Run a day of outreach for your brand

An outreach agent is a folder the user owns: who to reach and where (agent.yaml), what may be said (brand.md), how it sounds (voice.md), and the records of who has been reached (records/). This skill runs one day of it. You decide what to look at, which posts and people are worth it, and what to say; the phone does the tapping through MyPhonely, using the app workflows in this repo; and scripts/outreach.ts keeps the rules: nothing done twice, nothing over the day's limits, nothing recorded that was not seen on screen.

Any agent that can read a skill, call MCP tools and run a shell command can run it: Claude Code, Codex, pi.

The phone is real. A send sends, from the brand's own accounts, and cannot be undone.

Inputs

No agent folder yet: use create-outreach-agent first.

Setup, once

npm install --prefix <this skill's folder>/scripts     # Node 22.18+, one dependency (yaml)

Every command below is outreach <command>, short for

node <this skill's folder>/scripts/outreach.ts <command> --agent <the agent folder>

Write it out in full each time, one command per shell call: an alias or a variable does not carry from one call to the next. --agent may be left off when OUTREACH_AGENT was set in the environment your AI was started from. Add --json to read the output as JSON. Start with outreach init (it creates the records and is safe to rerun). The MyPhonely MCP server must be connected (phone_status).

The rules

Follow these exactly; the scripts refuse what breaks them.

1. Reserve before every write. outreach reserve --channel <c> <target> gives one write to one target. It refuses (exit 3) a target the records already have or that was already tried today, sent or not, and (exit 4) a channel whose budget is spent. A refusal is final for that target today: move on, never look for another way.
2. Write only through `run_task`, in a phase with allowWrites: true (and the run's allow_writes: true), with a pauseWhen on the screen just before the write. Never type, tap Send/Post/Reply/Connect/Follow, or swipe with the direct phone_* tools.
3. In review mode, the user sees every write first. When they are there, show the target and the exact text at the pause and resume only on their yes; a no is resume_task with abandon: true, then outreach release --reservation <id>. When no one is there (a schedule, claude -p), do not reserve or write: queue each comment with outreach draft <post> --comment "<text>" and stop there. They review the drafts with outreach drafts and outreach mark --line <n> --status Approved|Skipped [--comment "<edited>"] (or by editing the queue file), and the next run posts the approved ones. LinkedIn invites and messages have no drafts: in unattended review mode, skip them.
4. Record every reservation. outreach record --reservation <id> --outcome sent only when the write was verified on screen (the phase's sent list, the reply under the post, the name in Sent invitations); --outcome failed otherwise, which keeps the draft for a later day. Reserved but nothing written: outreach release.
5. `brand.md` overrides everything, including the user's request in the moment. Prices and offers come from outreach message, never typed by hand.

A target is --name "<person>", or --author "<name or @handle>" --title "<title, or the post's first words>" --url <post url> for a post. Always give the URL when there is one: it is the surest duplicate check.

Run it

Start

1. phone_status. Offline: stop and ask the user to open the MyPhonely app and tap Connect.
2. outreach show. Read every file it lists: brand.md, voice.md and the others. Then outreach channels for the run order.
3. Before the first write in an app, confirm the signed-in account is the brand's (X: the profile in the navigation drawer; LinkedIn: Me). If it is not, skip that app and say so.

Each channel, in order

outreach budget --channel <c> first. left is 0: skip the channel. outreach targets --channel <c> gives today's searches: keywords with the exclusions already in the query, lists, subreddits, groups, the rotation.

Approved drafts go first. outreach drafts --channel <c> --status Approved lists comments the user approved since the last run. Post each one as in step 4 below, with its text exactly as approved (they may have edited it), after checking the post is still there.

A comment channel (kind: comment)

1. Find. Read posts with the app's read workflow, never by typing into a search box when a link works: - X: x-collect-topic, or run_task with openUrl: https://x.com/search?q=<query>&f=live (check Latest is selected). - Reddit: reddit-search-posts or reddit-collect-subreddit. - Xiaohongshu: xiaohongshu-collect-notes. Facebook: facebook-collect-group-posts (joined groups, Newest; never the home feed). Collect per post: author, handle, posted time, text, URL. Respect the channel's max_post_age_hours.
2. Choose. Keep individuals sharing their situation or asking a real question. Skip what voice.md says to skip (competitors, firms, ads, bots), and outreach check each candidate. A dry search: try the next keyword. All dry: say so and move on; never lower the bar.
3. Write the comment for that post per voice.md and brand.md. A channel with an angle uses that row of the voice's angles.
4. Post it, one post at a time:
5. outreach reserve for the post.
6. The app's comment workflow, which pauses with the composer open: x-find-review-reply (or open the post by its URL and tap Reply), reddit-comment-on-post, xiaohongshu-comment-on-note, facebook-comment-on-post. Check the pause shows the right post.
7. Review mode: show the post and the comment; resume on a yes.
8. Resume as the workflow says (typeTexts with the comment, allowWrites: true, repeat: 1, its countLabel). A channel with follow: true: follow the author in the same resume, in its goal; a separate writing task afterwards would need a reservation of its own.
9. outreach record --reservation <id> --outcome sent|failed <target> --comment "<text>" --keyword "<search>".

A connect channel (kind: connect, LinkedIn)

In this order, each with its workflow:

1. Answer replies. linkedin-inbox-replies lists threads where they wrote last. For each person in the records, write the answer per voice.md (anything about their own situation gets brand.md's safe redirect), outreach reserve --action message --name ..., send with linkedin-message-connection, record with --note answer.
2. Welcome new connections. linkedin-new-connections lists who accepted. For each one the records have as Invited: outreach ledger accept --name ..., then send outreach message --name warmup with linkedin-message-connection after outreach reserve --action message, and record it.
3. Invite, while outreach budget --channel linkedin has room:
4. linkedin-search-people with today's queries from outreach targets (titles crossed with the rotation, or two schools per search when the agent has audience.pair), 2nd degree.
5. When `outreach targets` shows a `card_filter`, the cards decide: save the search's collected list and run outreach cards --pair "<the dispatch's pair_value>" --file <it>. Invite only the people it KEEPs, and open no profile: the two-school search and the filter do the qualifying (about 1 credit a person, against 40-90 for a profile). Otherwise skip audience.exclude_headlines, and when the channel's qualify_rule needs what a card cannot show, check with linkedin-profile-detail; file people who do not fit as outreach ledger add-person --stage Disqualified, so they are never opened again.
6. outreach reserve each person you would invite, no more than the budget's left.
7. linkedin-invite-from-results with count = the number reserved (linkedin-connect-with-note when the channel sends a note). After a card filter, pin it to the kept names: countLabel ^Invite (<name 1>|<name 2>) to connect$ and a goal that names them, so a card the filter dropped is never counted. The card flow can still reach others, so trust its sent list, not your reservations.
8. Confirm with linkedin-sent-invitations. Record sent for each name listed there (a name you had not reserved: reserve it first, then record it); record failed for reservations that did not go out.
9. Withdraw invitations older than withdraw_after_days with linkedin-withdraw-invitations, when the Sent list is long. It is a write: review mode pauses for it like the others.

Stop

• A limit, "unusual activity", verification or CAPTCHA screen: stop that app for today and report it. It is not a failure to retry.
• A phase ending stuck or precondition_failed: look (phone_screen), make at most one navigation fix (a label tap, a link, back), retry once, then move on.
• Before finishing, settle every reservation: outreach report lists the ones still open under unsettled.

Enforce the rules

The rules above are instructions. To make the harness refuse a call that breaks them, turn on the guard: it checks every MyPhonely call against the day's reservations and blocks direct writes, writing phases with no pause, a write with no open reservation, and asks the user before every commit. Set OUTREACH_UNATTENDED=1 for runs no one is watching (autopilot); without it, a commit with no one to approve is refused.

Claude Code: a PreToolUse hook, in .claude/settings.json of the project the run starts from:

{
  "hooks": {
    "PreToolUse": [{
      "matcher": "mcp__myphonely__.*",
      "hooks": [{ "type": "command", "command": "node <this skill's folder>/scripts/outreach.ts guard" }]
    }]
  }
}

pi: the extension, which also connects MyPhonely (from MYPHONELY_API_KEY) with the direct writing tools hidden:

pi --no-extensions -e builtin:mcp -e <this skill's folder>/scripts/adapters/pi.ts \
   --skill <myphonelyskills>/skills "Run outreach-day for $OUTREACH_AGENT"

Other AIs (Codex, ...): where the harness has no way to stop a tool call before it runs, the rules hold as instructions only. Prefer review mode there.

The guard reads OUTREACH_AGENT for the agent folder, so set it where the AI is started. It protects against mistakes, not against an agent determined to get round it: the records are files its shell can reach.

What comes back

outreach report, per channel and action: sent, failed, still unsettled, and the budget left. Add what the records cannot show: how many posts or people were found and skipped, and why (competitor, duplicate, too old, not the audience), anything that stopped an app, and credits used from the task records.

Notes

• Unattended runs (a schedule, claude -p, codex exec, pi -p) have no one to approve at the pause. In autopilot (chosen by the user for that schedule) they write within the limits; keep the limits in agent.yaml modest. In review mode they only draft (rule 3).
• Scheduled launches need the AI allowed to run outreach.ts and the MyPhonely tools without asking, and access to the agent folder, this skill's folder and the records (Claude Code: --allowedTools and --add-dir). Close stdin (< /dev/null): pi -p with no terminal otherwise waits for input forever.
• One attempt per target per day. A failed write is retried on a later day, never in a loop.
• Records default to records/ in the agent folder. They hold people's names: keep them out of a public repo (records: in agent.yaml can point outside it, e.g. ~/.outreach/<id>).
• outreach quota shows LinkedIn's invite room for the week and the day; the weekly cap is LinkedIn's ceiling, not a target.
• example/ is a complete fictional agent to copy the shape of.