Single Portable Markdown Skill File · Zero Config Required

Your agent finished the task.
You still can't explain how it works.

idontknow is an autonomous engineering skill for Claude Code, Cursor, Copilot CLI, Windsurf, and modern agents. It builds real, production-ready code directly in your repo while stopping to teach every meaningful decision along the way. Never finish a task you can't defend.

100%
Portable Markdown. Runs anywhere instructions are read.
0
Zero external dependencies or npm packages to install.
1:1
Implementation and learning happen in the exact same step.
.ledger
Compounds knowledge across commits in your repository.

What happens when your agent builds without teaching

Modern AI coding assistants are extraordinarily fast. That speed creates an invisible vulnerability: features ship into production before anyone builds a mental model of how they work.

Standard AI Autopilot (The Trap)
  • Prompt for unfamiliar tech: You ask for a Kafka consumer, OAuth flow, or CSS subgrid layout you've never authored before.
  • It works, so it gets merged: The tests pass, the PR is closed in minutes. Nobody verifies if you could explain a single line in a design review.
  • Breaks three weeks later: A production edge case triggers at 2:00 AM. The code was architected based on silent AI assumptions you never evaluated.
  • Crippling Imposter Debt: You find yourself frantically debugging a codebase you supposedly own but can't mentally trace.
With idontknow (The Senior Co-Pilot)
  • Map terrain & existing repo idioms: It inspects your actual code and package conventions before generating a single line.
  • Mental model first, syntax second: Explains the architecture and data lifecycle before touching sensitive files.
  • Full Explanation Contract: Every decision explains: What it is, why it exists, real alternatives, wider industry practice, and failure modes.
  • Compounding Ledger: Logs what was demonstrated in .learning-ledger.md so the next task starts further ahead.

One skill file. Three uncompromised commitments.

idontknow doesn't ask you what you know before it decides how much to teach. It builds real code and breaks down every key trade-off, every time, regardless of how confident you sound.

01 / REAL_CODE

Build for Real

In your repository, inside your established patterns, using what already works instead of quietly replacing dependencies. Nothing is thrown away after the explanation: the lesson and the production implementation are the exact same act.

02 / LEAN_QUESTIONS

Ask Only What It Must

It exhaustively scans existing manifests, configs, and directory structures before asking you anything. You are only consulted on ambiguous requirements code cannot answer, and every question comes with an opinionated recommendation.

03 / FULL_CONTRACT

Teach Every Decision

What it is, why it exists, how it works underneath, real trade-offs considered, standard industry practice vs. this repo's style, and realistic failure modes. Depth only scales back once your repository ledger proves you've retained it.

Six stages, and the ledger powers the next task

This is the exact sequence the skill executes when invoked on a real task. It ends by updating your repo's learning ledger, leaving you one loop further ahead every time.

1
Map the Terrain
Inspects stack, conventions & past ledger history.
2
Ask What Code Can't Tell
Only asks business constraints with recommendations.
3
Confirm Risky Choices
Stops for architecture, destructive, or prod changes.
4
Build in Your Repo
Direct production code matching existing style.
5
Teach Every Decision
Mental models, trade-offs & common failure modes.
6
Log What Stuck
Records verified retention in .learning-ledger.md.
Next task loop: Reads the ledger first. Concepts proven retained skip repetitive basics and immediately advance to deeper architecture.
[Stage 3 is the only manual gate]

Why I Created idontknow

A personal reflection on the double-edged sword of AI-assisted engineering and how we take back technical ownership.

J

Built by Jeethu

Creator of idontknow · Engineering Note

Over the past two years, AI coding assistants completely transformed how fast software gets written. We can prompt an agent, watch 500 lines of complex boilerplate materialize in seconds, and declare the ticket finished.

But I noticed a deeply uncomfortable pattern: the faster the agent wrote, the less I genuinely understood what was living in my repository. I was shipping database migrations, reactive state machines, and concurrency pipelines that I couldn't explain from first principles. If an interviewer asked, "Why did you use this exact concurrency primitive here?", the honest answer was, "Because Claude Code put it there."

"AI should not be an autopilot that turns engineers into helpless consumers. It should be a tireless staff mentor that levels up your capability with every single commit."

I built idontknow to resolve this contradiction. I didn't want a tool that just generated toy tutorials or lectured me in the abstract. I wanted a skill that would do the real work in my real repo, but refuse to let me sign off on code I cannot defend.

With idontknow, whenever I venture into an unfamiliar corner of computer science—whether that's low-level CSS layout algorithms, Web Workers, OAuth security handshakes, or Kubernetes manifests—the agent builds the solution cleanly, but stops to systematically transfer the mental model into my head.

The result is that every completed pull request doesn't just advance the codebase: it permanently advances my skills as a software engineer.

One universal skill. Pick your AI agent.

The skill content remains identical. Only the directory path and command you invoke change based on your preferred tool.

Claude Code Native Skill

Drop into your local repository or home directory for global availability.

.claude/skills/idontknow/SKILL.md
mkdir -p .claude/skills/idontknow && curl -sSL https://raw.githubusercontent.com/.../SKILL.md > .claude/skills/idontknow/SKILL.md

How to invoke: In Claude Code, ask explicitly by name: "use idontknow to implement the responsive navigation with container queries."

Cursor & Windsurf Project Rules

Place inside your workspace rules or prompt instructions folder.

.cursor/rules/idontknow.mdc or .windsurfrules
mkdir -p .cursor/rules && cp SKILL.md .cursor/rules/idontknow.mdc

How to invoke: Tag the rule in Composer or Chat: "@idontknow build the OAuth token refresh handler and teach every decision."

GitHub Copilot CLI & Agent Mode

Supply as project instruction context or terminal runner prompt.

.github/copilot-instructions.md
cat SKILL.md >> .github/copilot-instructions.md

How to invoke: Instruct Copilot: "Follow idontknow instructions to build the feature and explain failure modes."

Google Antigravity (AGY) Custom Skill

Install into workspace or global skills directory.

.agents/skills/idontknow/SKILL.md
mkdir -p .agents/skills/idontknow && cp SKILL.md .agents/skills/idontknow/SKILL.md

How to invoke: Prompt Antigravity: "Activate idontknow skill and implement this task."

Any Markdown-Compatible LLM

Works with Aider, Cline, Continue.dev, or raw system prompts.

Plain Markdown Context

There is zero proprietary syntax in the file. Copy the raw text below and paste it directly into your system prompt or session context.

Copy the complete skill file

The complete, uncompressed markdown file. Nothing hidden or omitted.

SKILL.md (Ready to paste)

    

Six simple steps to get it running

No complex installation scripts or package dependencies. You are minutes away from your first guided build.

01

Copy or download the SKILL.md file

Use the copy button above to place it in your clipboard or download it directly to your machine.

02

Place it where your AI agent expects skills

For Claude Code, put it in .claude/skills/idontknow/SKILL.md. For Cursor, put it in your rules directory.

03

Invoke it by name on a real task

Prompt your agent: "use idontknow to build [your feature]". It stays quiet until specifically requested.

04

Confirm high-stakes architectural gates

It performs routine code edits automatically, only pausing to confirm destructive actions, security decisions, or architecture.

05

Absorb the explanation contract

Read the trade-offs, alternative approaches, and failure modes taught alongside the code diff in the response.

06

Watch your repository ledger compound

Check .learning-ledger.md in your repo. Your agent remembers what you've learned for all future tasks.

Frequently Asked Questions

Everything you need to know about integrating idontknow into your team and workflow.

Only by the seconds it takes to read the explanation alongside the diff. It does not stop for routine code edits. By catching architectural flaws early and explaining failure modes before you ship to production, it actually cuts overall debugging and rework time significantly.

The skill checks your repository's .learning-ledger.md. Once you have demonstrated retention on a concept, future explanations are automatically trimmed to brief references so you never get bogged down by redundant explanations.

Yes! Because .learning-ledger.md is committed directly to git, teammates can see what patterns and conventions have been taught and demonstrated across the codebase.

No. idontknow is 100% pure markdown instructions executed directly by your local AI agent. No telemetry, no third-party servers, and no external tracking.

Copied to clipboard