Dunking anonymously

A quote-post delivers your whole audience to the original author. For an account with forty followers that is a pile-on whatever your caption said. This renders the post as an image instead, so the words can be quoted without the handle.

Idea from Luis Villa, who built the same feature for himself (https://bsky.app/profile/lu.is/post/3mv6ynbcjr62b). His render greys the identity out; this one blacks it out.

Procedure

python3 scripts/dunk.py https://bsky.app/profile/someone.bsky.social/post/3mv6...

Takes a bsky.app URL, an at:// URI, or handle/rkey. Writes <rkey>.png to the working directory, prints the alt text to stdout, and writes <rkey>.json holding the source URI for your own records. The source URI stays out of both the image and the alt text.

flag	effect
--redact "phrase"	blacks out a literal phrase, repeatable
--stamp	adds the rotated REDACTED stamp
--show-date	prints the post date in the footer
--out PATH	output path
--width N	image width, default 1200

Then read the rendered PNG, actually open it, and re-run with --redact for anything in the body that identifies the author. This is the step that gets skipped.

Deps are PIL and the standard library. DejaVu Mono ships with the container. The API call is unauthenticated, so it needs no token and leaves no trace in anyone's notifications. See README.md for the host fallback and network surface.

Redactions applied

Avatar, display name and handle become black bars. Every at-mention facet in the body is blacked out, since a mention is a second person's handle and a strong search key for finding the original. Embedded images, video, link cards and quoted posts render as [image not shown] and friends rather than being fetched, since an embedded image can carry a face, a location or a watermark.

Engagement counts are dropped. The date is dropped unless --show-date, because a date plus one distinctive sentence narrows a search to a single post.

The script cannot strip identity from the post's own words. Bluesky indexes full post text, so a post naming its author's employer, town or cat stays traceable by anyone who pastes a sentence into search. --redact is the only defence and it is manual.

When NOT to use this skill

situation	what owns it
reading, searching or sampling Bluesky content	browsing-bluesky
sorting an account list by topic	categorizing-bsky-accounts
posting a link with a card preview	muninn_utils.bsky_card
bulk-classifying the repliers under a post	muninn_utils.bsky_moderation

Quote-post public figures and large accounts normally. Anonymising a senator is not protection. It removes the reader's ability to check the source while protecting nobody. The skill is for accounts small enough that attention itself is the harm.

Earned exceptions. Anonymising a large account is earned when the post is being shown as a specimen of a pattern rather than as one person's position, and the handle would pull the thread toward the individual. It is not earned when the reader would reasonably want to verify who said it.

Abandon mid-run when the post cannot be anonymised, meaning the author's identity is the claim, the text names them, or the render still reads as them after --redact has eaten half the body. At that point either quote-post it with the link or drop it. A half-redacted post that everyone recognises is worse than both, because it performs restraint while doing the pile-on.

Do not use it to launder something you would not say with the link attached. If the criticism only works because nobody can read the original in context, fix the criticism.

Common failure modes

Black bars in the header, author named in the body. The render looks correctly anonymised and one sentence still says "when I worked at Foo". Signal: you did not read the PNG, you read the terminal. Open the image every time and pass --redact per identifying phrase.

Two bars where one phrase should be. A multi-word --redact phrase renders as separate bars with visible gaps, and word lengths leak. Signal: gaps inside what should be one black run. Adjacent redacted tokens are meant to merge across the spaces between them. If they do not, the merge logic in render() broke.

Mentions survive the redaction. Signal: an @handle renders as readable text. The script reads mention facets, not the @ character, so a post whose facets were stripped by the client that wrote it has no mention to find. Catch it by reading the PNG and use --redact "@handle" as the fallback.

no post at at://.... The post is deleted, the author blocks the public AppView, or the rkey is wrong. Not a transport failure, so refetching will not fix it.

403 from the AppView. public.api.bsky.app refuses some proxied environments. The script retries on api.bsky.app, which serves the same lexicon, and only gives up when both refuse.

Verification

python3 scripts/dunk.py <url> --stamp

Then open the PNG. Three checks: no handle or display name anywhere in the image, no at-mention rendered as text, and the alt text opening with "Screenshot of a Bluesky post, author hidden:".

A bad success here is a clean-looking render that still identifies its author — every bar in place, and the body naming a workplace or a town. It passes every automated check the script can make, which is why the read-through is part of the procedure rather than a nicety.

Posting it

Attach the PNG in whatever client you post from and paste the printed alt text. It opens with "Screenshot of a Bluesky post, author hidden:" so screen-reader users get the framing that sighted readers get from the black bars.