FREE TOOL · EDITORIAL REVIEW

Humanizador

Keep your voice, nuance and decisions when working with AI.

Read the original PDF, mark what you want to change and assess suggestions in one workspace. Export the review with precise references before returning to your manuscript.

User guide (Spanish)Source on GitHub
Version 0.6MIT · Open sourceRules + optional OllamaSpanish from Spain
PDF with personal notes and editorial suggestions beside a compact decision table
Real Humanizador review of IAHPT, shown with permission from its author, Roberto Canales Mora. The full book and review data are not distributed.

WHY IT EXISTS

When composition keeps the words but loses the intention

A class or a spoken explanation contains examples, doubts, conditions and a distinctive way of telling things. Turning it into chapters with AI can reduce a scene to a moral or turn personal experience into a general claim. Editing needs to recover what the author meant.

Humanizador helps organise that reading and review. It flags candidates, stores your notes and lets you decide on each proposal. Its purpose is responsible editing and a final manual review, not plagiarism, deception or evading AI detectors.

The document and its review, together

Read and annotate

Select text or mark an area of an infographic. Notes retain the page, position, context and image crops so you can return to the exact location.

Assess and decide

Notes and suggestions use different colours and two tabs. A compact table lets you edit a proposal, approve it, keep the original or reopen a decision.

Continue editing

One JSON export brings together notes, proposals, decisions and references. You can give it to another AI to prepare manuscript changes for your review.

Collapse the menu, resize or swap the panels, and change pages within the viewer. The original is immutable: approving a proposal stores your decision; it does not rewrite the PDF.

A catalogue with an explicit scope

30

Checks

Deterministic patterns that work without a model.

12

Contextual criteria

Optional reading with Ollama and proposal validation.

38

Pending specifications

Documented for future implementation or validation; not executable coverage.

The 80 entries include examples, exceptions and sources. Fidelity principles aim to preserve the subject, complete scenes, stance, humour, causal relationships, conditions and certainty. Context matters: a technical term or a deliberate pause may be appropriate.

A finding is an invitation to reread. False alarms and omissions are possible; general accuracy has not been measured on a representative corpus. Counting findings does not establish quality or authorship.

The catalogue does not learn automatically when changes are approved. Contributions become criteria with synthetic examples and counterexamples, tested before activation. Contributing editorial knowledge (Spanish).

USER GUIDE

From your PDF to a review you can continue

  1. Install and start. Download and extract the repository; run sh instalar.sh on Mac/Linux or .\instalar.ps1 in PowerShell. The wizard checks Node and prepares the application. The browser opens; an alternative port is used if needed.
  2. Create a project and upload a PDF. Start reading without running a model. Analysis requires selectable text; OCR is not included.
  3. Add your notes. Select text and use the note button or right-click. Mark an area for a figure.
  4. Request a review. Choose rules or Ollama if installed. Follow progress within the document.
  5. Decide in the table. Read the context, expand details with “+”, edit the proposal and approve or keep the original. You make the final decision.
  6. Export and return to the manuscript. The JSON includes references, crops and decision states. You decide whether to share it with an AI, then manually review the changes.
Page controls within the PDF viewer
Move between pages while reading, without returning to the header.
User guide (Spanish)Source on GitHub

Guided installation on Mac, Windows and Linux

Download the project ZIP, extract it and open a terminal in that folder. Git is another option. Node.js 24 or later with npm is required; the wizard explains how to install it if missing. The wizard and full manual are in Spanish.

Mac and Linux

In Terminal, from the project folder.

sh instalar.sh

For daily use

Start
sh arrancar.sh
Stop
sh parar.sh

Windows

In PowerShell, from the project folder.

.\instalar.ps1

For daily use

Start
.\arrancar.ps1
Stop
.\parar.ps1

The wizard shows progress, installs dependencies and asks how you want to review. Keep your settings, use rules alone or set up Ollama. Model downloads require confirmation and documents are preserved. The installation manual also explains what to do when PowerShell blocks scripts.

Ollama is optional

On Mac, Ollama requires macOS 14 or later: Apple Silicon can use the GPU, while Intel uses the CPU. On Windows, it requires Windows 10 22H2 or later; acceleration depends on the GPU and its drivers. Install and open Ollama, then select a model in the wizard. The application distinguishes an unavailable service from a missing model and provides setup help.

You can try qwen3.5:4b with fewer resources, without claims about its editorial quality. Models need several GB of disk space and additional memory for context. The historical qwen3.6:27b-q8_0 model is much larger and is not downloaded by default. Ollama setup, models and activity.

The interface uses React, Vite and PDF.js; the server uses Node and SQLite. Stopping pauses work so you can resume later. Data stays on the computer or server running the application. GitHub Pages hosts the manual; it does not process PDFs.

Humanizador was built with Codex, under the direction of Roberto Canales Mora. It is checked through automated tests and human review. Codex is a development tool: you do not need an account to use Humanizador. Recommendations still require your judgement.

Before you use it

Proposals need your review

Quote and fidelity checks reduce mechanical errors but do not guarantee correct interpretation. The model works on fragments; relationships between chapters and final editing require human reading.

The PDF is not rewritten or typeset

Humanizador stores references and decisions. The export supports manuscript editing. Scans, columns, tables and split words need particular care.

The code is free; your documents remain yours

Code and original documentation use MIT. Dependencies retain their licences. IAHPT excerpts shown in screenshots retain their copyright and appear with permission. Do not post private exports or documents in issues.

Report a problem with a synthetic example · Full illustrated guide (Spanish) · All projects