X — Keyword Comment

keyword + reply intent → search X tweets → read tweet content → generate contextual replies → post to reply area

Language

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

Objective

Search X by keyword, read each tweet's content, generate contextual replies based on a configured brand persona, and post them — all within a browser-act session.

Prerequisites

• config/keyword-comment-config.json has been filled in with actual product, persona, and tone values (all YOUR_* placeholders replaced before first run)

Session Rule

{SESSION} is a temporary, per-run session name used in all browser-act --session {SESSION} commands below. It is generated at execution start (e.g., xkc-{timestamp}) and not persisted across runs.

Pre-execution Checks

1. Tool Readiness

If browser-act has been confirmed available in the current conversation → skip this step.

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

2. Load Config

python -c "
import json, pathlib, sys
for base in ['.claude/skills/x-keyword-comment', 'output/x-keyword-comment']:
    cfg = pathlib.Path(base) / 'config/keyword-comment-config.json'
    if cfg.exists():
        print(json.dumps(json.loads(cfg.read_text(encoding='utf-8')), ensure_ascii=False, indent=2))
        sys.exit(0)
print('ERROR: config/keyword-comment-config.json not found', file=sys.stderr)
sys.exit(1)
"

Hold product.*, persona.*, tone.* fields in working memory for reply composition.

3. Browser Selection

List available browsers:

browser-act browser list

• If browsers exist → present the list to the user and let them choose which browser to use for this X session.
• If no browsers exist → guide the user to create one (e.g., browser-act browser create --type stealth --headed), then repeat the list step.

Once the user selects a browser, record its ID as {BROWSER_ID} for this run.

4. Open Session

Generate a unique session name (e.g., xkc-{timestamp}) as {SESSION}. Open the browser:

browser-act --session {SESSION} browser open {BROWSER_ID} https://x.com/ --headed

If the browser is already open with an active session, list sessions and reuse:

browser-act session list

Pick the session associated with {BROWSER_ID} and assign its name to {SESSION}.

5. Login Verification

If X login status has been confirmed in the current conversation → skip this step.

Otherwise: browser-act --session {SESSION} get markdown and check:

• Sidebar bottom shows @username, top navigation shows Home / Explore → logged in, continue
• Page shows a "Sign in" button with no logout entry → not logged in; inform the user that login is required and assist the login flow

User refuses or cannot log in → terminate execution.

Capability Components

This Skill's operational boundary = what the user can manually do in their browser. It only reads data already displayed to the logged-in user, never bypassing authentication or access controls. JS code is encapsulated in Python files under scripts/, invoked via browser-act --session {SESSION} eval "$(python scripts/xxx.py {params})". $(...) is bash syntax; use the bash tool for execution.

Below are all atomic capabilities discovered and verified during the exploration phase, listed by command template with parameters. Simply invoke them as needed — no need to read scripts/*.py source code or re-verify. Only inspect scripts when execution fails for troubleshooting. Combine freely as needed during execution.

AI Workflow: Pre-reply Warmup

Warm up the account before posting replies to simulate organic browsing behavior.

Skip condition: warmup already performed today and less than 4 hours ago, or user says "fast mode".

Step 1 — Check notifications and messages (2–3 min)

browser-act --session {SESSION} navigate "https://x.com/notifications"
browser-act --session {SESSION} wait stable
browser-act --session {SESSION} get markdown
sleep $((RANDOM % 31 + 60))   # 60–90 s
browser-act --session {SESSION} navigate "https://x.com/messages"
browser-act --session {SESSION} wait stable
browser-act --session {SESSION} get markdown
sleep $((RANDOM % 31 + 30))   # 30–60 s

Step 2 — Browse feed and like (3–5 min)

browser-act --session {SESSION} navigate "https://x.com/home"
browser-act --session {SESSION} wait stable
browser-act --session {SESSION} get markdown

Randomly pick 3–5 tweets from the feed. For each:

browser-act --session {SESSION} navigate "{tweet URL}"
browser-act --session {SESSION} wait stable
sleep $((RANDOM % 26 + 15))   # 15–40 s
# If content is relevant → like it:
browser-act --session {SESSION} state
browser-act --session {SESSION} click {Heart index}   # element with aria-label containing "Like"
browser-act --session {SESSION} wait stable
sleep $((RANDOM % 8 + 8))     # 8–15 s
browser-act --session {SESSION} navigate "https://x.com/home"
sleep $((RANDOM % 16 + 10))   # 10–25 s

Target: like 1–3 tweets; daily cap 20–30 likes (avoid fast bulk likes that trigger rate limits).

Step 3 — Keyword search browsing (2–3 min)

browser-act --session {SESSION} navigate "https://x.com/search?q={KEYWORD_ENCODED}&f=live"
browser-act --session {SESSION} wait stable
browser-act --session {SESSION} get markdown

Open 2–3 results, spend 25–60 s each reading the full tweet (as reply material).

Pre-action pause

sleep $((RANDOM % 61 + 60))   # 60–120 s — simulate "browse first, then reply"

---

DOM: Scan Replyable Tweets on Current Page

After navigating to the X search results page, scan all tweets with their reply button indices and content.

1. Navigate: browser-act --session {SESSION} navigate "https://x.com/search?q={KEYWORD_ENCODED}&src=typed_query&f=live"
   • {KEYWORD_ENCODED} is URL-encoded (spaces as %20)
   • f=live returns newest tweets; omit for Top tweets
2. Wait: browser-act --session {SESSION} wait stable --timeout 30000
3. (Optional) Scroll to load more: browser-act --session {SESSION} scroll down --amount 1500 → browser-act --session {SESSION} wait stable --timeout 10000 → re-scan
4. Scan: browser-act --session {SESSION} eval "$(python scripts/scan-search-tweets.py --limit {N})"

Parameters:

• --limit: max tweets to return, default 10

Output example:

{
  "totalReplyBtns": 8,
  "tweets": [
    {
      "i": 0,
      "tweetSnippet": "Breaking: Alibaba just killed the browser automation stack...",
      "authorHandle": "@AIGuideHQ",
      "authorUrl": "https://x.com/AIGuideHQ",
      "tweetUrl": "https://x.com/AIGuideHQ/status/2051969984847286536",
      "replyBtnIdx": 0
    }
  ]
}

replyBtnIdx note: This is the reply button's position index among all [data-testid="reply"] buttons currently on the page. After posting a reply, the DOM partially updates (new reply inserts), shifting subsequent indices — re-run scan-search-tweets.py after each reply to get fresh indices before the next one.

DOM: Click Reply Button (Open Editor)

Click the reply button for a specific tweet to open the reply input box.

browser-act --session {SESSION} eval "$(python scripts/click-reply.py {replyBtnIdx})"

Parameters:

• {replyBtnIdx}: the tweet's reply button index (positional argument, from scan-search-tweets.py)

Output example (success):

{
  "ok": true,
  "replyBtnFound": true,
  "totalReplyBtns": 8
}

Output example (out of range):

{
  "ok": false,
  "reason": "reply_btn_out_of_range",
  "total": 8
}

DOM: Type Reply Text and Submit (Operation)

Architecture note: X uses the Draft.js editor (public-DraftEditor-content). document.execCommand('insertText') only updates the DOM without triggering React internal state — the submit button stays disabled. You must use browser-act's native input command to simulate real keyboard input to activate the submit button. This is the only reliable method.

After clicking the reply button (click-reply.py), complete text input and submission:

1. Wait for editor mount: browser-act --session {SESSION} wait --selector '[data-testid="tweetTextarea_0"]' --state attached --timeout 10000
2. Get editor index: browser-act --session {SESSION} state → find aria-label=Post text role=textbox → note {EDITOR_IDX}
3. Input reply text: browser-act --session {SESSION} input {EDITOR_IDX} '{reply_text}'
4. Get submit button index: browser-act --session {SESSION} state → find button labeled Reply → note {REPLY_BTN_IDX}
5. Submit: browser-act --session {SESSION} click {REPLY_BTN_IDX}
6. Wait: browser-act --session {SESSION} wait stable --timeout 10000

Success signal: browser-act --session {SESSION} network requests --filter CreateTweet --method POST --status 200 returns at least 1 record.

Closing the editor: If the editor is empty, pressing Escape dismisses it directly with no dialog. If text has been typed and Escape is pressed (or the modal is otherwise closed), X shows a "Save post?" confirmation dialog (Save / Discard). To discard: browser-act --session {SESSION} state → find Discard button index → browser-act --session {SESSION} click {DISCARD_IDX}

Composite: Full Keyword Reply Flow

All operations remain on the X search page — no navigation to individual tweet detail pages required.

Config: Load config/keyword-comment-config.json and hold product.*, persona.*, tone.* fields in working memory before proceeding:

python -c "
import json, pathlib, sys
for base in ['.claude/skills/x-keyword-comment', 'output/x-keyword-comment']:
    cfg = pathlib.Path(base) / 'config/keyword-comment-config.json'
    if cfg.exists():
        print(json.dumps(json.loads(cfg.read_text(encoding='utf-8')), ensure_ascii=False, indent=2))
        sys.exit(0)
print('ERROR: config/keyword-comment-config.json not found', file=sys.stderr)
sys.exit(1)
"

1. browser-act --session {SESSION} navigate "https://x.com/search?q={KEYWORD_ENCODED}&src=typed_query&f=live" → browser-act --session {SESSION} wait stable --timeout 30000
2. Initial scan: browser-act --session {SESSION} eval "$(python scripts/scan-search-tweets.py --limit {N})" → candidate tweet list
3. If not enough candidates → browser-act --session {SESSION} scroll down --amount 1500 → browser-act --session {SESSION} wait stable --timeout 10000 → re-scan, merge results
4. Filter candidates by authorUrl / tweetSnippet (skip promotional or low-relevance tweets)
5. For each target tweet:
   • a. Generate reply: Use intent (caller-provided) + tweetSnippet + authorHandle + loaded config (product.*, persona.*, tone.*) to compose a 60–180 character ASCII reply. See references/quality-checklist.md (7-item checklist) and references/reply-composition.md (3 recommendation scenarios A/B/C).
   • b. browser-act --session {SESSION} eval "$(python scripts/click-reply.py {replyBtnIdx})" → open reply box
   • c. browser-act --session {SESSION} wait --selector '[data-testid="tweetTextarea_0"]' --state attached --timeout 10000
   • d. browser-act --session {SESSION} state → get editor index {EDITOR_IDX} → browser-act --session {SESSION} input {EDITOR_IDX} '{reply_text}'
   • e. browser-act --session {SESSION} state → get Reply button index {REPLY_BTN_IDX} → browser-act --session {SESSION} click {REPLY_BTN_IDX}
   • f. browser-act --session {SESSION} wait stable --timeout 10000
   • g. Verify: browser-act --session {SESSION} network requests --filter CreateTweet --method POST --status 200
   • h. Random interval: sleep $((60 + RANDOM % 120)) (60–180 s between replies)
   • i. Re-scan after each reply: re-run scan-search-tweets.py to refresh replyBtnIdx values before the next reply

Output per tweet:

{
  "authorUrl": "https://x.com/AIGuideHQ",
  "tweetUrl": "https://x.com/AIGuideHQ/status/2051969984847286536",
  "tweetSnippet": "Breaking: Alibaba just killed the browser automation stack...",
  "replyText": "The captcha point is real -- Playwright + Cloudflare means more glue than logic...",
  "posted": true,
  "skippedReason": null
}

Pagination

DOM Pagination: Search results load as an infinite scroll. Trigger more: browser-act --session {SESSION} scroll down --amount 1500 → browser-act --session {SESSION} wait stable --timeout 10000 → re-scan. Termination: totalReplyBtns does not increase across 2 consecutive scrolls, or target reply count is reached.

Success Criteria

posted == true for each tweet, confirmed by browser-act --session {SESSION} network requests --filter CreateTweet --method POST --status 200 returning at least 1 record (editor disappearing after submission is a secondary signal only).

Known Limitations

• Windows non-ASCII encoding trap: browser-act input {idx} '{text}' on Windows cmd (GBK active codepage) will corrupt non-ASCII characters (em-dash, full-width quotes, emoji) passed as arguments. Scripts call sys.stdout.reconfigure(encoding='utf-8', newline='\n'). Callers must also ensure UTF-8 terminal: run chcp 65001 or set PYTHONUTF8=1, or restrict reply text to ASCII-only characters
• Draft.js editor rejects execCommand: document.execCommand('insertText') only updates the DOM without triggering React state — submit button stays disabled. Must use browser-act native input command
• replyBtnIdx is not stable: After each reply the DOM partially updates; the new reply may insert near the top, shifting all subsequent indices. Must re-scan before every reply
• Reply rate: X's CreateTweet API rate limit is 300/15min, but account-level risk controls are far stricter. Over 20 replies/hour risks rate limiting, verification prompts, or suspension. Recommended: 60–180 s between replies, max 50 replies/day per account
• Account weight: Accounts with no avatar, no bio, few followers (< 50), and no post history may have replies silently shadow-banned
• Duplicate content filter: Sending identical or similar replies in a short window is automatically intercepted
• "Save post?" dialog: Pressing Escape with text in the editor triggers a save confirmation; must click Discard to close
• Platform ToS: X's Terms of Service explicitly restrict automated behavior; accounts risk rate limiting, warnings, or permanent suspension

Execution Efficiency

• Batch orchestration: For small counts (< 3) invoke directly; for larger counts write a bash loop script. Do not parallelize — rate limits apply per account
• Test before batch: Run the full flow (scan → post reply → verify CreateTweet) for 1 tweet first; only run the full batch after confirming it works
• Re-scan after each reply: replyBtnIdx changes with DOM updates; must re-run scan-search-tweets.py after every reply
• Error resumption: Save result per tweet (posted status + tweetUrl + tweetSnippet hash) incrementally; on failure, resume from breakpoint
• Interval jitter: 60–180 s random interval between replies
• Stop on risk signals: Immediately stop on: identity verification prompt, reply buttons disappearing, "You've reached your reply limit" message, or any suspension warning

Experience Notes

Path: {working-directory}/browser-act-skill-forge-memories/x-keyword-comment-x-keyword-comment.memory.md (working directory is determined by the Agent running the Skill, typically the project root or current working directory)

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

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

Normal execution does not write to the file. Do not record what keywords were used or how many results were returned — those are task outputs, not experience.