The autonomous workflow & a master cheat-sheet
The finale of your revision. We’ll retrace the full professional loop — from a ticket to a merged pull request — then hand you a single cheat-sheet and a short self-check so you walk into week three sure of your footing.
Weeks one and two peaked with a real product built the professional way. Let’s make sure the shape of that workflow is second nature — because week three automates big chunks of it.
01The end-to-end loop
The professional workflow is a loop that starts with a unit of work and ends with reviewed, merged code — every step tracked.
02FeatureDev’s discipline
Left alone, an agent rushes to code. FeatureDev (Anthropic’s plugin) forces a disciplined, staged march — explore, plan, and get sign-off before writing code, then build and verify. It is the trust-but-verify mindset turned into a repeatable process.
When you hand work to several agents at once, a shared plan is the only thing keeping them from colliding. FeatureDev’s “agree the plan first” habit becomes the plan.md that coordinates a whole team of agents in Class 12.2.
03What we actually shipped
The Pre-Legal SaaS pulled every thread together into one real product:
Tickets + a custom skill
A backlog PL-4–PL-7 in a stakeholder’s voice, and your own Cerebras skill for fast, structured LLM calls.
A working platform
Multi-user auth, saved documents, an AI chat that fills legal templates — FastAPI + Next.js + SQLite, shipped as reviewed PRs.
Even after its own review, Claude dropped a specified legal disclaimer. Verify against the ticket — the agent’s confidence is not proof. That instinct is what keeps a swarm honest.
04The master cheat-sheet
| When you want to… | Reach for |
|---|---|
| See or reclaim the context window | /context, then /compact or /clear |
| Give the agent standing orders | A lean CLAUDE.md |
| Connect an external system | An MCP server (Jira, GitHub, docs) |
| Teach a repeatable skill | A skill in .claude/skills/ |
| Add a shortcut you type | A slash command (Class 12.3) |
| Offload a noisy task, protect context | A sub-agent (Class 12.4) |
| Do something automatically on an event | A hook (Class 12.5) |
| Bundle & share a whole setup | A plugin + marketplace (Class 12.6) |
| Build for hours, hands-off | A Ralph loop / swarm (over Git!) |
| Undo a bad turn | Rewind / a clean Git commit |
05Quick self-check
Answer these out loud. If any is fuzzy, the linked page is your revision target.
Why does an agent “forget” unless you use CLAUDE.md?
Hint: what happens to the transcript each turn? (See 11.2)
MCP vs. skill vs. plugin — one line each.
Connect / teach / bundle. (See 11.3)
Where does “trust but verify” sit on the control↔autonomy dial?
Left of centre — and why? (See 11.1)
What is the safety net under YOLO mode?
If you can’t answer instantly, revise 11.3 before the finale.
If those four came easily, you have the whole of weeks 1–2 in your head. Time to go big: Class 12 opens with the mindset for the finale — controlled chaos.
✓ Key takeaways
- The pro loop: ticket → FeatureDev (plan/build/test) → reviewed PR → repeat.
- Staging (“agree the plan first”) becomes the shared
plan.mdthat coordinates many agents. - We shipped a real multi-user SaaS — and still had to verify against the ticket.
- Keep the cheat-sheet handy: it maps every “I want to…” to the right tool.
- Pass the self-check, then step into controlled chaos in Class 12.