VLM-Based OCR Pipeline for Scanned Document Corpora Agent Skill

VLM-based OCR pipeline: model selection, prompts, architecture, evaluation.

Context cost
5.8k tokensestimated from the bundle, loaded when it triggers
Bundle
2 files · 23.1 kBtext throughout, nothing executable
Licence
Source-availablethe author’s own terms
Last change
—no release on file
Servers it uses
Noneruns standalone

What it does

VLM-based OCR pipeline: model selection, prompts, architecture, evaluation.

What this skill changes about the agent has not been set out here as a list. Its SKILL.md, shown below, says it in the author’s own words.

data science

The skill itself

A skill is instructions the model reads, so you can see what you are getting: below is the start of its SKILL.md. The front matter loads with every session, the body loads when the skill triggers, and the whole file comes with the install.

SKILL.md14.7 kB · preview
--- name: vlm-ocr-pipeline argument-hint: "[describe your document corpus, languages, and compute environment]" ---
7# VLM-Based OCR Pipeline for Scanned Document Corpora
8
9## Instructions
10
11For a worked language-specific transcription prompt (pre-reform Cyrillic) and a per-page JSON output schema with uncertain_spans, layout_markers, and flags, see reference/prompt-and-schema.md.
12
13### 1. Model Selection
14
15- **Start from OCR benchmarks, not general VLM leaderboards.** OCRBench (Liu et al. 2024) tests across 29 document OCR dimensions; OCRBench v2 (Fu et al. 2025) extends to multilingual scripts and multi-page documents. General vision-language benchmarks (MMMU, VQAv2) do not predict OCR accuracy.
16- **Verify language support explicitly.** Confirm the target language appears in the model's training set with per-language accuracy data. Qwen3-VL (Bai et al. 2025) enumerates 39 languages with OCR benchmarks; models that claim "multilingual support" without language-specific evidence may fail on non-Latin scripts.
17- **Compare across models for your script family.** E-ARMOR (2025) compares five VLMs and two traditional OCR systems across 54 languages on a hand-annotated dataset. Use this or similar comparative studies rather than relying on a single model's self-reported scores.
18- **Assess quantization tradeoffs before committing.** Low-bit quantization (e.g., W3A16, W4A8) enables single-GPU deployment but can degrade VLM accuracy non-uniformly across modalities. Li et al. (2025) show that **language tokens are an order of magnitude more sensitive to quantization than vision tokens**; treating them equally during calibration over-weights the insensitive modality and hurts performance. Always compare quantized output against a full-precision baseline on a pilot sample before committing to a bulk run.
19- **Test the instruct variant against reasoning variants.** For OCR tasks, instruction-following models typically outperform chain-of-thought variants, which may add latency and fabricate content rather than improving transcription fidelity.
20
21### 2. Image Handling and Preprocessing
22
23- **Separate OCR input resolution from archival preservation.** For archival capture, FADGI (2016) prescribes 300-400 ppi across books, journals, and manuscripts (with 400+ ppi for 4-star compliance), and Metamorfoze (van Dormolen 2012) applies similar European standards. For VLM OCR input, resolution requirements are lower but not well-pinned by the KB: calibrate on a pilot, do not drop below ~150-200 ppi, and raise DPI for small type, faded ink, or pre-industrial typography.
24- **Prefer native image extraction over rasterization.** Scanned PDFs store each page as an embedded image at the original scan resolution. Extract these byte-for-byte using PDF library methods rather than re-rasterizing, which downsamples and re-encodes (Pitt OCR Best Practices). Reserve rasterization for VLM input when a specific DPI is needed.
25- **Test preprocessing on a sample before applying corpus-wide.** Learned image restoration combined with neural post-correction can yield large CER reductions on degraded historical documents — Guan et al. (2025) report 63.9-70.3% CER reduction for the full PreP-OCR pipeline (ResShift image restoration + ByT5 post-correction), not for classical deskew/binarization/contrast alone. Classical preprocessing alone may not help and can actively hurt: Machidon & Machidon (2025) find that grayscale conversion, binarization, and dilation did not improve OCR on degraded folkloristic scans and in some cases introduced artifacts that worsened recognition. Always run with and without preprocessing on a diverse sample and compare.
This is a preview of SKILL.md. Download or install the skill to get the whole thing.
In the file
SKILL.md2,004 words
Files2
LicenceSource-available
Why you can read it

Nothing in a skill executes. The client loads the text and the model follows it, so a skill can be audited the way a runbook is — by reading it.

What it costs in context

Skills are not billed by the call. They are paid for in context: every token the instructions occupy is a token your code, your diff and your conversation cannot use. Here is what this one takes and when it takes it.

≈50
always loaded
The name and description, so the model knows the skill exists and when to reach for it.
5,725
on trigger
The instruction body and 1 supporting file, read only when the skill fires.
2.9%
of a 200k window
Ten skills this size would take about 29% of the window before you open a file.
050k100k150k200k context window

5.8k tokens, estimated from the bundle at four bytes to the token, held for the rest of the session once it triggers. Heavy. Teams tend to install this one per project rather than globally, and load it only when the job comes up.

Servers bill, skills cost

A server charges by the month. A skill charges once per session, in context, and then keeps charging it for as long as the session lives.

Before and after

The same question, put to the same model twice: once as it comes, and once with these instructions loaded.

No worked example has been published for this skill yet.

The procedure it runs

The procedure has not been set out here step by step. It is in the body of its SKILL.md: the start of it is shown above, and the whole file comes with the install.

Prose, not code

These steps are written for a model to follow, not executed by a runtime. It can still be told to skip one, and it will say so when it does.

Servers it uses

None. This skill calls no MCP servers at all.

Everything it needs is in the instructions, so it works in a project with nothing connected — the model reads the file and changes how it works with what it can already reach.

Its SKILL.md does not declare which tools it asks the agent for, so what it touches is whatever the instructions themselves ask for — read them before you install it.

What it asks for
Writes filesnot declared
Network accessnot declared

This SKILL.md carries no allowed-tools line, so neither is declared. A skill grants no permissions of its own — it can only ask for tools your client already has.

What it will not do

Every skill is narrow, and the useful ones say where they stop. These are the jobs this one is the wrong tool for.

What this skill is not for has not been published here. Nothing is implied by that: it is a section the author has not filled in.

What is in the bundle

2 files, 23.1 kB on disk. A bundle is text throughout: the instructions the model reads, plus the templates it fills in.

  • SKILL.md14.7 kB
  • reference/prompt-and-schema.md8.4 kB
What is not in it

No dependencies and nothing executable: a skill is text the agent reads, so the bundle is 2 files you can review in full before installing. The Source-available licence covers the templates and examples as well as the instructions.

Install

Installing copies the bundle into your project. Nothing runs at install time — the files sit on disk until the model reads them.

# VLM-Based OCR Pipeline for Scanned Document Corpora · 5.8k tokens when loaded · macOS, Linux mkdir -p .claude/skills && t=$(mktemp -d) && curl -fsSL -o "$t/bundle.tar.gz" https://mcprush.com/api/skills/vlm-based-ocr-pipeline-for-scanned-document-corpo/bundle.tar.gz && tar -xzf "$t/bundle.tar.gz" -C "$t" && [ -d "$t/vlm-based-ocr-pipeline-for-scanned-document-corpo" ] && { [ ! -e .claude/skills/vlm-based-ocr-pipeline-for-scanned-document-corpo ] || { mv .claude/skills/vlm-based-ocr-pipeline-for-scanned-document-corpo "$t/.previous" && echo "Previous copy moved to $t/.previous"; }; } && mv "$t/vlm-based-ocr-pipeline-for-scanned-document-corpo" .claude/skills/ && echo "Installed .claude/skills/vlm-based-ocr-pipeline-for-scanned-document-corpo"
On Windows, or any OS with Node 20+
npx mcprush@latest skill add brycewang-stanford/vlm-based-ocr-pipeline-for-scanned-document-corpo

Windows PowerShell says 'running scripts is disabled on this system'? Run Set-ExecutionPolicy -Scope CurrentUser RemoteSigned once, or type npx.cmd instead of npx (codex.cmd, gemini.cmd, grok.cmd likewise). PowerShell 7 needs neither. Writes to .claude/skills/vlm-based-ocr-pipeline-for-scanned-document-corpo/ in the current project. For every project instead, run mkdir -p ~/.claude/skills && t=$(mktemp -d) && curl -fsSL -o "$t/bundle.tar.gz" https://mcprush.com/api/skills/vlm-based-ocr-pipeline-for-scanned-document-corpo/bundle.tar.gz && tar -xzf "$t/bundle.tar.gz" -C "$t" && [ -d "$t/vlm-based-ocr-pipeline-for-scanned-document-corpo" ] && { [ ! -e ~/.claude/skills/vlm-based-ocr-pipeline-for-scanned-document-corpo ] || { mv ~/.claude/skills/vlm-based-ocr-pipeline-for-scanned-document-corpo "$t/.previous" && echo "Previous copy moved to $t/.previous"; }; } && mv "$t/vlm-based-ocr-pipeline-for-scanned-document-corpo" ~/.claude/skills/ && echo "Installed ~/.claude/skills/vlm-based-ocr-pipeline-for-scanned-document-corpo" (macOS, Linux) — or add --global to the npx line. The curl line replaces .claude/skills/vlm-based-ocr-pipeline-for-scanned-document-corpo/ only after the download succeeds, so files left there by another skill of the same name, or by an older release, do not mix in. A copy already there is moved aside, not deleted, and the line prints where. It appears as vlm-ocr-pipeline in Gemini CLI, Codex, Copilot and Grok (its own name); Claude Code lists it by the folder name, vlm-based-ocr-pipeline-for-scanned-document-corpo.

Which clients pick it up on their own

A skill is a folder of text. Most of these read a skills folder and pick it up without being told; the rest take the same folder another way — an upload in the app, a parameter on an API call — and then choose it by its description just the same.

Claude Code.claude/skills/
Claude DesktopCustomize › Skills
ChatGPTSkills › Create › Upload
Cursor.cursor/skills/
VS Code.github/skills/
Codex CLI.agents/skills/
Gemini CLI.gemini/skills/
Grok Build.grok/skills/
Microsoft CopilotSkills › Upload a skill
PerplexityComputer › Skills › Upload a skill
DeepSeek Harness.dsh/skills/
Zed.agents/skills/
Devin Desktop.devin/skills/
Windsurf (legacy).windsurf/skills/
OpenAI Agents SDKInstructions or sandbox Skills
HTTP / APIa request parameter
This release
Versionnot versioned
Publishedno release date on file
PriceFree
Referencebrycewang-stanford/vlm-based-ocr-pipeline-for-scanned-document-corpo

Versions

Its author publishes no version number, so there is nothing here to pin to: what you install is the folder as it stands today. Instructions change more often than APIs do — a skill can be rewritten entirely without anything it depends on moving.

Pinning

Nothing to pin to: this skill carries no version number of its own. What you install is what the folder holds on the day you install it.

Reviews

—no reviews yet

Nobody has reviewed this skill yet. The rating is the mean of the reviews written here, so there is none until somebody writes the first.

Sign in to review it

A review is tied to an account so it can be answered, edited and taken down by the person who wrote it.

Sign in
Who can post

Anybody with an account, one review each. It carries the version that was current when you wrote it, and the publisher may reply once.

Publisher
Servers1
Claim this skill