Dyslex.ai
Dyslex.ai is a cognitive accessibility skills plugin for AI agents.
It helps agents reduce friction across the human-AI interaction loop: expressing intent, reconstructing meaning from imperfect wording, preserving technical detail, managing cognitive load, and presenting the next action clearly.
Dyslex.ai does not diagnose disability, provide medical advice, or claim medical benefit.
Quickstart
How It Works
Dyslex.ai is a skills-only plugin with an evidence-aware cognitive model.
The plugin contributes reusable Agent Skills through:
.codex-plugin/plugin.json
skills/
src/cognitive/
When a matching task appears, the agent can use the relevant skill to reduce a specific functional friction. The project maps evidence-informed mechanisms to interaction risks, then to concrete agent behavior.
scientific knowledge
-> functional model
-> interaction risk
-> adaptation strategy
-> skill behavior
The goal is not to replace the agent's reasoning or Superpowers workflow. The goal is to make that workflow easier to follow, safer to resume, and harder to misread.
Every skill starts with a Superpowers gate. If Superpowers is missing or uncertain, the agent must stop and ask the user to install Superpowers before applying Dyslex.ai.
Installation
Codex
Primary path, after public availability:
Search for:
CLI install, when available from the configured marketplace:
codex plugin add dyslex-ai@openai-curated
Fallback if Dyslex.ai is not visible in the public plugin directory yet:
codex plugin add dyslex-ai@npm:@firou91/dyslex.ai
Start a new Codex session after installation so the skills are loaded.
Local Package
The npm package is published as:
npm install @firou91/dyslex.ai
The CLI command is:
dyslexai doctor --verbose
The Basic Workflow
- Install Superpowers in the target agent.
- Install Dyslex.ai as a plugin.
- Start a new session so the skills are visible.
- Ask normally, including messy text, unclear instructions, raw errors, or long output.
- Dyslex.ai detects the relevant accessibility skill.
- The skill preserves commands, paths, package names, environment variables, and other technical tokens.
- The agent clarifies only material ambiguity before acting.
- The response is structured for readability and verification.
What's Inside
Dyslex.ai currently includes 16 skills, plus a shared cognitive model for evidence, functional dimensions, and adaptation rules.
Readability And Writing
accessible-response: reshape answers into clearer blocks without losing detail.spelling-and-writing-support: correct, proofread, or reformulate text while preserving tone.typo-tolerant-intent: infer likely intent from typos or malformed wording.identifier-safety: protect paths, commands, URLs, package names, variables, hashes, and regexes.
Debugging
error-decoder: turn raw errors, stack traces, and logs into clear explanations.terminal-output-reader: summarize noisy terminal output without hiding failures.diff-reader: explain changes, risks, and test needs from diffs or patches.code-explainer: explain source code in accessible layers.
Collaboration
instruction-decoder: unpack dense prompts, issues, specs, or task descriptions.prompt-clarifier: ask only the questions that materially change the result.task-sequencer: break work into ordered steps with dependencies and checks.decision-comparator: compare options and provide a clear recommendation when needed.number-and-date-clarifier: clarify ambiguous dates, units, versions, decimals, and ranges.
context-reentry: resume interrupted work after pauses, compaction, or task switching.documentation-navigator: identify which docs matter and what is required versus recommended.accessible-superpowers: present Superpowers workflows more accessibly without weakening them.
Philosophy
Dyslex.ai is built around a simple rule: accessibility should reduce cognitive friction without reducing technical precision or user autonomy.
That means:
- adapt to functional needs, not diagnostic labels;
- preserve exact technical tokens;
- separate facts from assumptions;
- ask fewer, better questions;
- make success criteria explicit;
- keep warnings and failures visible;
- avoid medical claims;
- avoid pseudo-neuroscience;
- treat unsupported ideas as hypotheses or preferences;
- respect Superpowers as the workflow authority.
The project is not trying to make agents softer or less rigorous. It is trying to make rigorous work easier to read, resume, and verify.
Scientific Model
The first research and architecture report is available in DYSLEX.AI — Scientific Interaction Model.
The model avoids diagnosis-based modes. It uses functional dimensions such as:
- working-memory load;
- spelling variability;
- lexical-access uncertainty;
- reading-speed load;
- numerical ambiguity;
- sequencing and planning load;
- interface and tool-action burden;
- over-adaptation risk.
Identity
Here is how we connect the very minimal visual identity to Dyslex.ai's role, name, and philosophy.
Read the full visual identity notes in Identity.
Support
Publishing and maintaining Dyslex.ai across public agent ecosystems costs time and API tokens. Making an accessibility plugin publicly available usually requires repeated review passes, test runs, model-specific validation, and ongoing compatibility checks.
Codex is already covered as the first public deployment target.
Community support will guide the next deployment targets:
- 1,000 GitHub stars: Public deployment for Claude Code.
- 5,000 GitHub stars: Public release of the plugin for Cursor and other planned hosts.
You can support the project by:
- starring the repository;
- testing Dyslex.ai in real workflows;
- reporting unclear behavior;
- suggesting accessibility improvements;
- sharing reproducible prompts where the skills help or fail.
Documentation