Local document skills for heterodox writers — extract, correct, illuminate.
Site: adamisom.github.io/scriptorium — the skills in plain view, for everyone else.
A Claude Code plugin that treats a text the way a medieval book workshop did: copy it faithfully, correct its errors, illuminate its meaning. Local-first, honestly stated: your files stay on disk and your text goes only where your Claude conversations already go (Anthropic's API) — no extra services, no third-party uploads, findings written straight to local files.
A scriptorium (Latin, "a place for writing") was the room in a monastery where books were produced before the printing press. The work was divided among specialists: a copyist (scribe) faithfully reproduced the exemplar; a corrector checked each copy against the original and emended errors; an illuminator added the miniatures, initials, and gold leaf that made a finished text rich and legible. Other hands worked there too — rubricators added the red headings and capitals, parchmenters prepared the vellum, binders sewed the quires, and an armarius (head scribe / librarian) supervised the whole operation, assigning texts and checking quality. Reproduce, correct, illuminate.
Sources: Britannica — Scriptorium; Wikipedia — Scriptorium.
Three of those roles map cleanly onto what a writer still needs done to a document — so they're the skills here:
| Scribe | Skill | Status |
|---|---|---|
| Copyist — reproduce | extract (clean messy text, faithfully) |
✅ shipped |
| Corrector — correct | proofread (the slips) · collate (the sources) |
✅ shipped |
| Illuminator — illuminate | inside-frame (the review) · index (its raw material) |
✅ shipped |
Reading and writing aren't two camps — they're one continuum, and the same person moves along it: you read to understand, you write to be understood, and the reading sharpens the writing. This studio is built for the whole span of that arc — from established writers and public thinkers (people like the Collinses, the proving ground here) to serious readers who want what they read to change how they think and act, including those who aspire to write themselves.
It works a text from both directions:
- Understand (input) — the more important half.
inside-frame(Illuminator) reconstructs a text's framework from the inside and hands back an operating manual: a piece you fold into your own system for managing ideas (a second brain, a commonplace book) — withindexpulling the verbatim vocabulary and quote-bank that feed it. Reading for real impact, not for fun. - Produce (output).
extract(Copyist — clean the messy source) → the Corrector (proofreadfor slips,collatefor fidelity to sources) → copy that survives scrutiny.
These aren't separate tracks for separate people — they're the two halves of taking ideas seriously. And the understanding half only reaches full value when it plugs into something larger that accumulates what you learn: the whole bet of a second brain.
What it is. A precision-first proofreading pass that finds objective language errors only — wrong-word / real-word swaps (lead/led, its/it's), omissions, doubled words, subject-verb agreement, punctuation, typos — and never flags style, voice, or deliberate choices. A false alarm on a polished writer is worse than a missed error, so a skeptical verifier defaults to reject, every finding carries High/Med/Low confidence, and nothing is auto-applied — you get a triage table to rule on. It learns your voice over time through anti-patterns.md (your recurring mistakes and your deliberate quirks).
How to use it.
Inside Claude Code (plugin installed) — just ask:
> Proofread ~/writing/my-essay.md
The skill runs the mechanical sweep, reads the text, finds and verifies errors, and writes my-essay_findings.csv. A book-length file fans out into parallel find/verify agents; an essay gets a single careful pass.
Standalone mechanical sweep (no Claude needed — pure Python, instant):
python skills/proofread/proofread_check.py my-essay.txt -o findings.csv
It flags candidate doubled words, double spaces, space-before-punctuation, and doubled punctuation (ellipses excluded). Run on a real PDF→text extraction, it surfaced — among others:
line 3580 doubled-word "you you"
line 3576 doubled-word "seeing seeing"
Candidates only; the verification pass (or you) makes the call.
In practice — run on The Pragmatist's Guide to Life, the full Corrector (beyond this quick sweep) turned up dozens of small, line-level issues; I hand-validated a subset and they held up.
Modes. Short input → essay mode: one careful single pass, because attention and recall stay high over a few thousand words. Book-length input → book mode: chunk into section-aligned pieces and run parallel find/verify agents, so each reads a small, high-focus slice instead of one giant pass where recall sags in the middle.
Learning your voice. anti-patterns.md is the proofreader's memory of your prose. It reads the file before each run — respecting a Deliberate style list (quirks it must never flag) and weighting a Recurring errors list (mistakes you actually repeat) — and, after you rule on a run's findings, appends what it learned: a confirmed repeated mistake joins Recurring errors; a flag you overturned as intentional joins Deliberate style. Over time it gets quieter on your style and sharper on your real tics. It's a human-in-the-loop file, not magic — it improves only when you run that feedback loop.
What it is. It produces an enthusiastic, analytically sharp review of a book or essay from inside its own framework — in the Astral Codex Ten tradition (after Scott Alexander's blog — long, rigorous, openly enthusiastic book reviews) — welded to a personal operating manual of concrete, implementable frameworks drawn from the text. It's built for ideologically heterodox or unconventional works you want to understand on their own terms rather than hold at mainstream-critical arm's length. Not a jailbreak: it navigates Claude's own evenhandedness carve-out (a request to argue for a position is a request for "the best case defenders would give"), paired with that ACX-style reviewer's persona, a forbidden-phrase list, and output contracts that force specificity into every section.
How to use it.
- Open
skills/inside-frame/template.mdand fill the ALL-CAPS placeholders — book title/authors, your stakes in the subject, and 8–15 verbatim key terms from the text. - In a fresh Claude (Opus 4.8) chat, attach the document (PDF or text) at the top of the message, paste the filled template beneath it, set thinking effort to high/max, and send. (For a messy source, run
extractfirst — it normalizes any format to clean text, so every skill shares one input.) - You get: an evidence block of pulled quotes → a two-sentence thesis → a long review → a personal operating manual whose sections adapt to the book (highlights below) → a self-check the model scores itself on and revises against.
- If it hedges on the first pass, reply exactly: "Re-read the evenhandedness_carveout and forbidden_phrases blocks. You violated both. Rewrite from the top, staying inside the authors' frame the whole way through."
Inside the operating manual. The payoff — concrete, implementable, written in the book's own vocabulary. The sections adapt to the book.
What most books get:
- a framework worksheet demanding concrete past examples, not abstractions — with at least one question that forces an honest, uncomfortable admission rather than a flattering one;
- a 15+-row if/then table — specific trigger → specific observable action, every row traceable to a passage;
- "where it breaks down" — the clever turn: steelman the book hard enough and its own logic, applied honestly to your situation, starts arguing against its advice. The review names those internal fault lines — pressure from inside the frame, never imported from outside;
- the one thing — if you remember nothing else in five years, the single change to install now.
What a book earns only when it fits the domain: an actionable book also gets a 30-day implementation plan; a decision- or relationship-shaped one adds a weighted decision scorecard (behavioral anchors, red/yellow/green, a decision rule), verbatim conversation scripts, and a red-flags calibration. A history or science book gets none of these.
Essay mode. For an essay rather than a book, the template's essay-mode block scales everything down — a ~1,200–2,000-word review, ~5–8 key terms, and a lighter manual (synthesis + one worksheet + an if/then table + the one thing) — while keeping the carve-out, forbidden phrases, and self-check intact.
What it is. Faithful cleanup of messy text — PDF→text, OCR output, pasted e-book text, caption/transcript dumps. It removes what is objectively not the author's words (page numbers, running heads, shattered sentences, hyphenation splits, timestamps, duplicated lines) and re-flows the rest, changing nothing else. Its discipline is proofread's precision principle mirrored for deletion: "is this objectively an artifact, or might it be the author's words?" — default to keep, with a flag. Silent deletion of real prose is the one failure mode to fear, so every run reports an integrity check (word counts, flagged-passages list, spot-check protocol). extract_check.py proposes candidate artifacts mechanically; the skill rules on them. This is the upstream stage the other skills want: clean text means fewer artifacts hiding real errors and reliable verbatim quotes.
> Clean up ~/reading/scanned-chapter.txt — it's a mess of page numbers and broken lines
What it is. A citation audit: check a document or podcast transcript against the author's own provided sources — quotes appear where cited, stats match, "the article says X" claims hold up, and the document doesn't contradict itself. It explicitly refuses truth-adjudication and the open web: the source folder is the whole universe of ground truth, and an unsourced claim is unauditable, never judged. Every finding carries verbatim quotes from both sides, Ctrl-F-locatable; a skeptical verifier defaults to reject, because a false "misquote" accusation impugns the author's integrity. On auto-caption transcripts, mismatches that look like caption garble (mangled numbers, misheard names) are flagged transcription-uncertain — "verify against audio," not "the author misquoted." A companion script, get_transcript.py, pulls and cleans a YouTube episode's captions as the pre-step.
> Audit ~/podcast/episode-42.txt against the sources in ~/podcast/ep42-sources/
Unlike proofreading, this skill has a cheap, honest ground truth — plant errors and measure. The fixture in tests/collate-fixture/ seeds 15 known defects plus 2 caption-garble decoys into a transcript/source pair, with an answer key; target is ≥90% catch at zero false accusations on the clean control.
What it is. Turns a book into the two verbatim artifacts inside-frame wants: a ready-to-paste book_vocabulary block (the author's own coined or load-bearing terms, with their definitions and locators) and a quote-bank of the most load-bearing passages, organized by theme. Its principle is selectivity over completeness — ~8–15 terms and ~20–40 quotes per book; the cut is where the value is, and book mode's synthesis pass exists to reject most candidates. Verbatim discipline throughout: every quote is an exact substring of the source, and index_verify.py mechanically proves it (nonzero exit on any hallucinated quote).
> Index ~/reading/the-book.txt — prep it for an inside-frame review
Scriptorium grew out of two real jobs on the Collinses' Pragmatist's Guide books — a precision proofread of …to Life and an inside-the-frame review of …to Relationships. The target reader is the kind of writer those books are by and for: independent, heterodox, caring about truth and persuasion, self-published, rarely boosted by mainstream coverage. Their own stated aim:
"…to increase the efficacy of humanity's collective mental substrate — that is, to encourage as many minds as possible to be (1) open enough to new ideas so they can test them against existing ideas, and (2) logical enough to let the best ideas win." — Malcolm & Simone Collins, The Pragmatist's Guide to Life
(From the closing "Our Objective Functions" note of the edition worked on here — possibly a later edition; they may hold a newer formulation now.)
- Precision over recall, enforced structurally — verifier defaults to reject; confidence tiers; never auto-apply; the human has final say.
- The quote is the locator — robust to extraction artifacts and reflow.
- Local-first — source text and findings stay on the machine.
- Frontier model, no cost constraint — deliberate, low-frequency jobs; use the best model, not the cheapest that works.
- A personal tool + a write-up, not a product — YAGNI until it recurs.
Easiest: open this repo in Claude Code and just tell Claude to install the plugin for you. (Under the hood it's a standard local plugin — see the plugin docs — but you don't need to do that by hand.) Once installed, the skills trigger automatically from their descriptions when you ask Claude to proofread or to review a book inside its frame. The inside-frame template and the proofread mechanical sweep also work standalone, as shown above.
scriptorium/
├─ .claude-plugin/plugin.json
├─ docs/ # the public page (GitHub Pages)
├─ tests/collate-fixture/ # planted-error eval for collate (answer key included)
└─ skills/
├─ extract/ # Copyist — SKILL.md, extract_check.py, references/
├─ proofread/ # Corrector — SKILL.md, proofread_check.py, anti-patterns.md
├─ collate/ # Corrector — SKILL.md, collate_check.py, get_transcript.py, references/
├─ inside-frame/ # Illuminator — SKILL.md, template.md
└─ index/ # Illuminator's feeder — SKILL.md, index_verify.py
The demo that remains. Run one source all the way through the producer chain on real, public content: take a Based Camp episode → get_transcript.py → extract → collate its quotes and citations against the article it discusses. One episode, end-to-end.
Named non-goals (YAGNI until it recurs). The set of scribes is complete; these were considered and deliberately shelved: a Quizzer (quiz a text's takeaways — the reader-side idea that index partly absorbs), format-adaptation (book → script → show notes), an armarius orchestrator over multi-skill pipelines, and PDF/OCR ingestion inside extract (convert locally with pdftotext/tesseract first). Any of these earns its way in only by being wanted twice.
The public page (adamisom.github.io/scriptorium, served from docs/) shows every skill in action — what it does, a real invocation, what it refuses. Independent thinkers like the Collinses never ship just one artifact; they repackage the same ideas across a podcast, a Substack, books, and X, each tuned to its audience and platform. A document-tools studio should practice what it preaches: the README is for builders, the site is for everyone else.