What Changed — The Mentor Plugin That Finally Calls You Out
A new open-source Claude Code plugin, dont-let-me by hcsum, turns your agent into a personal mentor that knows your long-term goals and your worst habits. It's been running for months in the author's own workflow, and now it's available as a plugin for Claude Code, Codex, and OpenCode.
The core idea: you spend 10 minutes answering questions about what you're actually working toward and how you get in your own way. The plugin then loads into every session, reads along while you work, and nudges you when you're falling into your own traps.
What It Means For You — Concrete Impact on Daily Claude Code Usage
This isn't a todo list app. It's a behavior alignment system that lives inside your coding agent. Here's what it does differently:
1. It forces honest goal-setting. The /mentor:init setup pushes back if you answer vaguely. You define your actual long-term goals, not generic aspirations.
2. It captures your anti-patterns while you're still honest. The author's examples:
[ship] don't let me build in private forever, waiting for it to be ready[explore] don't let me keep opening new threads while nothing is finished
3. It watches what's not moving. Items that carry over day after day get flagged as avoidance, not scheduling problems. The plugin treats procrastination as a first-class signal.
4. It loads passively. It reads along during unrelated work. When you're doing something on your don't-let-me list, it says so — once per pattern. You can wave it off.
Try It Now — Commands, Config, and Prompt Examples
Installation
claude plugin marketplace add hcsum/dont-let-me
claude plugin install mentor@dont-let-me
Setup (one-time, ~10 minutes)
/mentor:init
This walks you through questions about your goals and anti-patterns. Be honest — the plugin holds you to what you write.
Daily Use
/mentor:check
Run this anytime for an alignment check. It tells you whether what you're doing serves your stated goals.
Natural Language Queries
You can also just talk to it:
mentor, what's on my plate today?
It won't read your whole todo list. It picks the few items that serve your main goals and names the first move on one — small enough to start now.
add "ship the landing page" to my todos
Prompts without the keyword 'todo' or 'mentor' also trigger the mentor stance. You can brain-dump:
I want to redo the pricing page, learn Rust, call the accountant, fix the signup bug, and write that post I keep putting off
The plugin files each item against your goals, assigns priority, and merges duplicates. Items that serve none of your goals get flagged and put low.
Under the Hood: How It Works with CLAUDE.md
Because Claude Code loads the plugin into every session, it effectively appends its stance and your profile to your existing CLAUDE.md context. The plugin uses two files:
stance.md— the mentor's behavior instructionsprofile.md— your goals and anti-patterns
These are loaded alongside any CLAUDE.md instructions you already have. The mentor stance takes priority when it detects you're violating your own rules.
Why This Works — The Psychology of Self-Accountability
The plugin leverages a simple truth: you're most honest about your weaknesses when you're not actively failing at them. By writing down your anti-patterns during setup (when you're calm and reflective), you create a contract your future self can't weasel out of.
The "one nudge per pattern" design is intentional — it avoids nagging while ensuring you can't ignore the signal. And the "what's not moving" detection treats stale items as avoidance rather than scheduling, which reframes the problem from "I'm busy" to "I'm avoiding this."
Caveats
The author notes that while the plugin works on Codex and OpenCode, Claude Code gives the best results because Claude models handle the nuanced nudges better. If you're using a weaker model, the "mentor" may feel robotic or miss context.
What's Next
The plugin is open-source on GitHub (hcsum/dont-let-me). The author plans to add more anti-pattern templates and possibly integrate with project tracking tools via MCP.
Source: github.com
