AI gave us the greatest answer engine ever built.
But there is one brutal limit:
It cannot answer a question you never ask.
If you do not know a concept exists, you cannot ask about it. The blind spot, better question, different perspective, or next opportunity stays invisible.
Scout reveals it.
npx @gomzkov/scout
How it works
Scout is a behavior layer for Claude Code, Cursor, and Codex. Your agent gives you the full answer first. When Scout finds a high-value gap, it adds one short pointer. It adds a second only when both materially matter:
- Risk: the hidden constraint, failure mode, or assumption
- Ask: the next question that unlocks a better answer
- Lens: the perspective that changes the picture
- Direction: the work that matters more
- Opportunity: the feature, product, tool, or idea sitting nearby
No gap, no block. Scout stays silent until it has something worth saying.
The goal: 10x the value of the conversation without 10x the noise.
Examples
Your agent answers the request. Scout adds the gap:
Software
You: Help me design an expense approval workflow.
Scout:
- Risk: The happy path breaks when the approver is unavailable or submits their own expense. Define delegation, reassignment, and self-approval rules now.
Product
You: How should we improve onboarding for our analytics product?
Scout:
- Direction: Optimize for the first useful decision, not completed setup. Measure time to first insight before adding more onboarding steps.
Creative
You: Help me outline a documentary about abandoned shopping malls.
Scout:
- Lens: The buildings are the setting, not the story. Follow one person whose life changed with the mall and let the decline unfold through them.
Scout discovered Scout
Scout started as a private skill I built for myself. Then it surfaced an adjacent opportunity I had not asked about: turn the behavior into a standalone product so other people could use it.
That suggestion created this repository. Scout found the product hiding next to the work.
Where it helps
- Developers: catch logic holes, missing states, security gaps, and failure modes
- Founders and product teams: find the user need behind the feature and the product behind the workaround
- Researchers: uncover missing vocabulary, assumptions, sources, and opposing models
- Creative people: discover stronger angles, formats, constraints, and adjacent ideas
- Decision makers: expose hidden tradeoffs, second-order effects, and better options
Install
Run the interactive installer with Node.js 22 or newer:
npx @gomzkov/scout
Choose:
- Claude Code, Cursor, Codex, or all three
- A global installation or the current project
- Always-on or on-demand mode when installing for Claude Code
Running the installer again updates Scout without duplicating configuration. Existing settings and instruction files are preserved.
Non-interactive install
npx @gomzkov/scout install --agent all --scope global --mode on
Available values:
| Option | Values |
|---|---|
--agent |
claude, cursor, codex, all |
--scope |
global, project |
--mode |
on, demand for Claude Code or all |
Agent behavior
| Agent | Global | Project |
|---|---|---|
| Claude Code | Skill plus optional always-on hook | Project skill plus optional always-on hook |
| Cursor | Personal skill, loaded by matching or /scout |
Skill plus always-applied project rule |
| Codex | Skill plus a managed ~/.codex/AGENTS.md block |
Skill plus a managed project AGENTS.md block |
Start a new agent session after installing.
Uninstall
npx @gomzkov/scout uninstall
For an unattended uninstall:
npx @gomzkov/scout uninstall --agent all --scope global --yes
The uninstaller removes only Scout's files, its Claude Code hook, and its marked Codex instruction block. Other hooks, settings, skills, rules, and AGENTS.md content stay untouched.
Restart any open agent sessions after uninstalling.
Configuration
| Variable | Default | Purpose |
|---|---|---|
SCOUT_DISABLE |
0 |
Set to 1 to mute the Claude Code always-on hook without uninstalling |
Installer safety
- Zero runtime dependencies.
- No install or postinstall script. Nothing changes until you run the CLI.
- Writes are atomic and refuse symlinked targets.
- Install and uninstall are idempotent and preserve unrelated configuration.
- The package tests global and project flows for all supported agents.
High signal by design
- Answer first. Scout never replaces or buries the real answer.
- One line by default. A second appears only when it materially changes the outcome.
- No manufactured insight. Nothing worth saying means no Scout block.
- Calibrated to you. Scout skips what you know and searches beyond your current frame.
Edit the installed SKILL.md to tune the pointer budget, skip rules, and activity level.
Development
npm ci
npm test
npm pack --dry-run
npm test compiles the installer and exercises install, update, and uninstall flows inside a temporary directory.
No comments yet
Be the first to share your take.