rubicon/ai-skills

identity-theft

Overview

This skill rewrites supplied prose in a selected fictional character's voice while preserving the input format and protecting every non-prose construct. One personality-agnostic engine drives all characters; each character is one dossier under references/personalities/. The roster grows by adding dossiers — the engine never changes.

The engine is a strict transformer with an output gate, not a Markdown compiler wearing a fake mustache. The model transforms; a bundled script verifies; output ships only if it passes. The transformation is not guaranteed to preserve everything — the accepted output is, because invalid output is rejected.

Primary use: easter-egg documents (README.swanson.md beside a repo's real README.md). Secondary use: paste text in chat, get converted text back.

Invocation

• A personality is required. It is named at invocation ("swanson-ize this", "convert README.md as Yoda"). There is no global default personality. If none is named, ask which character before doing anything.
• Aliases select the dossier. Each dossier's header declares trigger aliases; match the user's request against them across the roster.
• Modes are per-personality. A dossier may declare modes; one is marked default_mode. Select a non-default mode only when the user names it. Ron Swanson defaults to The Full Swanson.
• Input is pasted text, HTML, or Markdown in chat, or a file path.
• Output format always matches input format. Markdown in → Markdown out. HTML in → HTML out with every tag and attribute intact and only text nodes converted. Plain text in → plain text out.
• File input writes a sibling file <name>.<suffix>.<ext> — e.g. README + suffix swanson + md → README.swanson.md. The suffix is declared by the dossier. Never write in place unless the user explicitly asks for it.

Workflow

Run this sequence for every conversion:

1. Load the dossier for the selected personality from references/personalities/<id>.md. Read only that one file — other personalities stay out of context. Apply its speech patterns, vocabulary, riff themes, anti-patterns, preamble guidance, and selected mode.

2. Classify every span of the source into the three content classes (Protected / Conservative / Free) defined below. Do this before writing a single word.

3. Transform the Free and Conservative spans in voice; copy Protected spans through byte-identically; prepend the preamble per the placement rules.

4. Write the result to a temporary file — never directly to the sibling path.

5. Run the output gate from this skill's directory:

   python3 scripts/validate_output.py SOURCE TEMP
   

   SOURCE is the original file (or a temp file holding the pasted source); TEMP is the transformed temp file. Exit 0 = pass; exit 1 = fail with a per-region report.

6. On pass, move the temp file to the sibling path <name>.<suffix>.<ext>. On fail, do not create the sibling file. Report the damage the gate named — region type, occurrence index, expected content, what was found — and retry once with stricter adherence, correcting exactly the regions the gate flagged. If the retry also fails, keep the temp file for diagnostics only and report the failure. Never ship failed output and never silently downgrade to writing anyway.

7. Chat inputs validate the same way. Pasted Markdown or HTML is written to a temp file, validated against the supplied source (also via a temp file) before the result is returned. Plain text with no recognized protected constructs needs no region validation.

The gate enforces structural invariants only: protected-region type, order, count, and exact contents (compared region-by-region, not one concatenated hash); link-class rules; zero unresolved same-document anchors; matching output format, line-ending style, and final-newline state. Factual integrity, quote classification, legal-text recognition, and voice quality are enforced by these instructions and by review — the script does not check them, so honor them yourself.

Preamble

The preamble is the engine's wrapper — the character saying "here you go." It is not part of the transformed body; mode liberty rules govern the body only. It appears in every output, every personality, every mode.

• Length: 1–3 lines, in character, funny. Brief & Blunt uses exactly one short line.

• Placement:

  Input	Preamble goes
  Plain text	At the beginning.
  Markdown without frontmatter	At the beginning, as a blockquote.
  Markdown with YAML frontmatter	Immediately after the closing frontmatter delimiter, as a blockquote.
  HTML document	First child of <body>, as a <blockquote>.
  HTML fragment	Before the first convertible text node.

Emit the transformed document plus this preamble and nothing else. Add no "Here's the Ron Swanson version" framing, no trailing meta-commentary, no notes about what you did.

Content handling — three classes

Classify every span before transforming.

Class 1 — Protected (byte-identical)

Executable, referential, legal, and machine-readable content. Copy through character-for-character, in original order, unchanged:

• fenced and indented code blocks
• inline code spans
• <code> / <pre> / <kbd> / <samp> / <script> / <style> / <textarea> / <template> contents
• HTML comments
• YAML frontmatter
• badges / shields
• HTML tags and attributes (including alt, title, aria-label — accessibility text is preserved in HTML)
• file paths
• version numbers
• license and legal text
• link destinations (see Link classes)

Code is data, not prose. A fenced block keeps its info string; an indented block keeps its indentation; an inline span keeps its backticks and contents.

Class 2 — Conservatively transformed

Headings, table headers, navigation labels, and instructional prose (requirements, installation steps, warnings). Transform under the meaning floor and factual-integrity rules; the result must remain followable and scannable. Exception — headings: reproduce every heading verbatim and in original order, including duplicate headings; do not rename, merge, or deduplicate them, because same-document anchors depend on exact heading text (see Anchor repair for the one permitted change).

Class 3 — Freely transformed

Descriptive and expressive prose. Full personality treatment, subject to factual integrity and anti-caricature. Default liberty is full editorial license — the character may riff, add asides, and add in-character commentary; a dossier or mode may restrict this.

Class boundaries

• Markdown image alt text is convertible (class 2). HTML attributes are not (class 1) — alt, title, and aria-label in HTML stay verbatim.
• Natural-language instructions ("run the installer, then restart the service") are prose (class 2), not protected commands. Command protection is structural: it applies to explicitly delimited machine-readable content only (code fences, spans, the tags above).
• Table structure is protected: the delimiter row (| --- | --- |), column count, and cell layout stay exact; only cell text that is class 2/3 prose may be voiced. Task-list checkbox syntax (- [x] / - [ ]) and its state are protected; only the trailing label text is eligible.

Link classes

• External URLs and non-fragment destinations: byte-identical, always. This covers inline destinations, reference-style definitions, autolinks, bare http(s):// URLs, and HTML href/src. Only link text is voiced.
• Relative links with a path component (guide.md#setup): external to the current document; unchanged.
• Same-document fragment links (#installation): may change only to repair a renamed heading's anchor, and the repaired fragment must resolve to exactly one heading in the output.
• Fragments inside code or other protected regions: untouched.

Anchor repair

• Because headings are reproduced verbatim (class 2 exception above), anchors normally need no repair. Repair applies only if a heading legitimately changed.
• Compute slugs with GitHub's heading-ID behavior. Use the GitHub-compatible slug implementation already vendored in the runtime — gfm_slug in scripts/validate_output.py — not a hand-rolled algorithm.
• Duplicate headings receive renderer-compatible numeric suffixes (-1, -2, …) exactly as the slug function assigns them.
• The gate confirms every rewritten same-document fragment resolves uniquely; an unresolved fragment fails validation.

Quotes

• Attributed quotations — a named person, a testimonial, cited material, or any blockquote followed by an em-dash attribution — are reproduced byte-for-byte, including punctuation and the attribution line. This is a hard rule, not a hope.
• Decorative callout text converts — its existing wording may be re-voiced in tone. But no blockquote gains new sentences, asides, or commentary: converting a callout means re-voicing the text that is there, never appending to it or changing its factual content or length beyond tone. Admonitions, notes, and warnings additionally keep meaning-floor followability.
• Unsure → untouched. If you cannot tell whether a blockquote is a real quotation, leave it exactly as written.

Factual integrity (universal invariant)

Voice changes tone, never content or claims. Preserve factual propositions, names, dates, quantities, comparisons, and qualification levels. In-character additions must be clearly rhetorical and must not:

• introduce personal authorship or attribution claims absent from the source (do not turn an anonymous "we built this" into "I built this");
• fabricate instructions, remedies, or steps that replace the source's actual guidance (do not replace "check connectivity" with invented advice);
• assert reliability or quality claims the source does not make ("works most of the time");
• convert qualified claims into absolute ones;
• attribute fictional opinions to real people or organizations.

Quantifiers are qualifications: "multiple machines" does not become "every machine you own", "some users" does not become "all users".

Do not editorialize about the document's own content or the task itself (never call a documented section "a fake problem"). Voiced commentary stays in-character narration about the subject matter, not meta-commentary that undermines the document's credibility.

Voice quality (anti-caricature)

• Character comes from sentence rhythm, worldview, priorities, and rhetorical habits — not catchphrase insertion.
• No signature phrase more than once per document. Not every paragraph gets a character reference. Rate-limit persona markers rather than leaving frequency to chance.
• Scale characterization to source length: a short document gets a light touch, not a saturated one.
• When voice conflicts with the reader's ability to use the document, the voice yields — clarity wins.

Intellectual property

• Canonical dialogue may be studied for analysis but is never copied into dossiers or emitted as stock output.
• Do not reproduce or substantially paraphrase memorable copyrighted dialogue. Capture abstract stylistic traits, not quotation banks.
• All dossier examples are newly written.

File fidelity

• Preserve LF vs CRLF line endings and the presence or absence of a final newline. The gate fails output that changes either.
• Do not normalize whitespace outside converted prose.

Markdown dialect (v0.1.0)

Supported: GitHub Flavored Markdown — YAML frontmatter, fenced and indented code blocks, tables, task lists, autolinks, reference-style links, and raw HTML blocks.

Conservatively preserved (treated as Protected, not transformed): MDX, template languages (Hugo/Jekyll shortcodes), Mermaid, math, and app-specific extensions.

Non-goals (v0.1.0)

• Generation of canonical dialogue or exact impersonation
• Modification of code, configuration, or executable commands
• Full MDX / template-language support
• Semantic rewriting of legal text
• Transformation of HTML accessibility attributes
• In-place file modification by default
• Chunked large-document processing

Known limitations

• Intended for README-scale documents. Large documents are processed in one pass and may exceed practical context or voice-consistency limits.
• MDX, template languages, and app-specific Markdown extensions are conservatively preserved, not transformed.
• If validation fails after one retry, no final output file is produced.
• The gate's guarantee is structural only; factual integrity, quote classification, legal-text recognition, and voice quality rest on these instructions and on review.