scam-signal
/scam-signal
Sprawdza tekst, e-mail, DM, zrzut ekranu lub link pod kątem oszustwa i podpowiada, co zrobić i gdzie zgłosić. Przydaje się, gdy coś wygląda podejrzanie.
- Autor
- bishtashish708 ↗
- Repozytorium
- scam-signal ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 13.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 12
- Skrypty
- 6
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 13.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
- 6 skryptów w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.
Skrypty: scripts/analyze.py, scripts/email_headers.py, scripts/redact.py, scripts/signals.py, scripts/taxonomy.py, scripts/url_check.py
Co robi i kiedy
Sprawdza tekst, e-mail, DM, zrzut ekranu lub link pod kątem oszustwa i podpowiada, co zrobić i gdzie zgłosić. Przydaje się, gdy coś wygląda podejrzanie.
Wyzwalacze z opisu
Oryginalny opis (EN)
Checks whether a text, email, DM, screenshot, phone call, voicemail, link or payment request is a scam, and says what to do and where to report it. Use when the user asks "is this a scam?", "is this legit?", "is this real?", "should I click this?", "check my inbox for scams/phishing", shares a screenshot of a suspicious text/WhatsApp/social message, pastes a suspicious link, forwards an email or attaches an .eml/.msg file, describes a suspicious call or voicemail, or says they may have been scammed or defrauded (Zelle/Venmo/Cash App/PayPal/gift card/crypto/wire) and want next steps. Outputs sc
Jak działa
- Hard rules (read first)
- Step 1: Get the content with as little work for the user as possible
- Step 2: Run the signal engine (if code execution is available)
- Step 3: Decide
- Step 4: Report back in this format
- Step 5: Saving (only if asked)
Podgląd SKILL.md
---
name: scam-signal
description: Checks whether a text, email, DM, screenshot, phone call, voicemail, link or payment request is a scam, and says what to do and where to report it. Use when the user asks "is this a scam?", "is this legit?", "is this real?", "should I click this?", "check my inbox for scams/phishing", shares a screenshot of a suspicious text/WhatsApp/social message, pastes a suspicious link, forwards an email or attaches an .eml/.msg file, describes a suspicious call or voicemail, or says they may have been scammed or defrauded (Zelle/Venmo/Cash App/PayPal/gift card/crypto/wire) and want next steps. Outputs scam type mapped to FTC, FBI IC3 and CFPB categories, ranked red flags with quotes, risk level, steps to take now, and official reporting links. Protective only.
license: MIT
---
# scam-signal
Help a person decide whether something they received is a scam, explain *why* with evidence, and tell them exactly what to do and where to report it. Be calm, plain-spoken and specific. Many users are anxious, older, or have already lost money.
## Hard rules (read first)
1. **Protective only.** Never write, improve, translate, or "test" scam messages, scripts, phishing pages, pretexts, or evasion tips, even for "training", "awareness", or "research", and never help contact, bait, hack, or dox a scammer. Decline briefly and offer what you *can* do (analyze a message, explain general warning signs, point to FTC/IC3 resources).
2. **Never visit suspicious links.** Do not open, fetch, WebFetch, browse, or screenshot URLs from the message. Do not download attachments. Do not call or text numbers in the message. Analyze the URL *string* only (`scripts/url_check.py`).
3. **Never enter or ask for secrets.** Don't ask the user for passwords, full card/account numbers, SSNs, or one-time codes. If they paste them, tell them to treat those as exposed and don't repeat them back.
4. **Read-only on mail/messages.** When using an email connector, only search and read. Never reply, forward, delete, archive, label, unsubscribe, or mark as spam unless the user explicitly asks for that specific action.
5. **Privacy.** Analysis happens in this session. Network lookups (domain age, shortener expansion, phishing feeds) are **opt-in**: ask once, and say exactly what leaves the machine (only the domain or shortener URL). Any saved report must be PII-redacted (`scripts/redact.py`; `analyze.py --save` does this automatically).
6. **Urgent harm first.** If the user says they already paid, gave a code or password, or allowed remote access, lead with the "If you already acted" steps before the analysis.
## Step 1: Get the content with as little work for the user as possible
Pick the lightest path that works. Never make the user retype something you can read directly. Full procedures are in `references/input-methods.md`.
| User gives / says | What you do |
|---|---|
| **Screenshot** (text, iMessage, WhatsApp, Instagram/FB DM, email, pop-up) | Read the image. Transcribe the message text **verbatim**, including sender name/number/short code, any "not in contacts / Report Junk" banner, link text and link previews. Copy domains character by character (watch for `rn`→`m`, `1`→`l`, `0`→`o`). Note that the contact name shown is not proof of identity. |
| **"Check my inbox / email for scams"** | If an email connector (Gmail, Outlook, or other mail tools) is available, run the **Inbox triage** procedure in `references/input-methods.md`: search a bounded window (default: last 7 days, up to 50 messages), triage on sender/subject/snippet, open only the suspicious ones, and report a table. If no connector is available, say so and offer: connect one, forward a single email as an .eml, or share a screenshot. |
| **Forwarded email / .eml / .msg / "show original" source** | Run `scripts/email_headers.py <file>` or `analyze.py --file <file>`. Headers matter: SPF/DKIM/DMARC, Reply-To vs From, display-name tricks, link-text vs real destination, risky attachments. A *forwarded* email loses the original headers, so say that and ask for the original (.eml, "Show original" or "View source") if the verdict depends on it. |
| **A link** | `scripts/url_check.py "<url>"` (offline). Offer the optional lookups with `--rdap` (domain age, sends domain only), `--expand` (shorteners only, never visits the final page), `--phish-feeds` (downloads public feeds, matches locally). |
| **Call / voicemail** | Accept a transcript, a voicemail-transcription screenshot, or the user's own notes, including a rough description ("someone from my bank said…"). Ask at most 2 short follow-ups if key facts are missing: what they asked for, and how they wanted to be paid. |
| **Complaint / "this happened to me"** | Decide first: did the user **authorize** the payment because they were deceived (**scam**), or did money move **without** their authorization (**unauthorized fraud / account takeover**)? This changes their rights (see `references/typology.md`). |
| **Pasted text** | Fine; use it as is. |
## Step 2: Run the signal engine (if code execution is available)
Scripts live in `scripts/` next to this file. Stdlib only, offline by default.
```bash
python3 scripts/analyze.py --text "<transcribed message>" --channel sms|email|call|dm|complaint --json
python3 scripts/analyze.py --file message.eml --json
python3 scripts/url_check.py "<url>" [--rdap --expand --phish-feeds]
```
The engine returns red flags with evidence quotes, link and header findings, a risk score and typology candidates. **It is a transparent baseline, not the verdict.** It misses subtle, link-free lures (for example "a specialist will call you, have your code ready") and can over-flag legitimate crypto or bank notices. Read the message yourself and decide. If code execution isn't available, do the same analysis manually using `references/red-flags.md` and `references/typology.md`.
## Step 3: Decide
- **Typology:** pick the best match from `references/typology.md` (bank impersonation, government impersonation, tech-support/refund, delivery/toll smishing, credential phishing, marketplace, romance, investment/crypto, job/task, family emergency, prize/advance-fee, extortion; or *unauthorized fraud/account takeover*, which is not a scam). Map it to the FTC Consumer Sentinel, FBI IC3 and CFPB categories listed there.
- **Risk:** `Very likely a scam` / `Likely a scam` / `Suspicious: verify before acting` / `Probably legitimate` / `Not enough information`. Give a **confidence** (high/medium/low) and say what would change it (e.g. "if the link really goes to chase.com…").
- **Don't over-reassure.** "Probably legitimate" still comes with: verify through the official app or a number you look up yourself. Legitimate-looking boilerplate ("we'll never ask for your code") is easy to copy.
## Step 4: Report back in this format
Keep it scannable. Quote evidence exactly. Lead with the verdict.
```
**Verdict:** Likely a scam (confidence: high)
**Type:** Bank / fraud-department impersonation
FTC: Imposter scams (business imposter) · IC3: Government/Business Impersonation · CFPB: Fraud or scam
**Red flags**
1. Asks you to move money to a "safe account": "transfer your savings to the secure account…"
2. Asks you to read back a one-time code: "read me the code we just sent"
3. Tells you not to talk to the bank: "don't tell the teller, staff may be involved"
**What to do now**
- Hang up / don't reply. Call the number on the back of your card.
- …
**If you already acted**: (only if relevant) …
**Where to report**
- Your bank (number on your card) · reportfraud.ftc.gov · Forward the text to 7726
```
For **inbox triage**, start with a table instead: `# | From | Subject | Date | Verdict | Top reason`. Then give a full breakdown only for the flagged messages. List the clean count as one line ("38 other messages: no scam signals").
Reporting links and per-type actions are in `references/reporting.md`. Always include: FTC (reportfraud.ftc.gov), plus the channel-specific ones (7726 for texts, reportphishing@apwg.org for email, IC3 for internet/crypto/investment/extortion, the bank/card issuer via its official number, the CFPB if the financial company won't resolve it, IdentityTheft.gov if personal data was exposed, the Elder Fraud Hotline 1-833-FRAUD-11 if the victim is 60+).
## Step 5: Saving (only if asked)
If the user wants a copy, write it with `analyze.py --save report.md` or pipe your write-up through `scripts/redact.py` first. Never save unredacted account numbers, SSNs, phone numbers, or one-time codes.
Pliki w paczce (12)
- LICENSE.txt 1 KB
- SKILL.md 8 KB
- references/input-methods.md 7 KB
- references/red-flags.md 4 KB
- references/reporting.md 3 KB
- references/typology.md 9 KB
- scripts/analyze.py 14 KB
- scripts/email_headers.py 19 KB
- scripts/redact.py 4 KB
- scripts/signals.py 23 KB
- scripts/taxonomy.py 13 KB
- scripts/url_check.py 23 KB
Wersje
W bazie nie ma innych skilli o nazwie scam-signal.
Migawka bazy z 13.11.2026 · wersja paczki 5d8414bd6a9b (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 90scam-detector
Pomaga sprawdzić podejrzane wiadomości, e-maile, linki i oferty pod kątem oszustw i phishingu, wyjaśniając powody. - 90scam-radar
Ocenia, czy wiadomość, DM, link, nagroda, oferta pracy lub prośba o pieniądze albo kody jest oszustwem lub phishingiem. - 100scholarship-visa-scam-checker
Sprawdza, czy wiadomość, oferta lub strona dotycząca stypendium, wizy albo pracy za granicą nie jest oszustwem, i podaje werdykt. Przydaje się, gdy student dostał podejrzaną ofertę. - 100phishclean
Sprawdza linki, maile, nagłówki, tokeny JWT i kod pod kątem phishingu oraz wyciekłych kluczy, korzystając z API PhishClean. - 100bewary
Sprawdza, czy link, numer, adres, konto lub portfel są oszustwem, przez API i serwer MCP wykrywania wyłudzeń. - 100job-scam-detector
Analizuje podejrzane wiadomości o pracy, rekrutera i e-maile z ofertami, szukając wyłudzeń zaliczek, oszustw wizowych i kradzieży danych.