Glasses

Put your glasses on.
The nerd who asks for evidence before your AI writes expensive code.

AI coding agents are good at producing code. They are also good at inventing
abstractions, dependencies, and future requirements nobody asked for.

Glasses makes the agent stop and answer:

What evidence requires this complexity now?

No evidence means no extra machinery. Correctness, security, accessibility,
data integrity, validation, observability, and useful tests are never removed.

Choose the nerd you need

You need to…	Use	What happens
Create or simplify a plan	`/glasses:glasses`	Challenges unsupported complexity and produces an executable plan.
Find and fix a code problem	`/glasses:pointer`	Diagnoses once, recommends a correction, and asks before editing.
Research a technical decision	`/glasses:homework`	Creates a research plan, brings sources, and asks before running validation.
Capture notes or project docs	`/glasses:scribe`	Takes approved notes, maps project docs, and preserves visible memory.

They can hand work to each other when needed without repeatedly bouncing the
same decision around.

Install in Claude Code

Run these commands in your terminal:

claude plugin marketplace add DanielGouveiah/glasses
claude plugin install glasses@daniskills

Restart Claude Code. That is all.

Glasses now activates automatically when Claude creates a plan, ADR, design, or
architecture proposal. Its default mode is high.

Use it

Review a plan

/glasses:glasses

You can choose a mode:

/glasses:glasses medium
/glasses:glasses high
/glasses:glasses ultra
/glasses:glasses off

Mode	Meaning
`medium`	Remove obvious excess, but allow named near-term flexibility.
`high`	Keep complexity only when current evidence requires it. Default.
`ultra`	Require evidence for every new component and prefer native solutions.
`off`	Disable automatic Glasses review for the session.

Diagnose existing code

/glasses:pointer the failing checkout test

Pointer:

Inspects the relevant code and evidence.
Asks one numbered question at a time if needed.
Shows the problem, options, and recommendation.
Asks permission before changing code.
Applies the accepted correction once, verifies it, and stops.

Research a decision

/glasses:homework best local model for structured extraction on 6GB VRAM

Homework:

Uses your conversation and project as context.
Keeps an explicitly general question general; project context applies only
when you ask it to.
Always asks a numbered Q1 to confirm objective, scope, and exclusions.
Shows a bounded research plan and waits for approval.
Researches reliable sources and returns a recommendation.
Suggests the cheapest bounded validation plan.
After separate approval, runs its named experiments until the budget or
stopping condition is reached.

One validation plan may include several stages or matrix cells. Homework does
not ask again for each approved cell, but it must ask before expanding the
scope or budget.

Capture notes and documentation

/glasses:scribe document what we decided
/glasses:scribe map this project

Scribe:

Reads existing notes and docs before writing.
Takes session notes when explicitly requested.
For broad project notes, first proposes the files to create or update.
Writes only after the user requested or approved that specific note work.
Records visible project memory only; it does not manage hidden AI memory.

Available depths are skim, study (default), and thesis.

What Glasses changes

A Glasses plan clearly states:

required behavior and acceptance tests;
the simplest design supported by evidence;
complexity removed and retained;
decisions answered during review;
triggers that would justify more machinery later.

Implementation then stays inside that plan. If new evidence changes the design,
Glasses revises the plan explicitly instead of silently expanding the code.

Other AI coding agents

The repository includes adapters for Gemini CLI, Codex, GitHub Copilot CLI, and
OpenCode. Claude Code currently has the most complete installation flow.

Install the Gemini extension with:

gemini extensions install https://github.com/DanielGouveiah/glasses

The portable policies live in:

CLI harness support

Harness	Auto context	Manual route	Stepped questions	Approval gates	Scribe notes
Claude Code plugin	Session and subagent hooks	Skill routes	Yes, `Q1`, `Q2`	Yes	Yes, approval-only
GitHub Copilot plugin	Session and subagent hooks	Skill routes	Yes, `Q1`, `Q2`	Yes	Yes, approval-only
Codex plugin	Skills directory	Skill routes	Yes, `Q1`, `Q2`	Yes	Yes, approval-only
Gemini CLI extension	Static Glasses context	`/glasses:*` commands	Yes, `Q1`, `Q2`	Yes	Yes, approval-only
OpenCode adapter	Chat system transform	Adapter policy	Yes, `Q1`, `Q2`	Yes	Yes, approval-only
Benchmark harness	Prompt policy by arm	None	Scripted clarifications only	Not interactive	Not measured

The shared routing policy is generated from
skills/glasses/scripts/workflow-routing.mjs for JavaScript harnesses. Static
harnesses keep the same routing rules in the Glasses skill and command files.

Configuration

To change the automatic mode, create ~/.config/glasses/config.json:

{
  "defaultMode": "high"
}

Manual /glasses:glasses use defaults to high, even if configuration says
otherwise.

Benchmark

The completed pilot contains 32 isolated Codex sessions across eight tasks and
four arms: baseline, Ponytail, Glasses High, and Ponytail with Glasses High.
Ponytail is a comparison arm, not a dependency.

This pilot is not yet an efficacy claim. Independent acceptance testing and
blinded complexity review are still required. See the
benchmark design and scoring rules.

Develop locally

git clone https://github.com/DanielGouveiah/glasses.git
cd glasses
npm test
npm run validate
claude --plugin-dir ./skills/glasses

License

MIT