This page is the guide for anyone landing on the deployed site rather than the repo directly. The workshop itself, five modules plus the fixture they run against, lives in the repo, whose own README covers the same ground in more detail once you've cloned it. Nothing here substitutes for cloning it; this is orientation, not the content.
See it before you clone anything
A real, condensed transcript of an actual attempt at Module 04's core exercise: reproduce a bug, root-cause it, fix it, prove it, in that order. Not a mockup: a script extracts the text straight from the real run, unedited. The raw version is linked from that same page.
The idea: the harness is the classroom
Every exercise in this workshop runs inside your own coding-agent tool (Claude Code, Codex, Cursor, or equivalent), not on a page you read. A module states a task and a rubric up front, you drive your agent through it, and feedback comes from your actual attempt: what the transcript shows, not a generic answer key. Reading a module's README is orientation. The module completes when something you did passes its stated gate.
This is a hypothesis, not a settled finding: that the next skill past "get a plausible answer once" is a set of four disciplines (precise instructions, curated context, bounded tool access, supervised loops) that compound, plus the judgment to tell which one is actually broken when something doesn't work. The build log is where that hypothesis gets argued with in public, including the places it's wrong.
What you walk away with
Two different things get built during each module, and they're both real: your own judgment, and a reusable artifact your agent can load next time. Neither is a guarantee, both are what the exercises are designed to produce; see each module's own README for the evidence behind its specific gate.
- You get practiced judgment: writing an instruction precise enough to be reproducible, deciding what a model actually needs to see, scoping what an agent can reach, telling a loop's terminal state apart from a loop that just stopped, and diagnosing which layer is broken when a setup doesn't work.
- Your agent gets the packaged takeaways: a prompt template, a
context-budgeting Skill, a sub-agent or harness-config definition, a loop template, a
diagnostic playbook, each one built from an exercise actually completed, not written
from theory. Concrete form and reference implementations are in the repo's
.claude/directory once you clone it.
The runbook
- Prerequisites: comfortable with git, the CLI, and reading a diff; already using at least one coding-agent harness regularly, installed on your machine.
-
Clone it:
git clone https://github.com/coderturtle/terminal-velocity.git -
Start at
modules/README.md. It has the full arc, in order, and a table of what each module's gate requires and what you keep from it. - Work through the five modules in order: prompt, context, harness, loop, then the synthesis capstone. Each assumes the ones before it; the capstone only makes sense once you've done all four.
- No module completes by reading it. Every module states a required gate, an artifact you produce or an action you're observed doing. If a module ever reduces to "read this, then move on," or a link here is broken, that's a defect: open an issue.