| name | mugshot |
| description | Guess which model or family likely wrote a passage from its stylistic fingerprints, and show the matched tells. Profiles common styles — gpt-ish ("Certainly!", "I'd be happy to", bold headers, numbered listicles), claude-ish (leading "I'll/Let me/Here's", em-dashes, "Great question"), and the universal AI tells (delve, tapestry, "it's not just X, it's Y") — then names the most likely suspect with a confidence band and the prints that matched. Use to ID the probable author of a chunk of AI text. Explicitly a heuristic, not forensic proof. |
mugshot
we know your prints.
When this skill is active, treat a chunk of text as a suspect at a line-up. Read
its style, match it against the known profiles, and name the family that most
likely wrote it — then show the prints you lifted, so the call is auditable, not
magic. Where tell flags the giveaways, mugshot uses those giveaways to name a
suspect.
Rules
- Read the prints, not the content. Authorship lives in style, not subject:
openers, hedges, headers, list shapes, punctuation habits — not what the text
is about.
- Know the usual suspects. gpt-ish: "Certainly!", "I'd be happy to",
"It's important to note", "However, it's worth", bolded section headers,
numbered listicles, "In conclusion", "Overall,". claude-ish: leading
"I'll …", "Let me …", "Here's …", em-dash asides, "Sure,", "Great question",
warm hedging. generic-AI: the universal tells any model leaves — delve,
tapestry, "navigate the complexities", "it's not just X, it's Y", "in today's
fast-paced …".
- Weigh the prints. A blatant opener counts more than a stray em-dash. One
weak print is noise; a cluster of strong ones is a confession. Lean on the
weight and the spread between suspects, not raw hit counts.
- Always show the prints. Never just announce a verdict. List the specific
matches that drove it so a reader can disagree.
- State the confidence — and the caveat. Give a band (low / medium / high)
from how far the top suspect beats the rest, and say out loud that this is a
guess. When nothing strong matches, the honest verdict is inconclusive /
could be human — say that, don't force an accusation.
What mugshot is NOT
- It is not forensic proof. Even with a model in the loop it's a hunch
dressed up as a line-up, not a chain of custody.
- It is not court-admissible. No watermark, no logprobs, no ground truth —
just style.
- Models drift and mimic each other. Today's gpt-ish phrasing is tomorrow's
everyone-ish phrasing; fine-tunes and system prompts move the tells around.
- A clever human can fake any style. Anyone who has read enough model output
can stuff a paragraph with "Certainly!" and frame an innocent.
Read a verdict as "this smells like X," never "X wrote this."
Example
Text: Certainly! I'd be happy to help. It's important to note that we
should weigh the options. In conclusion, the plan is sound.
Mugshot: most likely gpt-ish (high confidence). Prints lifted:
"Certainly!", "I'd be happy to", "It's important to note", "In conclusion".
Heuristic, not proof — a human could have typed every one of those.
Output
Deliver the verdict and the lifted prints, not a narration of the line-up. Don't
explain that you ran mugshot.py — name the suspect, the confidence band, and
the matched prints, keep the "heuristic, not proof" caveat, and stop. No
preamble, no extra closing.
Companion tool
mugshot.py in this folder runs the line-up. By default it asks a real model
(an LLM authorship/stylometry pass) when a provider key is configured, and falls
back to the offline regex heuristic otherwise — printing a note that setting
an API key (or --llm) unlocks real attribution. Force the path you want:
mugshot.py --llm — force the model-backed pass (fails loudly, exit 2, on a
provider error; no silent fallback). --provider / --model pick the backend.mugshot.py --parlor — force the offline regex heuristic, no network.
Every output mode reads the same verdict dict ({verdict, confidence, scores, prints}) and works for both paths: readable verdict (mugshot.py), every
matched print with offsets (mugshot.py --report), the full ranked scoreboard
(mugshot.py --all), or the structured dict (mugshot.py --json).
The parlor lineup is extensible: merge custom prints with --patterns FILE
(repeatable) or the MUGSHOT_PATTERNS env var (os-path-separator-joined paths),
each a JSON {"suspects": {"<name>": [[weight, "label", "regex"], ...]}}.
It pairs with tell: tell finds the prints (how AI does this read?), mugshot
names the suspect (whose prints are these?).
