open_my_profile
no demo yetauto-login runner: beta
Try asking
- Open my TikTok profile
- open my profile
- go to my TikTok profile page
- show me my profile
ghost skill run ghost/tiktok open_my_profile TikTok — user's-own-account actions: nav, account switching, self-stats, self-uploads. ToS-safe surface only.
High terms-of-service risk. The vendor's terms prohibit automated access or bot-created accounts. Published for research: do not run it against live accounts. Not vendor-approved.
8actions not benchmarked yet
No clean demo yet. Recordings will start on HOME and show the app launching.
Ghost in the Droid skill: TikTok automation for your own account. ToS-safe surface only.
8 actions, all deterministic hard-skill:
| Action | What it does |
|---|---|
open_my_profile | Navigate to your profile tab |
open_inbox | Navigate to your inbox |
get_active_account | Return current active @handle |
get_unread_count | Read the unread-notifications badge count |
get_my_stats | Scrape your handle + display-name + followers/following/likes |
list_accounts | Enumerate all logged-in accounts on this device |
switch_account | Switch active account to a given @handle |
upload_video | Upload a video via share intent: caption, hashtags, optional TTS, draft or post |
pip install ghost-in-the-droid>=1.6.0 # 1.6.0 is not on PyPI yet: until it ships, git clone the repo and `pip install -e .`
ghost skill install ghost/tiktok
# Simple: read your own stats
ghost "how many followers do I have on tiktok" --device asus
# Switch accounts
ghost "switch to my my_other_account tiktok" --device asus
# Upload a draft
ghost "upload cat.mp4 to tiktok as a draft with caption 'look at this' and #cat #fyp" --device asus
Or programmatically via MCP:
run_skill(name="ghost/tiktok", action="get_my_stats")
# → {"stats": {"handle": "yourhandle", "following": 62, "followers": 10, ...}}
run_skill(name="ghost/tiktok", action="switch_account", params={"handle": "my_other_account"})
# → {"switched_to": "my_other_account"}
run_skill(name="ghost/tiktok", action="upload_video", params={
"video_path": "/sdcard/DCIM/cat.mp4",
"caption": "look at this",
"hashtags": ["cat", "fyp"],
"action_mode": "draft",
})
com.zhiliaoapp.musically ≥44.0.0Fingerprint mismatch: the engine falls back to Tier 2 Adapt (LLM re-mines the flow using SOFT.md as reference).
files (video upload)notifications (unread badge read)skill.yaml declares 10 popup detectors auto-dismissed before every step:
FB/Contacts access, notifications enable, TikTok Shop relevance, draft resume overlay, rate-app nag, security checkup, invisible promo overlays, etc.
Unknown popups: the engine takes a screenshot and raises UnknownPopupError so the catalog can be extended.
docs/skill-schema-v2.mddocs/skill-taxonomy-soft-vs-hard.mddocs/contributing-skills.mdSOFT.md| App version | Android | Device | Evidence |
|---|---|---|---|
| 44.3.3 | 12 (run artifact fingerprint) | asus/ASUS_I002D (ROG, physical, 1080x2400) | 1 signed self-run in runs/ (get_my_stats, pass, 2026-10-05) |
| 44.3.3 | as above | same ROG | maintainer live runs, not signed artifacts: 7/8 actions green, later 8/8 with upload_video 10/10 steps (2026-10-05) |
Untested: emulators (TikTok is not installable there without a Play account), other OEMs, other screen sizes, locales other than en-US.
None in the signed self-runs. Known from maintainer runs: upload_video once stopped at 9/10 steps because the hashtag autocomplete list covered the Drafts button.
FB friends and contacts access, Facebook access, TikTok Shop relevance, shared collections, notifications, security checkup, rate-the-app, draft resume overlay, avatar upsell (BACK), precise-location upgrade. Storage permission is allowed (os_popup_policy) for uploads.
uiautomator dump fails there. Navigation to Profile uses a coordinate tap on the bottom bar (y 2214 on 1080x2400) and only reads the tree on the static profile page; other screen heights need a new coordinate.dumpsys reports SplashActivity for 30 s or more after the feed is visible; wait on the Profile content-desc instead.helpers.py falls back to coordinates.UnknownPopupError with a screenshot so the catalog can grow.SOFT.md is the narrative an agent reads when the exact steps no longer match your app version and it has to adapt the skill.
This document describes the TikTok skill at a semantic level — what each action does, what it depends on, and how to recognize success. LLMs use this to adapt the skill to a device or app-version the shipped hard-skill doesn't match.
Package: com.zhiliaoapp.musically
Verified: 2026-08-30 on ASUS ROG (Zenfone 10-class, Android 15, 1080x2400), TT 44.3.3
Skill class: HARD (all 8 actions are deterministic; no HITL checkpoints)
Scope: user's-own-account, ToS-safe surface only. No mass-outreach, mass-scrape, engagement automation.
TT bottom-nav has 5 tabs, left → right:
content-desc="Home")content-desc="Friends")content-desc="Inbox")content-desc="Profile")Tabs are clickable=true and always keyed by content-desc:. Empirical y-center on 2400h screens: ~2214. X-centers: Home=108, Friends=~324, Inbox=756, Profile=972.
open_my_profile() — LOW complexitycom.zhiliaoapp.musicallycontent-desc="Profile" (bottom-nav landmark; use this rather than dumpsys — dumpsys reports SplashActivity for 30s+ even after the home feed is visible)content-desc="Profile"content-desc="Profile menu" appears (the hamburger in top-right of profile page)open_inbox() — LOW complexitySame shape as open_my_profile but tap content-desc="Inbox". No stable post-tap landmark on all TT variants — some land on a notification list, some on a chat list. Accept success on presence of the Inbox tab confirming we made the tap.
get_active_account() — MEDIUM complexityopen_my_profile()@<handle> — the FIRST match is the active handle (profile header renders it under the display name)Handle pattern: @[a-zA-Z0-9_.][a-zA-Z0-9_.]{1,29} — 2 to 30 chars, letters/digits/./_.
get_unread_count() — MEDIUM complexitycontent-desc="Inbox, 4 unread" — regex-extract the digit"4" at bounds roughly (648-864, 2100-2220) — bottom-right of the Inbox tab areaget_my_stats() — MEDIUM complexityopen_my_profile()@xxxtext="my god" style, has chevron ▼ for account switcher)Following, Followers, Likes. Parse 1.2K → 1200, 3M → 3_000_000.text="Drafts: N" in the drafts-tab previewlist_accounts() — MEDIUM complexityThe account-switcher opens by tapping the profile-header display-name area (the text="my god" clickable, not the handle). No stable id — use bounds y ∈ (380, 480) and x-center 540 as fallback.
open_my_profile()text="Switch account" to appear (drawer-open landmark)selected="true" marks the currently-active account.Empirical row spacing: ~176px per row. Row-y = 1086 (first) + 176 × N.
switch_account(handle) — MEDIUM complexityopen_my_profile()handle in the drawer rowsselected=true → no-op success (close drawer)@{handle} now appears as active handleupload_video(video_path, caption, hashtags, inject_tts, action_mode) — HIGH complexityThe full flow — see helpers.py:prepare_video_with_tts / push_and_scan_video and skill.yaml:upload_video for the deterministic sequence. Key semantic beats:
inject_tts, overlay TTS voiceover on the video BEFORE the ADB flow starts/sdcard/DCIM/upload.mp4, trigger media scan, capture _id from MediaStoream start -a android.intent.action.SEND -p com.zhiliaoapp.musically -t video/mp4 --es android.intent.extra.STREAM content://media/external/video/media/{media_id} — this jumps directly to TT's post-page with the video pre-loaded#tag1 #tag2 ...)The upload flow is the "hardest hard-skill" reference example. It exercises: launch_intent, pre_hooks, wait_for_text, if_visible conditional branching, retry wrapper on flaky sub-taps, multiple python_action: targets, and composite any_of/all_of validators.
TT throws many popups mid-flow:
The skill's popup_detectors: array in skill.yaml catches all these — engine auto-dismisses before every step. For popups NOT in the catalog: engine's fallback (BACK → screenshot → UnknownPopupError) captures them for you to add.
If your fingerprint doesn't match (hardware != * in skill.yaml, or app_version < 44.0.0), the engine falls back to Tier 2 Adapt:
helpers.pyKey things LLM should re-verify on adapt:
Open your signed-in profile with the Profile menu visible before running an account action. The read-only logged_in probe checks that positive selector on the current screen. A different screen, language or UI version can return missing/unknown even while a session exists; the runner then stops with needs_state rather than guessing. This skill has no automatic login provider. auto-login runner: beta.
Try a sample prompt or run an action by name. Replace any parameters in braces with your own values.
open_my_profileauto-login runner: beta
Try asking
ghost skill run ghost/tiktok open_my_profile open_inboxauto-login runner: beta
Try asking
ghost skill run ghost/tiktok open_inbox get_active_accountauto-login runner: beta
Try asking
ghost skill run ghost/tiktok get_active_account get_unread_countauto-login runner: beta
Try asking
ghost skill run ghost/tiktok get_unread_count get_my_statsauto-login runner: beta
Try asking
ghost skill run ghost/tiktok get_my_stats list_accountsauto-login runner: beta
Try asking
ghost skill run ghost/tiktok list_accounts switch_accountauto-login runner: beta
Try asking
ghost skill run ghost/tiktok switch_account upload_videoauto-login runner: beta
Try asking
ghost skill run ghost/tiktok upload_video Each release is an immutable git tag. The tree hash pins the exact skill directory the installer verifies.
| Version | Released | App versions | Tree | Size | Provenance |
|---|---|---|---|---|---|
2.0.1latest | never | >=44.0.0 | pending | 60.9 KB | untagged |