X (Twitter) — DM Auto Chat (End-to-End)

Full X DM automation Skill: inbox scan → conversation read → persona-based reply → send; also supports search-and-outreach. The calling Agent generates reply text based on persona; this Skill handles all mechanical operations.

Language

All process output to user (progress updates, process notifications) follows the user's language.

Objective

Encapsulate "refresh DM list → identify pending replies → read context → reply with persona → send" and "search user → enter chat → send first message" into callable end-to-end capabilities.

Prerequisites

• Browser is open at X site, logged into X account ([aria-label="Account menu"] present)
• The 4-digit DM passcode for the current account is available (required for E2E encryption)
• Caller has prepared a "persona description" (used to generate replies), e.g.:
  • "You are BrowserAct outreach team. Tone: friendly, concise, professional. Goal: invite creators to collaborate."
• Optional: list of target user search queries (for outreach scenario)

Pre-execution Checks

1. Tool Readiness

If browser-act has been confirmed available in the current session → skip.

Invoke browser-act via Skill tool to load usage. If installation or configuration issues arise, follow its guidance to resolve then retry.

2. Open DM Entry + Comprehensive State Check

browser-act --session <name> navigate https://x.com/i/chat
browser-act --session <name> wait stable --timeout 15000
browser-act --session <name> eval "$(python scripts/check-page-state.py)"

Return format:

{
  "url": "https://x.com/i/chat/pin/recovery?from=%2Fi%2Fchat",
  "logged_in": true,
  "need_passcode": true,
  "on_inbox": false,
  "on_conversation": false,
  "has_panel": false,
  "has_composer": false,
  "inbox_count": 0
}

Decision matrix:

• logged_in: false → inform user to log in first; wait; retry this step
• need_passcode: true → proceed to step 3 below
• on_inbox: true and inbox_count > 0 → ready, enter business flow
• on_inbox: true but inbox_count === 0 → account has no DM conversations; outreach scenario can still proceed, pending-reply scenario has nothing to do

3. DM Passcode Unlock (when need_passcode is true)

1. If caller has provided passcode in advance → use it directly; otherwise ask user for 4-digit DM passcode via AskUserQuestion tool (do not use plain text prompt — must call AskUserQuestion)
2. browser-act --session <name> state — find indexes of 4 <input maxlength=1 pattern=[0-9]*> elements (usually 4 consecutive)
3. Enter each digit: browser-act --session <name> input <idx1> "<d1>", <idx2> "<d2>", <idx3> "<d3>", <idx4> "<d4>"
   • Must use browser-act input (CDP real keyboard events), cannot use eval to set value — X ignores non-real keyboard input
4. browser-act --session <name> wait stable --timeout 10000
5. Re-run check-page-state.py, confirm need_passcode: false and on_inbox: true
6. 3 consecutive failures still showing need_passcode: true → inform user passcode may be wrong; terminate

Business Flows

Choose Scenario A, Scenario B, or both. Each scenario is an ordered AI Workflow (not a single JS).

Scenario A: Scan unread DMs → Persona-based reply

Flow: Scan inbox → Filter unread & latest peer messages → Per-conversation: read context → Generate reply with persona → Send → Next

Steps:

1. Scan inbox:

   browser-act --session <name> eval "$(python scripts/scan-inbox-merged.py)"
   

   Returns items[], each containing conversation_id / conversation_url / peer_screen_name / peer_display_name / peer_can_dm / latest_message_preview / latest_message_from_self / unread, etc.

2. Filter pending-reply conversations: from items, select conversations meeting all conditions:

   • unread === true (has unread) or latest_message_from_self === false (peer's latest message not yet replied)
   • peer_can_dm === true (recipient allows DM)
   • is_muted !== true and is_deleted_by_viewer !== true
   • Optional caller filters: only reply to specific screen_names, exclude already-replied (use external JSONL ledger)

3. For each pending-reply conversation (strictly serial, random sleep 8-15 seconds between each):

   a. Open conversation:

   browser-act --session <name> navigate https://x.com<conversation_url>
   browser-act --session <name> wait stable --timeout 15000
   

   b. If passcode re-triggered → re-unlock (usually won't re-trigger within same session)

   c. Read context:

   browser-act --session <name> eval "$(python scripts/read-conversation.py)"
   

   Returns messages[], each with direction (self/peer), text, timestamp_text, links, images.

   d. (Optional) Load full history: If caller needs longer context, loop:

   browser-act --session <name> eval "$(python scripts/scroll-load-history.py)"
   

   Until reached_top: true, then re-read with read-conversation.py.

   e. Generate reply: Calling Agent combines persona, message history to generate reply text. Reply content is entirely the caller's decision; this Skill does not participate in generation. Suggested inputs:

   • Persona prompt (provided by caller)
   • Recent N messages (typically messages.slice(-6))
   • Peer name (peer_display_name / peer_screen_name) for address
   • Return one string reply_text, length < 10,000 characters

   f. Send reply:

   1. browser-act --session <name> eval "$(python scripts/check-composer.py)" → record last_message_id
   2. browser-act --session <name> state — find <textarea placeholder=Message> index TA_IDX
   3. browser-act --session <name> input <TA_IDX> "<reply_text>" (must use CDP real keyboard, cannot use eval)
   4. browser-act --session <name> wait --selector '[data-testid="dm-composer-send-button"]' --state attached --timeout 5000
   5. browser-act --session <name> eval "document.querySelector('[data-testid=\"dm-composer-send-button\"]').click(); 'clicked'"
   6. browser-act --session <name> wait stable --timeout 15000
   7. Verify: browser-act --session <name> eval "$(python scripts/verify-sent.py '<reply_text>' --prev-last-id <last_message_id from step f1>)"
      • sent: true and composer_cleared: true → success, record result
      • sent: false → record failure, do not retry (prevents duplicate sends); proceed to next conversation

   g. Random delay: sleep 8-15 seconds (avoid anti-abuse limits)

4. Batch completion: Summarize results (success count / failure count / conversation_id per item); return or write to external log file.

Scenario B: Search users → Start new conversation → Send first message

Flow: Search candidates → Filter sendable → Enter conversation → Generate first message → Send

Steps:

1. Search target users (one search per target, 1-2 second interval between searches):

   browser-act --session <name> eval "$(python scripts/search-users.py '<search_query>')"
   

   Returns users[], each with user_id / name / screen_name / can_dm / can_dm_reason / verification fields.

2. Filter users who can receive DMs:

   • can_dm === true and !suspended and !protected
   • can_dm_reason === "Allowed"
   • If screen_name is already in send history → skip (deduplication)

3. For each target user (strictly serial, sleep 10-20 seconds between each):

   a. Calculate conversation URL:

   browser-act --session <name> eval "$(python scripts/open-conversation-by-user.py '<user_id>')"
   

   Returns conversation_url (e.g., /i/chat/{smaller_id}-{larger_id}).

   b. Navigate to conversation:

   browser-act --session <name> navigate https://x.com<conversation_url>
   browser-act --session <name> wait stable --timeout 15000
   

   c. Handle passcode (may appear on first DM entry) → unlock

   d. Verify composer ready:

   browser-act --session <name> eval "$(python scripts/check-composer.py)"
   

   composer_ready: true → record last_message_id; false → skip this user

   e. Generate first message: Calling Agent generates first outreach text first_text based on persona + target user info (screen_name / name / verification type). Suggested content:

   • Brief self-introduction (caller identity)
   • Personalized reason for reaching out to this specific user
   • Clear call-to-action
   • Keep length < 500 characters (first messages that are too long are more likely to be flagged as spam)

   f. Send: Follow the 7 sub-steps in "Scenario A step 3f", substituting first_text for reply_text.

   g. Random delay: sleep 10-20 seconds

4. Batch completion: Summarize results.

Capability Components (callable individually)

In addition to the Scenario A / B end-to-end flows, the following components can also be called directly:

Composite: Inbox scan (API + DOM merged)

browser-act --session <name> eval "$(python scripts/scan-inbox-merged.py)" Returns merged conversation list with peer screen_name + message preview + unread flag.

API: Fetch inbox from API only (with pagination)

browser-act --session <name> eval "$(python scripts/fetch-inbox-api.py --cursor-id {cursor_id} --graph-snapshot-id {snap} --limit {N})"

DOM: Read current conversation messages

browser-act --session <name> eval "$(python scripts/read-conversation.py)"

DOM: Scroll to load message history

browser-act --session <name> eval "$(python scripts/scroll-load-history.py)"

DOM: Check composer state

browser-act --session <name> eval "$(python scripts/check-composer.py)"

DOM: Verify message was sent

browser-act --session <name> eval "$(python scripts/verify-sent.py '<expected_text>' --prev-last-id <last_id>)"

API: Search X users (with DM permission)

browser-act --session <name> eval "$(python scripts/search-users.py '<query>')"

JS: Calculate conversation URL from user_id

browser-act --session <name> eval "$(python scripts/open-conversation-by-user.py '<user_id>')"

JS: Comprehensive page state check

browser-act --session <name> eval "$(python scripts/check-page-state.py)"

Success Criteria

End-to-end Scenario A:

• sent: true rate >= 90% for each pending-reply conversation
• Failed conversations have clear reason recorded (wrong passcode, composer unavailable, 429, etc.)

End-to-end Scenario B:

• All filtered sendable users enter conversation page (composer_ready: true)
• First message sent: true rate >= 90%

Atomic components: see success criteria in each atomic Skill (scripts in this directory fully reuse the atomic implementations).

Known Limitations

X Platform DM Limits (verified through exploration)

• E2E encryption passcode required: Must enter 4-digit passcode to unlock DMs; wrong or disconnected passcode loses message history. Passcode input only works via browser-act input (CDP real keyboard); eval setting value does not work
• Message bodies are E2E encrypted: GraphQL API response message events are base64 T-protocol encrypted binary; plaintext is only readable from the browser's already-unlocked DOM. This Skill must run in an already-logged-in and unlocked browser
• Peer DM permissions (can_dm_reason enum, observed values): Allowed — can send; InboxClosed — recipient closed DM; other values (possibly Blocked, NotFollowing, etc.) treat as cannot send
• Non-follower DMs go to Message Requests: First message to a user who doesn't follow you goes to their Message Requests; they must accept before it moves to Primary
• Send rate (anti-abuse, no official docs): Empirical max ~5-10 messages per minute; 8-15 second random delay between messages; exceeding threshold triggers HTTP 429 or UI block
• Message length cap: 10,000 characters per message (X official limit)
• Timestamp precision: DOM only gives X display format ("30m" / "6:25 PM" / "May 8"); no ISO datetime
• Attachment messages not covered: Sending images / GIFs / voice / video / quote tweets not implemented; this Skill handles plain text only

Additional Skill Limitations

• Does not participate in reply content generation: Reply text generation (persona application, context understanding, personalization) is entirely the calling Agent's responsibility; this Skill is the operation layer
• Does not maintain cross-session state: Per-run reply history, blocklists, and progress need the caller to record in external files (JSONL)
• Group conversations: peer_* fields take only the first non-self member; fine-grained replies in group conversations are not supported
• Message Requests sub-inbox: Currently only scans Primary inbox; Message Requests are not read; scanning Message Requests requires navigating to a different page — not implemented in this version

Execution Efficiency

• Batch processing: One run processes one batch (N conversations or N target users) then returns; no long-running resident loop — let the caller decide the scheduling cadence
• Strictly serial: All DM operations for the same account must be serial — no parallel; parallel operations accelerate anti-abuse triggering
• No retry on failure: DM send failures are usually permission / rate / network issues; retrying risks duplicate sends — record uniformly and skip
• Resume from breakpoint: Batch tasks use JSONL to record {target, status, timestamp, error?} per item; resume from breakpoint on interruption
• Small-scale validation first: Before bulk runs, validate the full pipeline with 1-2 items, then scale to full batch
• Reuse browser session: Use the same browser-act session (e.g., --session x-dm) for the whole batch; passcode unlock and login state persist within the session, no need to re-unlock for each item

Experience Notes

Path: {working-directory}/browser-act-skill-forge-memories/x-dm-automation-x-dm-auto-chat.memory.md (working directory is determined by the Agent running the Skill)

Before execution: If the file exists, read it first — it records unexpected situations encountered during past executions (e.g., a strategy has become ineffective, a selector changed, a rate threshold discovered); adjust strategy order accordingly.

After execution: If an unexpected situation is encountered (strategy became ineffective, page redesigned, anti-scraping upgraded, better path discovered, new can_dm_reason enum values), append a line: {YYYY-MM-DD}: {what happened} → {conclusion}

Normal execution does not write to the file. Do not record what keywords were used, which conversations were replied to, or how many messages were sent — those are task outputs, not experience.