Loop engineering

1. 01

   Pick one label that means “do this”

   In Contextli, dictate a task and add the label claude to it. That label is your agent’s inbox. Everything you dictate without it stays a normal note that no agent touches, so you can still think out loud.

   Use a verb-like label, not a topic. A label called “ideas” tells the agent what you were thinking about. A label called “claude” tells it to act.

2. 02

   Connect Contextli to Claude Code (once)

   On the computer that will stay on, run this in a terminal:

   terminal

   Copy

   claude mcp add --transport http contextli https://mcp.contextli.com/mcp

   Open Claude Code, type /mcp, choose contextli and sign in with your Contextli email and the 6-digit code it sends you. The agent can now list, filter and label your notes. Using Codex or another MCP client? Add the same address as an MCP server there.

3. 03

   Connect where the results go

   Add an email connector (for example Gmail) so the agent can write to you. The loop only works if you read its output away from the desk, and email is the one inbox you already check on your phone.

   Optional but worth it: connect ClickUp, Jira or Linear, so each task gets a ticket that holds the technical detail and your email stays short.

4. 04

   Start the loop

   Start a fresh Claude Code session, so every connector is loaded, and paste the prompt. Keep this one session for everything: because it has been running the whole time, a later “make that button bigger” lands on the right task.

   claude code: the loop prompt

   Copy

   /loop 10m Check Contextli for notes labeled "claude" that don't have the label "claude-queued". For each one, oldest first:
   1. Add the label "claude-queued" so it never runs twice.
   2. Log it: create a ticket (or add a line to loop-log.md) with the request in my words and what done looks like.
   3. Do the work. Record progress on the ticket.
   4. Email me only when it's done, needs my approval, or you have a question. Plain English, with a link or proof of the result.
   5. If a note is feedback on an earlier task, apply it to that task.
   Ask me before anything you can't undo: deleting data, spending money, publishing, or emailing anyone except me.

5. 05

   Keep the computer awake

   The loop runs inside your session, so the machine must stay on and the session open. On a Mac, run caffeinate -i in a separate terminal tab and keep the lid open (or use a desktop). On Windows, set sleep to “never” while plugged in.

6. 06

   Test it in one minute

   Dictate Email me the word hello., add the label claude, and put your phone away. If “hello” lands in your inbox within about 10 minutes, the loop works. From then on: talk, label, walk away, read email when you have a minute.

• Mark on pickup, not on finish. The agent adds claude-queued the moment it takes a note. Its query is “labeled claude, without claude-queued”, so nothing is missed and nothing runs twice.
• Never filter by date. Mine once asked for “notes since” a timestamp; the filter returned nothing while two new notes were waiting. The label check has no such gap.
• Ticket before work. The request is written down in your words before anything changes, so there is always a record to compare the result against.
• Report in the same run. Every finished task is emailed in the check that finished it. Results left in the chat window are results you never see.
• Proof in the report. A link to the live page, a screenshot, the ticket id. “Deployed” is a claim; a link you can tap is evidence.
• Track what was handled by id. Never by “unread”. If you open a reply on your phone first, an unread filter silently hides it from the loop.

• Plain English or it does not get read. Agents write updates for engineers. One line in the prompt, “technical detail goes in the ticket, email me plain English”, turned walls of file names into decisions I can clear on my phone.
• One inbox for replies. Feedback arrives as a voice note or an email reply. The loop checks both, and a rule tells it to apply feedback to the earlier task rather than start a new one.
• Cap the helpers. The loop can spin up helper agents for big jobs. More than three at once and they start stepping on each other’s files.
• Money needs a ceiling. After one batch job spent far more on API calls than I expected, the rule became: anything that will cost more than a few dollars emails me an estimate first.
• Restart weekly. Recurring loops expire after seven days. A loop that quietly stopped on day eight looks exactly like a quiet week.

“The real test: can you say what got handled today without opening the laptop once? If not, it is not a loop yet. It is a faster notepad.”