Learn Anything — PathMX Learning Labs

What if a useful conversation with an agent became a learning path you could keep, change, and continue tomorrow? Our OpenAI Build Week 2026 project uses Codex and PathMX to turn ordinary Markdown into focused, playable learning experiences.

The submission has three parts:

1. This Learning Labs repository records the experiments, decisions, evaluations, and evidence.
2. The public PathMX Learning Starter gives a learner a durable workspace with the agent skills included.
3. The playable examples explore practice, simulation, and creation through the same Markdown-first format.

Canonical public companions:

• PathMX Learning Starter — the learner-facing starter for durable personal learning Paths
• PathMX Skills — the canonical agent skills for authoring PathMX and guiding personal learning (bootstrap.md is the one-prompt entry point)

Start with the public Learning Labs site or the three-minute submission walkthrough.

Working claim

Codex is the conversation and orchestration surface. PathMX is the durable, playable learning surface. The learner leaves with a living curriculum, evidence of learning, and a next step—not only a chat transcript.

How Codex and GPT-5.6 were used

We worked in Codex with GPT-5.6 throughout Build Week. It helped us research learning patterns, author and revise Markdown Sources, implement interactive components, diagnose Player behavior, run verification, maintain the task graph, and synthesize evaluation findings. The repository itself is the durable record of that collaboration: see the task dashboard, technical eval review, and change log.

The team chose the product direction, set the learning goals, reviewed the experiences as learners, and accepted or rejected the results. Codex and GPT-5.6 accelerated the work; they did not replace those decisions.

What existed before Build Week

Mark Johnson created the PathMX methodology, framework, and Player before the event. During Build Week, the team built or meaningfully extended the Learning Labs workspace, learner Starter, agent skills, interactive examples, reusable components, evaluation harness and evidence, and the workflow for preserving human-agent decisions. PathMX is currently in Labs, with its Core source kept private while its APIs and product boundaries stabilize. Core changes supported the public work but are not the submission's primary technical claim.

The PathMX Core progress log records that boundary without exposing the private Core repository.

Start with one Codex prompt

Open an empty writable folder in Codex Desktop and send:

Follow the instructions at https://raw.githubusercontent.com/pathmx/pathmx-skills/main/bootstrap.md. Create a new learning space in ./learning-space and help me learn [your topic or goal].

That hosted file is the canonical bootstrap. The agent checks or installs Bun, sets up the current PathMX Labs release, creates a repository from the Learning Starter, asks a short set of learner questions, and launches the Player. See the judge and learner walkthrough for expected checkpoints and fallbacks.

Try it without rebuilding this project

No OpenAI API key, credits, or plugin build is required for the judge path.

1. Open the public Learning Labs site.
2. Follow the submission walkthrough.
3. Open a featured example and use its visible Player controls.
4. Inspect the linked Markdown Source in this repository to see the durable artifact behind the experience.

The one-prompt flow above tests the learner-facing agent skills without compiling or packaging a plugin. For repository inspection, see PathMX Skills and the public Learning Starter, which ships the same repo-local skills under .agents/skills/.

Local setup and test

Prerequisite: Bun. Until the submission is frozen, local verification follows the latest published PathMX Labs CLI.

bunx @fellowhumans/pathmx@latest build paths/index.path.md -o .pathmx-check --clean
bunx @fellowhumans/pathmx@latest play

The first command is the reproducible build check. The second launches the Player for local review. The primary judge path does not require credentials; AI image generation in unrelated experiments is optional and is not part of the required test.

Repository shape

AGENTS.md
.agents/skills/pathmx/SKILL.md
paths/
├── index.path.md
├── tasks/index.tasks.md
├── labs/index.path.md
├── research/index.path.md
├── work-log/index.path.md
└── work-log/changes.log.md

• paths/index.path.md is the project hub.
• paths/tasks/ is the Codex-managed work queue and evidence trail.
• paths/labs/ contains experiments that may fail.
• paths/research/ collects useful external references.
• paths/work-log/ contains dated notes, handoffs, and the playable changes log required before shared pushes.

Private information

Do not commit Devpost invitation links, credentials, private contact details, learner records, raw transcripts, or Codex task identifiers. Share sensitive team administration through the agreed private channel.

Status

Active Build Week execution. Current ownership, blockers, next moves, and completion evidence live in the task dashboard.

Current limitation: the demonstrated loop is local and single-player. Shared and multi-user Paths remain future work.