Intro
AI belongs in the delivery loop only where it respects the system around it. My workflow is intentionally boring: capture facts locally, preprocess deterministically, route reasoning to a task-fit model, verify against human and runtime evidence, then compound what worked into reusable standards. Lightweight models handle faithful extraction; heavier synthesis stays a separate, deliberate step. Prompts and outputs stay bounded; private material stays behind explicit boundaries; nothing important lives only in chat history.
Orchestrated agents, not one endless chat
The same org shape shows up in how I use coding agents. Mixed intent is where copilots fail - infrastructure, production change, and unrelated edits in one thread blur context and permissions. I run a small roster instead: an orchestrator that routes intent, and domain specialists that own a repo, stack, or client boundary. The orchestrator names the owner and stops; it does not implement another domain's code because the window is open.
The session loop
Eight steps, same order as the diagram. Skills and rules are not a separate essay - they show up where the loop says they belong.
-
1) Intent
Prompt
What you want done - one clear task. Mixed intent in a single thread is where copilots fail.
-
2) Orchestrator
Route only
Names the owning specialist and stops. It does not implement another domain's code because the window is open.
-
3) Domain skill
Playbook on disk
Open the markdown skill for the repo, stack, or client boundary. The editor injects the catalog; the agent reads the file, not chat memory.
-
4) Agent
Specialist
Executes in the workspace that owns the code. Always-on rules apply every turn - secrets, commits, boundaries - so no skill has to repeat them.
-
5) Workflow skill
Plan · verify
At most one workflow playbook: plan, increment, browser verify, review, or git. Cross-repo work gets a priority slot first; execution still happens in the owning workspace.
-
6) Ship
Land the change
Deliver a narrow version, inspect real output, refine from concrete failure - not a longer prompt.
-
7) FYS
Fix Yourself
When a mistake repeats, name it in one line and patch the one file the next session reads - skill Current line, rule, or trap note. Guardrails compound on disk.
-
8) CCC
Close Chat Checklist
Fixed shutdown: what changed, update Last and Current, land pending FYS, confirm backup. Chat is scratch space; disk is the system of record.
IT WORKS!
I run this loop every day - and use the same framing in consultancy: one clear owner per task, human gates where it matters, and guardrails that compound on disk instead of fading in chat history. After I send a prompt I can close the chat, open a new one with fresh context, and type a focused ask for project X - because CCC, FYS, and skills on disk did the handoff, the orchestrator routes to the right specialist and that agent knows what to do without me re-explaining the stack.