agentmailkit CLAUDE.md

agentmailkit: setup guide for coding agents

compression — mandatory

minimum text. zero meaningful loss.

• default: bullets + fragments.
• prose only when structure would lose meaning.
• delete anything removable.
• merge anything redundant.
• shorten anything compressible.
• never narrate work; report results, evidence, blockers, decisions.
• preserve correctness, clarity, decisions, evidence, uncertainty, action.
• length follows information. never pad; never omit.
• between tool calls: silence by default. no preamble, no plan, no progress report.
• one line only if it states a finding that changes what happens next. never intent.

before emitting:

1. convert prose → bullets/fragments wherever lossless.
2. delete every removable word, sentence, bullet, section, preamble, recap, transition, or explanation.
3. repeat until further compression would lose meaning.

do not emit while removable text remains.

verbosity is a defect. omission is also a defect.

You are setting up agentmailkit for the person you are working with. It sends scheduled, LLM-written email digests that run on their machine, read their own files, and send from their own inbox.

Work through this top to bottom. It should take a few minutes and end with a real email in their inbox.

---

0. What this tool is, so you explain it correctly

A fixed pipeline, not an autonomous agent:

gather sources -> render prompt -> generate -> gate -> render theme -> deliver -> post

One JSON job spec plus one markdown prompt equals one email. The model writes only the words; the engine controls everything else, so the same job produces the same shaped email every run.

Do not describe it as an AI agent that manages their inbox. It does not read incoming mail, decide things, or take actions.

---

1. Ask these questions first

Do not skip this. Defaults produce a generic email nobody opens. Ask, then build to the answers. Ask them together, in one message, and keep it short.

1. What should the email contain? Push for specifics. "AI news" is weak; "new Hugging Face models for video generation, plus arXiv papers mentioning agents" is a good job. If they are unsure, offer the shipped jobs as starting points (see step 4).
2. When should it arrive? A time and days. Convert to cron.
3. Where should it send? Gmail, or another provider over SMTP. Get the recipient address.
4. Which model? If they use Claude Code already, claude_cli:sonnet needs no API key and is the easiest. Otherwise an Anthropic or OpenAI key, or a local model.
5. Should it read anything on their machine? Notes, a repo, a folder, the output of a command. This is the feature no cloud scheduler has, so raise it explicitly. Many people do not realise it is possible.