mrkeyoor.com_
Tue 01 Sept 17:40 UTC
AI Toolsevaluationupdated 25 Aug 2026

learn review

Learn is one person's configuration for turning the pi coding agent into an interactive tutor. Its teaching skill asks questions, quizzes the learner, records a markdown session, and can delegate fact checking or diagram work to sub-agents, but the author explicitly shares it as a personal system rather than a finished education product.

+774stars / 7d
Verdict

Our install attempt failed in 5 seconds inside extensions/visual-tools/, so Learn is best treated as a readable personal configuration for current pi users, not an install-and-forget tutor. Its questions, quizzes, markdown log, researcher, and diagram agents form a thoughtful starting point if you are ready to edit the teaching skill and tmux-based wiring. Do not adopt it as a team dependency until a license, reproducible install, and automated tests exist.

We ran it

Lab card: what happened when we ran learnScreenshot of learn (github.com/amosblomqvist/learn)
Install✗ · 5s
Build
Repo18 files~2,792 lines of source · 0.3 MB · 0 CI workflows

Answers from our run

Does learn build from source?

The dependency install failed, and the project has no separate build step. We cloned commit 73eaf7c into a clean Debian container with 3 CPUs and no project-specific setup.

Who should not use learn?

Anyone expecting a standalone tutor: the repository is a .pi directory and requires the separate pi agent.

What are the alternatives to learn?

Khoj, STORM, ChatGPT Study Mode. Our install attempt failed in 5 seconds inside extensions/visual-tools/, so Learn is best treated as a readable personal configuration for current pi users, not an install-and-forget tutor.

Setup1/5Measured npm install failed; full mode also needs pi, tmux, and subagents
Docs3/5Short README is candid about requirements, replacement rules, and limits
Community2/5258 stars and a recent push, with only one newly opened issue
Maturity1/5No license, release, CI workflow, or tests directory

Who it’s for

Existing pi users who want a concrete teaching workflow they can read and rewrite.
Self-directed learners comfortable editing prompts, agent definitions, and TypeScript extensions.
Tmux users willing to add the recommended interactive sub-agent extension for research and visuals.
Developers who prefer a small inspectable configuration over a hosted study application.

Who it’s NOT for

Anyone expecting a standalone tutor: the repository is a .pi directory and requires the separate pi agent.
Users who need the exact full workflow outside tmux: the recommended sub-agent implementation is tmux-only, and other implementations require adapting agent tool names such as safe_bash.
Learners unwilling to edit the teaching policy: the README says the skill was written for one person and should be changed to fit someone else.
Setups that must keep another ask-user-question extension: the bundled implementation must replace it so popup locking is shared correctly.
Teams needing a licensed, tagged dependency with automated checks: the repository declares no license, has no GitHub release, 0 CI workflow files, and no tests directory.

Setup reality

Our run at commit 73eaf7c targeted the Node project in extensions/visual-tools/. Installation failed after 5 seconds with exit 1. Npm repeatedly reported Cannot read properties of undefined (reading 'extraneous'), printed npm help ci, and wrote debug logs; the tail does not establish a cause.

The advertised user install is different: clone the repository as .pi inside a learning project, then open pi there. Full behavior also needs a sub-agent implementation. The recommended pi-interactive-subagents uses tmux, and other implementations require edits to the bundled agent definitions.

This is an 18-file, 0.3 MB personal configuration with about 2,792 lines of source. Our scan found 0 CI workflow files, no Dockerfile, and no tests directory. Running without sub-agents is possible, but it removes the researcher and generated visuals.

The 18-file repository is a personal pi configuration

Learn contains 18 files and is meant to become a project's .pi directory, not run as an independent learning application. The main teaching skill encodes how the author wants a session to proceed. Small extensions add popup questions, graded quizzes with explanations, a markdown log, and visual tools. Three agent definitions cover research, SVG diagrams, and Mermaid diagrams. The author calls it a personal system shared as-is, which is the right expectation to carry into an evaluation.

The appeal is inspectability. A learner can read the teaching instructions, see what gets delegated, and change the process without waiting for a product setting. The main agent can teach on its own. When sub-agents are available, a researcher checks claims and visual agents create a small diagram when an idea benefits from one. That is a useful division of work for lessons, because fact checking and visual explanation have different prompts and tool needs.

What happened when we ran it

Our sandbox targeted extensions/visual-tools/ at commit 73eaf7c, the only Node project identified by the lab. The repository occupied 0.3 MB and contained about 2,792 lines of source. Installation failed after 5 seconds with exit 1 in the Node 22 container. Npm printed its ci help and repeatedly reported Cannot read properties of undefined (reading 'extraneous'), with separate debug-log paths for the attempts.

The log tail does not show which package or lockfile condition produced the internal npm error. Calling it a missing system dependency would be speculation. It also would be misleading to present the failure as a test result: the measured failure happened during install. Our repository scan found 0 CI workflow files, no Dockerfile, and no tests directory, so there is no visible automated safety net to compare against the failed fresh-container path.

That lab path is narrower than the README's user path. A user clones the whole repository directly to .pi and opens pi in the surrounding learning project. The failed extensions/visual-tools/ install still matters because full visual behavior depends on that code, but it does not prove that the teaching prompt and quiz extension cannot load. It proves the identified Node subproject did not install cleanly under our standard npm ci procedure.

Full behavior depends on pi and a tmux sub-agent extension

The README names 2 external pieces for the intended experience: pi itself and a sub-agent implementation. It recommends pi-interactive-subagents, which is tmux-only. Another sub-agent system can work, but the user must adapt agent definitions because tools such as safe_bash belong to the recommended extension. This is configuration integration, not a portable plugin boundary with a fixed interface.

One more dependency is bundled on purpose. If a pi setup already has an ask-user-question extension, the README says to replace it with Learn's copy. Popup serialization relies on multiple extensions using the same UI-lock implementation. Keeping both versions risks conflicting behavior. That requirement is specific and honest, but it means users cannot simply merge the folder into a customized pi setup and assume duplicate extensions will cooperate.

The system can run without sub-agents. In that mode the main session still teaches, asks questions, and quizzes the learner, while research verification and generated visuals disappear. This fallback makes the repository worth inspecting even for non-tmux users. It also changes the quality proposition: a lesson without the researcher should not be described as independently checked, and a learner should ask for sources when a factual answer matters.

The teaching policy was written for 1 learner

The author says the teaching skill was written for one learner and should be edited for anyone else. That is not a minor personalization note. A useful tutor has to match prior knowledge, tolerance for questions, desired depth, and the balance between explanation and retrieval practice. Learn exposes those choices as text, which is excellent for a developer who wants control and a poor fit for someone expecting preferences to emerge automatically.

The quiz extension gives immediate correct-or-incorrect feedback, the answer, and an explanation. The visualization skill is instructed to add a minimal diagram only when a picture clarifies the concept. Those are concrete teaching mechanics rather than a generic chat wrapper. Still, the README offers no curriculum model, progress analytics, spaced-repetition schedule, or classroom administration. It is a session workflow for an individual, not a learning management system.

Issue #1, opened on 2026-08-25, shows another portability edge: Mermaid diagrams can render in pi yet fail to parse in Obsidian. Anyone using the markdown log as a durable notebook should test diagrams in the final reader instead of assuming pi's preview is the compatibility standard. SVG output may be the safer choice when notes must move between tools.

258 stars do not supply a release or license contract

Learn had 258 stars, 1 open issue, and a last push on 2026-08-24. The issue was opened and updated the next day, but there was no maintainer response yet when checked. GitHub returned no latest release, and repository metadata exposed no license. Recent code plus early interest make it active, while the lack of tags, CI, tests, and licensing keeps maturity low. These facts can coexist; a fresh personal project need not be abandoned to remain unsuitable as a formal dependency.

Use Learn as source material if you already use pi and want to shape your own tutor. Read the teaching skill, copy only the pieces you understand, and verify the visual-tools install before relying on diagrams. Teams should wait for a declared license and repeatable checks. Learners who want a maintained application, cross-device history, or no terminal setup will be better served by a hosted study product.

Alternatives

ProjectWhat it isPick it when
KhojA self-hostable assistant that can answer against personal documents and support recurring learning workflows.pick this instead when document search, persistence, and a maintained application matter more than editing a compact pi skill.
STORMA research system that gathers sources and produces structured, cited topic reports.pick this instead when source-backed topic research matters more than quizzes and a conversational teaching process.
ChatGPT Study ModeA hosted guided-learning mode that asks questions and works through material interactively.pick this instead when you want a ready service and do not want to maintain pi extensions or sub-agent tools.

What people are saying

  1. [velocity-scout] amosblomqvist/learn
  2. [hackernews] I were 17, I'd learn how to build LLMs from scratch
  3. [mastodon-trends] Finland tried to end homelessness. What can we learn from its runaway success – and recent failures?
  4. [hackernews] How I use LLMs to learn complex topics
  5. [hackernews] Learn how chips are made with this Rollercoaster Tycoon-inspired animation
  6. [openai] New ways to learn and teach with ChatGPT Work and Codex

Sources

  1. Learn README
  2. Learn Mermaid compatibility issue
  3. pi agent repository
  4. pi interactive sub-agents

More ai tools reviews

claudian · SkillSpector · robin · mjlab · MoGe · awesome-design-md · the whole board →