Gym mode for agentic coding

Instead of writing 100% of the code for you, the agent scaffolds a logical unit, hands it off, watches you implement it, and reviews your work before resuming.

_{___ Keep your edge ___}

---

Heavy AI usage makes you stupid.

Science. says. so.

Anthropic too.

When BIG BAD AI COMPANY™ warns us about the negative effects of its own product, we should probably pay attention.

Much like sitting on your ass all day makes you weak and sad, keeping your brain in powersave mode all day makes you lazy and dumb.

The first you fix by going to the gym.

The second you fix by using SpotMe.

---

How it works

1. Enable SpotMe at the start of a session: /spotme:on [lite|medium|hard] [--every N]
2. Every N code-writing actions, the agent scaffolds the next unit instead of completing it
3. You implement the marked section (# SPOTME: ...) directly in your editor
4. /spotme:done → agent checks your work and gives brief, calibrated feedback
5. Agent resumes the original task

Commands

Command	Description
/spotme:on [lite|medium|hard] [--every N]	Enable gym mode. Default: medium, every 2
/spotme:off	Disable — agent writes code normally
/spotme:status	Show current state
/spotme:rep	Request an exercise on-demand
/spotme:done	Submit your implementation for review
/spotme:hint	Get one targeted hint
/spotme:solve	Concede — agent completes the exercise
/spotme:skip	Skip this exercise, no note

Difficulty levels

Level	Agent writes	You write
lite	Signature + docstring + structure	Just the body
medium	Signature + # SPOTME: spec comment	All logic
hard	Plain English spec comment only	Everything

Install

Claude Code

SpotMe ships with a small installer CLI that generates a local Claude plugin marketplace and registers it. Install the CLI with bun:

bun add -g github:bge-kernel-panic/spotme

This puts a spotme binary in ~/.bun/bin. Make sure that directory is on your PATH.

Then register the Claude plugin:

spotme install claude

This will:

1. Generate the marketplace under ~/.spotme/claude-marketplace/
2. Run claude plugin marketplace add and claude plugin install spotme for you (or print the commands if claude is not on your PATH)

Restart Claude (/reload-plugin is not enough, as it won't install the pre-tool hooks)

The MCP server starts automatically, the PreToolUse hook intercepts Write/Edit/MultiEdit to count code writes, and the /spotme:* commands become available.

Useful flags:

• --yes — skip confirmation prompts
• --manual — write the plugin files but don't run any claude CLI commands (prints them instead)

OpenCode

Add to your opencode.json:

{
    "$schema": "https://opencode.ai/config.json",
    "plugin": ["spotme"]
}

Or use the installer CLI to do it safely (preserves comments in .jsonc):

spotme install opencode               # writes to ~/.config/opencode/opencode.json
spotme install opencode --scope project  # writes to ./opencode.json

Pi

Install as a Pi package:

pi install git:github.com/bge-kernel-panic/spotme

SpotMe integrates as a Pi extension with programmatic slash commands and custom tools (spotme_exercise, spotme_end, spotme_status). Commands that don't require LLM reasoning (/spotme:on, /spotme:off, /spotme:status) execute instantly via the Pi UI. Commands that need LLM input (/spotme:done, /spotme:solve, /spotme:skip, /spotme:hint, /spotme:rep) inject exercise details directly into the prompt for seamless review.

Skill only (any harness that supports AgentSkills)

Copy SKILL.md into your harness's skills directory. This gives the prompt layer without the automated tool interception — commands still work, but the counter-based trigger won't fire automatically.

Local Development

Build

git clone https://github.com/bge-kernel-panic/spotme.git
cd spotme
bun install
bun run build      # builds both bun targets (OpenCode/Pi) and node targets (CLI/MCP)

Other useful scripts:

Script	Purpose
bun run build	Build everything (bun + node targets)
bun run build:node	Build only node targets (dist/cli.js, dist/claude-mcp.js) — what the Claude installer needs
bun test	Run the test suite (engine, prompts, installer)
bun run typecheck	TypeScript check (no emit)
bun run lint	ESLint

The node-target bundles (dist/cli.js, dist/claude-mcp.js) are committed to git so bun add -g github:... installs work without a build step. Rebuild and re-commit those whenever you change the CLI, the MCP server, the installer, or the prompts.

Test locally on OpenCode/Pi

# 1. Clone SpotMe locally
mkdir ~/temp && cd ~/temp
git clone https://github.com/bge-kernel-panic/spotme.git

# 2. Create a new directory somewhere
mkdir ./test_spotme && cd ./test_spotme

# 3. Run scripts/test-local-branch.sh from the new folder
../spotme/scripts/test-local-branch.sh

Then, in test_spotme, open your agent harness (OpenCode, Pi...) and verify spotme commands exist. Finally, checkout the SpotMe branch you need.

Test locally on Claude Code

Install your local build directly with bun link, or push to a branch and pull from GitHub:

# From the spotme repo:
bun run build:node
git add -A && git commit -m "..." && git push

# On the test machine:
bun remove -g spotme            # if previously installed
rm -rf ~/.bun/install/cache     # clear bun's git cache
bun add -g github:bge-kernel-panic/spotme
rm -rf ~/.spotme/claude-marketplace   # clear stale plugin artifacts
spotme install claude --yes
# Then /reload-plugins in Claude Code

Name

The agent is your spotter. It sets up the lift, stands by while you push, catches you if you call for help. The work is yours.