  butler/                    ← durable orchestrator state (STATE.md + task records)
.cursor/                     ← the SAME harness ideas ported to Cursor (tool-agnostic proof)
CLAUDE.md                    ← the real project-memory file that governed the build
docs/                        ← the write-ups: what each pattern is and why it worked
examples/                    ← real, unmodified artifacts behind the two hardest lessons
  db-guard/                  ← the EF Core .Contains()/OPENJSON footgun + the fix
  migration-deploy-gate/     ← a migration that built green but never deployed

How to read it

• If you're learning agentic engineering: start with docs/00, then walk the pillars in order. Each doc points at the concrete artifact in .claude/ so you can see the pattern, then the real code/config that implements it.
• If you're an experienced SWE: skim the pillar tables, then go straight to the artifacts. The highest-signal reads are .claude/commands/cms-butler.md, .claude/skills/verify-staging-migrations/SKILL.md, .claude/skills/subagent-orchestration/SKILL.md, and the two examples/.
• If you want to fork it: the .claude/ tree is a working harness. Adapt the rules/, keep the hooks/ and the butler pattern, and replace the domain- specific skills/commands with your own.

What this repository demonstrates

For a reviewer or hiring manager — this is the engineering methodology behind a real, shipped production system (.NET 8 / Blazor Server, ~15 data contexts, dozens of modules), built AI-assisted under my direction. What it evidences:

• Systems thinking — factoring repeated work into a generic engine (page-harness, one loop that drove 9 pages) instead of copy-paste; layering the whole workflow as rules → skills → orchestration → gates.
• Turning production failures into permanent safeguards — the EF .Contains()/ OPENJSON footgun and a migration that built green but never deployed each became a rule, a subagent, and an evidence-first gate (docs/04 + examples/).
• Engineering judgment — evidence-first verification ("no verdict without pasted output"), matching model cost to the cost of being wrong, strict subagent containment so multi-agent work stays controlled.
• Directing AI agents as a force multiplier — designing the context and harness that let agents ship safely, which is a distinct and increasingly valuable skill from writing every line by hand.

I authored the CAMII-specific harness (the Butler orchestrator, db-query-guard, the migration gate, page-harness, the rules/, the hooks, CLAUDE.md). Community skills are marked with an origin: field and composed into the workflow — not claimed as mine (see Provenance below).

Author: · GitHub @jttd3v · <LinkedIn / email>

---

Provenance & honesty

This is a real working set, so it mixes original work with community building blocks — and it says which is which:

• Authored for CAMII CMS (the original engineering): the butler orchestrator, the db-query-guard agent and ef-core-contains-guard rule, the verify-staging-migrations gate, the page-harness engine and its per-page instances, the project rules/, the hooks, CLAUDE.md, and the .cursor/ harness.
• Third-party / community skills, kept because they were genuinely part of the workflow, with their attribution intact: any .claude/skills/*/SKILL.md whose frontmatter carries an origin: field (e.g. origin: ECC, origin: community). These are not presented as original work.

If you use this as a portfolio, be equally upfront that the build was AI-assisted — directing this machinery well is itself the skill on display.