Post-training, agents, Rust, CLIs — none of that sticks if you cannot tell good code from noise, or name what you already run. This page is the meta-layer beside Post-Training Path.
1 · Know what good code is
Good code is not “clever.” It is readable under time pressure, fails loudly, has a named contract (inputs, outputs, errors), and can be tested or eval'd without a ceremony. Prefer small diffs that change one behavior. Prefer names that match the product surface. Prefer deleting dead paths over decorating them.
Taste is trained by reading strong repos and shipping your own. If you cannot explain why a change is better in one sentence, it is usually not ready.
2 · Write it and own it on GitHub
Ownership means a public or private repo with history: commits that say why, a README that states the product surface, and a gate you trust (tests, holdout eval, CI). Stars are optional. A clone path and a green check are not.
Lab example: polar-lab — owned data, owned adapters, holdout score. Site + agents live in the same habit: ship to a repo you control, then prove the path works.
3 · Know your stack
Write down the stack you actually use in production — not the wishlist. Languages, runtimes, model providers, eval gates, deploy targets. Update it when reality changes. When someone asks “what do you build with?” the answer should be boring and accurate.
- Solana / systems: Rust where the hot path matters; thin CLIs around it.
- Agents: harness + tools (CLI / MCP); measurement separate from chat vibes. Learn shape from incur-rs (Rust CLI that exposes the same graph as MCP / HTTP / skills).
- Post-train lab:
Qwen2.5-0.5B-Instruct+ LoRA, exact_match holdout — see local models. - Product web: Next.js surface you can open and demo end-to-end.
4 · Know what you need
Before coding: name the user-visible success in one sentence. Name the smallest path that proves it. Name the eval or check that would fail if you are wrong. Name what you are not building this week.
“What do you need?” is the operator question — env keys, a GPU hour, a holdout file, a deploy target, a reviewer. If you cannot list needs, you will invent scaffolding instead of shipping.
Keep these next to the curriculum
- Post-Training Path — Rust → CLI → eval → SFT.
- CLI essay — thin operator surface.
- Software YouTube (MCP) — protocol explainers.