Viralman setup

viralman-setup Skill

Unified credential setup for all viralman channels. One invocation = one channel. Run once per channel you want to activate.

Trigger phrases

• /viralman-setup
• "viralman 셋업", "viralman setup"
• "set up viralman", "viralman 설정 도와줘"
• "viralman credentials", "viralman 자격증명 설정"
• "gitmail 셋업", "set up gitmail" (if a channel name is part of the phrase, auto-route to that channel)
• "twitter for viralman", "set up twitter for viralman"
• "reddit for viralman", "set up reddit for viralman"
• "linkedin for viralman", "set up linkedin for viralman"

Korean:

• "viralman 자격증명 저장해줘", "API 키 저장", "OAuth 등록"
• "viralman 키 등록", "twitter 자격증명 저장", "reddit 인증"
• "viralman 깔아줘", "viralman 설치해줘", "viralman 부트스트랩", "viralman 기본 세팅"

English:

• "save my viralman credentials", "register viralman api key"
• "configure viralman auth", "set up viralman tokens"
• "install viralman", "set up viralman from scratch", "bootstrap viralman", "make viralman work"

Chinese:

• "保存 viralman 凭证", "配置 viralman 密钥"
• "安装 viralman", "把 viralman 装好"

Japanese:

• "viralman の認証情報を保存", "viralman の API キーを設定"
• "viralman をインストール", "viralman をセットアップ"

Boundaries (read before starting)

• Never accept secrets in chat without warning. Show the warning in Step 2 once; after the user confirms, proceed silently with --set.
• Never run read -s yourself — it must land in the user's terminal.
• Never read, echo, or cat ~/.viralman/.env. Only --show-keys is safe.
• Do not WebFetch any platform dashboard (logged-in surfaces).
• One channel per invocation. Do not touch other channels.
• No live sends or posts after setup — read-only verification only.

---

Step 0 — environment check (auto-bootstrap if needed)

Before touching credentials, verify viralman itself runs.

1. Locate the viralman binary — try in this order: a. which viralman 2>/dev/null — use it if present b. test -x ~/.local/bin/viralman — use it if present c. git rev-parse --show-toplevel is the viralman repo root (confirm via name: viralman in .claude-plugin/marketplace.json) AND <root>/.venv/bin/python exists → use it d. ~/.claude/plugins/cache/*/viralman/*/.venv/bin/python — newest version e. None of the above → bootstrap is required

2. If bootstrap is required, run this in place:

   • Decide REPO location (CWD git toplevel → plugin cache → --path → clone into ~/viralman)
   • python3 -m venv .venv (require Python 3.10+)
   • .venv/bin/pip install --upgrade pip
   • .venv/bin/pip install flask
   • On Python ≤ 3.13: .venv/bin/pip install -e . (editable). On Python 3.14+: skip editable (the shim dispatches).
   • Write ~/.local/bin/viralman shim + chmod +x
   • If ~/.local/bin is not on PATH, print the addition guidance (do not auto-edit shell rc)
   • Run ~/.local/bin/viralman --no-browser --port 8765 in the background for 1.5 s, confirm response, then stop it
   • Print the verification OK message

3. If already installed, proceed to Step 0.5 (auto-update).

Boundaries:

• No sudo
• No auto-edit of shell rc files (advisory only)
• No global pip install outside the venv
• git clone never targets paths outside $HOME (exception: explicit --path)

---

Step 0.5 — auto-update (when already installed)

If a newer version is published, tell the user and update before continuing to Step 1.

1. Detect the current version based on the install path found in Step 0:

   • Plugin install (~/.claude/plugins/cache/.../viralman/): version field in that folder's .claude-plugin/plugin.json
   • pipx install (~/.local/pipx/venvs/viralman/): pipx list --short 2>/dev/null | awk '$1=="viralman"{print $2}'
   • Local clone: version in <repo>/.claude-plugin/plugin.json

2. Check the latest version (one network call):

   curl -fsSL --max-time 4 \
     https://raw.githubusercontent.com/art8engine/viralman/main/.claude-plugin/plugin.json \
     | python3 -c 'import json,sys; print(json.load(sys.stdin)["version"])'
   

   On network failure or 4 s timeout, skip silently (offline is fine).

3. If the versions differ, give the user a one-line notice and run the update for that install type:

   • Plugin install: cannot be done directly. Ask the user to run the two slash commands below in order, then wait (Claude Code does not have a single /plugin update <name> — must do marketplace refresh + reload as two steps).

     /plugin marketplace update viralman
     /plugin install viralman
     

     Then start a new session, or /reload-plugins (if available) to apply.
   • pipx install: pipx install --force git+https://github.com/art8engine/viralman
   • Local clone: cd <repo> && git pull --ff-only && .venv/bin/pip install --upgrade . (skip editable on 3.14+, keep the shim)

4. After updating, re-check that the new version matches latest. If yes, proceed to Step 1. If no, print manual update guidance but proceed to Step 1 anyway.

Boundaries:

• Any version-check failure (curl error, parse error) → skip silently. Never block Step 1.
• Never force-reinstall without the user's go-ahead — pipx --force only when a newer version was actually detected.

---

Step 1 — pick a channel

If the user's phrase already names a channel (e.g. "gitmail 셋업", "set up twitter for viralman"), treat the channel as decided. Skip the question and jump to that channel's branch (Step 3a/3b/3c/3d).

If $ARGUMENTS already contains the category, skip this. Otherwise ask once:

Which channel do you want to set up?

  1. gitmail  — cold email outreach to GitHub stargazers (most common)
  2. twitter  — automated X/Twitter posting (optional — default falls back to compose URL)
  3. reddit   — subreddit posting
  4. linkedin — LinkedIn posting

Reply with the number or the name.

Accept 1–4 or names. If the answer is still unclear after one follow-up, stop.

---

Step 2 — plain-text warning (conditional)

Trigger if: --plain flag given, or the user pasted something that looks like a token (ghp_…, sk-…, long alphanumeric, Bearer prefix).

Print once:

⚠ You pasted a token in plain text in this chat. The contents may be retained in the LLM context and the conversation log.

Safer alternative: read -rs -p '<KEY>: ' s && printf '%s' "$s" | viralman save-creds --stdin <KEY>; unset s; echo

If you continue anyway, the token will be saved via viralman save-creds --set <KEY>=<VALUE>.

Ask "Proceed? y/n". On yes → use --set for this and all remaining keys. On no → present the read -s pipe pattern for each key instead.

---

Step 3a — gitmail branch

Needs three credential bundles: GitHub, SMTP, and one LLM provider.

GitHub token — direct the user to https://github.com/settings/tokens?type=beta, Fine-grained token, Public Repositories read-only. Save:

read -rs -p 'GITHUB_TOKEN: ' s && printf '%s' "$s" | \
  viralman save-creds --stdin GITHUB_TOKEN; unset s; echo

Verify: curl -fsS -H "Authorization: Bearer $(grep ^GITHUB_TOKEN= ~/.viralman/.env | cut -d= -f2-)" https://api.github.com/rate_limit | python3 -c 'import json,sys; print(json.load(sys.stdin)["resources"]["core"]["limit"])' (expect 5000).

SMTP — any provider. Gmail shortcut:

viralman save-creds --set SMTP_HOST=smtp.gmail.com --set SMTP_PORT=587 --set SMTP_SECURITY=starttls
viralman save-creds --set SMTP_USER=<gmail> --set SMTP_FROM=<gmail> --set SMTP_FROM_NAME='<name>'
read -rs -p 'gmail app password: ' s && printf '%s' "$s" | \
  viralman save-creds --stdin SMTP_PASSWORD; unset s; echo

For SendGrid/Mailgun/SES: set SMTP_HOST, SMTP_PORT=587, SMTP_USER=apikey, SMTP_FROM, then pipe SMTP_PASSWORD the same way.

LLM provider — check for Claude Code first (which claude && claude --version). If found, no key needed. Otherwise save one of:

# Claude
read -rs -p 'ANTHROPIC_API_KEY: ' s && printf '%s' "$s" | \
  viralman save-creds --stdin ANTHROPIC_API_KEY; unset s; echo
# OpenAI
read -rs -p 'OPENAI_API_KEY: ' s && printf '%s' "$s" | \
  viralman save-creds --stdin OPENAI_API_KEY; unset s; echo
# Gemini
read -rs -p 'GEMINI_API_KEY: ' s && printf '%s' "$s" | \
  viralman save-creds --stdin GEMINI_API_KEY; unset s; echo

Public unsubscribe base (required for real sends, optional for dry-runs) — every outgoing mail carries an Unsubscribe: <base>/u/<token> link. If VIRALMAN_UNSUBSCRIBE_BASE is unset and the user runs a real send, gitmail aborts with a clear error rather than emit a localhost link recipients can't click. Save once:

viralman save-creds --set VIRALMAN_UNSUBSCRIBE_BASE=https://your-domain.example.com

The host needs the dashboard's /u/<token> route reachable on the open internet. If the user doesn't have a public deployment yet, point them at options like a tunnel (cloudflared tunnel, ngrok) or skipping the real-send step until they do — dry-runs work without it.

Verify end-to-end: viralman gitmail analyse "A quick test project". Done: "gitmail is hooked up — go to http://localhost:8765/gitmail and start a dry-run job."

---

Step 3b — twitter branch

First ask whether they need API posting or the compose-URL default is enough (one tweet at a time, no setup). If they pick the default, exit here.

For API posting there are two paths. Recommend OAuth 2.0 unless the user already has 4 OAuth 1.0a keys saved.

Default — OAuth 2.0 PKCE via dashboard (recommended)

One browser click; refresh tokens auto-renew. No 4-token paste.

1. App setup — go to https://developer.twitter.com/en/portal/dashboard, create an app under any project. In the app's User authentication settings: enable OAuth 2.0, type Confidential client, app permissions Read and write, callback URL http://127.0.0.1:8765/oauth/twitter/callback. Save Client ID + Client Secret.

2. Save the client values:

   viralman save-creds --set TWITTER_HANDLE=<handle>
   viralman save-creds --set TWITTER_OAUTH2_CLIENT_ID=<id>
   read -rs -p 'TWITTER_OAUTH2_CLIENT_SECRET: ' s && printf '%s' "$s" | \
     viralman save-creds --stdin TWITTER_OAUTH2_CLIENT_SECRET; unset s; echo
   

3. Run the OAuth flow — start the dashboard locally, then open the start URL:

   .venv/bin/python -m dashboard.server --host 127.0.0.1 --port 8765 &
   open http://127.0.0.1:8765/oauth/twitter/start
   

   X's "Authorize app" page → redirected back to 127.0.0.1:8765/oauth/twitter/callback, which exchanges the code, persists TWITTER_OAUTH2_BEARER and TWITTER_OAUTH2_REFRESH to ~/.viralman/.env, then renders a green "connected" page.

4. Verify:

   viralman save-creds --show-keys | grep TWITTER_OAUTH2
   

   Expect 4 keys: TWITTER_OAUTH2_CLIENT_ID, TWITTER_OAUTH2_CLIENT_SECRET, TWITTER_OAUTH2_BEARER, TWITTER_OAUTH2_REFRESH.

viralman post-twitter prefers the OAuth 2.0 bearer; on a 401 it auto-refreshes using the refresh token and persists the rotated pair.

Legacy — OAuth 1.0a 4-key (fallback)

Keep this only if the user already has the 4 keys configured. New users should use OAuth 2.0 above.

App permissions Read and write, generate Keys and tokens. Regenerate the Access Token after setting Read+Write — pre-upgrade tokens are read-only.

viralman save-creds --set TWITTER_HANDLE=<handle>
# four secrets, each via read -s | viralman save-creds --stdin:
#   TWITTER_API_KEY, TWITTER_API_SECRET, TWITTER_ACCESS_TOKEN, TWITTER_ACCESS_SECRET

Verify: viralman check-creds --platform twitter → twitter OK — @<handle> (id=...).

Common failures

• 403 = app still Read-only (regenerate tokens / re-authorize).
• 401 after a previous success → OAuth 2.0 refresh token expired or app permissions revoked; re-run the OAuth flow above.
• 429 = monthly free-tier cap.

Done: "X is hooked up. /viral --only x posts via the v2 Tweets endpoint."

---

Step 3c — reddit branch

Direct the user to https://www.reddit.com/prefs/apps. Create a script type app (not web, not installed) named viralman, redirect URI http://localhost:8765. Note the CLIENT_ID (short string under app name) and CLIENT_SECRET ("secret" field).

Save non-secrets:

viralman save-creds --set REDDIT_CLIENT_ID=<client_id>
viralman save-creds --set REDDIT_USERNAME=<username>
viralman save-creds --set REDDIT_USER_AGENT='viralman/0.1.0 by <username>'

Save secrets via read -s:

read -rs -p 'reddit client_secret: ' s && printf '%s' "$s" | \
  viralman save-creds --stdin REDDIT_CLIENT_SECRET; unset s; echo
read -rs -p 'reddit password: ' s && printf '%s' "$s" | \
  viralman save-creds --stdin REDDIT_PASSWORD; unset s; echo

Note: Reddit 2FA breaks PRAW password auth. Use a dedicated account without 2FA, or disable 2FA on this account.

Verify: viralman check-creds --platform reddit → reddit OK — u/<username>.

Common failures: 401/invalid_grant = whitespace in secret (re-run read -s), wrong app type, or 2FA active.

Done: "Reddit is hooked up. /viral --only reddit --subreddit <name> will post."

---

Step 3d — linkedin branch

LinkedIn requires a browser OAuth flow. Tokens expire in 60 days; re-run Steps 3–5 (below) to refresh without repeating app setup.

App setup — user goes to https://www.linkedin.com/developers/apps, creates an app tied to a company/org page (free personal page is fine). In the Products tab, request both "Sign In with LinkedIn using OpenID Connect" and "Share on LinkedIn" (auto-approved). In the Auth tab, add redirect URL http://localhost:8765/callback and note Client ID + Client Secret.

Save client_id: viralman save-creds --set LINKEDIN_CLIENT_ID=<id>

Save client_secret:

read -rs -p 'LINKEDIN_CLIENT_SECRET: ' s && printf '%s' "$s" | \
  viralman save-creds --stdin LINKEDIN_CLIENT_SECRET; unset s; echo

OAuth flow — construct and print this URL (fill in the saved client_id):

https://www.linkedin.com/oauth/v2/authorization?response_type=code
  &client_id=<LINKEDIN_CLIENT_ID>
  &redirect_uri=http%3A%2F%2Flocalhost%3A8765%2Fcallback
  &scope=openid%20profile%20email%20w_member_social
  &state=<random_8_chars>

User opens URL, clicks Allow, browser redirects to a localhost URL that won't load. They copy the code= parameter from the address bar.

Token exchange — have the user run:

read -rs -p 'LinkedIn auth code: ' code && \
  python3 - "$code" <<'PY'
import sys; sys.path.insert(0, "scripts/lib")
from creds import load
import urllib.parse, urllib.request, json, subprocess
c = load()
data = urllib.parse.urlencode({"grant_type":"authorization_code","code":sys.argv[1],
    "redirect_uri":"http://localhost:8765/callback",
    "client_id":c["LINKEDIN_CLIENT_ID"],"client_secret":c["LINKEDIN_CLIENT_SECRET"]}).encode()
resp = json.loads(urllib.request.urlopen(urllib.request.Request(
    "https://www.linkedin.com/oauth/v2/accessToken", data=data,
    headers={"Content-Type":"application/x-www-form-urlencoded"})).read())
print("token_received_chars:", len(resp.get("access_token","")))
subprocess.run(["viralman save-creds","--stdin","LINKEDIN_ACCESS_TOKEN"],
    input=resp["access_token"], text=True, check=True)
PY
unset code; echo

Capture person URN:

viralman check-creds --platform linkedin
# output: linkedin OK — <name> (sub=<id>)
# hint: set LINKEDIN_PERSON_URN=urn:li:person:<id>
viralman save-creds --set LINKEDIN_PERSON_URN=urn:li:person:<id>

Re-run viralman check-creds --platform linkedin — hint should disappear.

Common failures: redirect URI mismatch (Bummer error on auth page); 401 invalid_token = token expired, re-run OAuth steps; 403 ACCESS_DENIED = "Share on LinkedIn" product not added.

Done: "LinkedIn is hooked up. Token expires in 60 days — re-run /viralman-setup linkedin (skip app setup, just redo OAuth steps) to refresh."

---

Step 4 — final key-list confirmation

After any channel completes, run viralman save-creds --show-keys and confirm that the expected keys for that channel are present:

Channel	Required keys
gitmail	GITHUB_TOKEN, SMTP_HOST/PORT/SECURITY/USER/FROM/PASSWORD, + one LLM key
twitter	TWITTER_HANDLE + (OAuth 2.0: TWITTER_OAUTH2_CLIENT_ID / CLIENT_SECRET / BEARER / REFRESH) or (legacy: TWITTER_API_KEY / SECRET / ACCESS_TOKEN / ACCESS_SECRET)
reddit	REDDIT_CLIENT_ID/SECRET, REDDIT_USERNAME, REDDIT_PASSWORD, USER_AGENT
linkedin	LINKEDIN_CLIENT_ID/SECRET, LINKEDIN_ACCESS_TOKEN, LINKEDIN_PERSON_URN

If any expected key is missing from the list, flag it and offer to re-run the relevant sub-step.

---

Step 5 — Pre-approve viralman commands (optional but recommended)

Claude Code's permission layer prompts the user (and the auto-mode classifier may flag-and-block) on each fresh viralman <subcommand> invocation, and viralman's most common gitmail flows trigger this on every collect/send call. The fix is a one-time edit to ~/.claude/settings.json — but the agent must NOT self-edit it. Claude Code's harness blocks agent-driven permission self-grants. The agent must surface the snippet and stop.

After Step 4 completes, present this block verbatim to the user (Korean default; mirror the user's language if they've been speaking another):

권한 프롬프트를 매번 통과하지 않으시려면 ~/.claude/settings.json 의
permissions.allow 배열에 아래 한 줄을 추가해 주세요 (한 번만 paste, 새
세션부터 적용):

  "Bash(viralman:*)"

이 한 줄로 viralman 의 모든 subcommand (gitmail, twitter-reply,
save-creds, post-*, check-creds, ...) 가 커버됩니다.

⚠ auto-mode classifier 는 별개 레이어라, 대량 발송 류 명령
(예: viralman gitmail recipients --max-users 500) 은 위 룰을
추가하셔도 한 번 권한 다이얼로그가 뜰 수 있습니다. 그땐 한 번만 Allow
누르시면 됩니다.

이 스킬은 settings.json 을 직접 편집하지 않습니다 — Claude Code 의 harness 가
agent 의 self-permission-grant 를 차단하므로, 위 내용을 사용자께서 직접
paste 해주셔야 합니다.

Then stop — do not run a follow-up command, do not offer to "edit it for you", do not call any tool that touches ~/.claude/settings.json. Wait for the user to confirm they've pasted (or to skip).

This step is the canonical mitigation for the cross-project permission friction described in the gitmail / twitter-reply skill pre-flights. Users who skip this will keep hitting the permission dialog on every run; that's their call, not the agent's to bypass.