About Vibe Code Textbook

Vibe Code Textbook is a site about writing software with AI coding agents. It is maintained by one person, with no vendor behind it; the articles are produced from the sources each one names, under that person's direction, and how this site is made has the details.

Editor

C. B. Zakarian (a pen name) is the owner and editor of Vibe Code Textbook. C. B. Zakarian holds a Ph.D. in Information Systems, with a focus on data science, from the University of North Carolina at Greensboro (2016), and is an assistant professor.

C. B. Zakarian decides what Vibe Code Textbook publishes and when, and the figures on each page are checked against the sources it names before publication. Pages the editor has reviewed in full show the date of that review, and errors are corrected when they are found.

What "vibe coding" means here

The phrase started as a joke about letting the model drive. This site takes the tools seriously and the joke lightly: the agents are genuinely useful, they are changing monthly, and most of what is written about them is either marketing or a screenshot with no follow-up. The aim here is to write it down properly.

What is published

Harnesses are walkthroughs of the agent runners themselves: what each one does, how it is configured, what its permission model actually permits, and where it breaks, read from the documentation and the source at a named version.

Workflows are repeatable ways of working with an agent, drawn from the harnesses' documentation and published practice: spec-first, tests-first, review loops, context files, and the habits that stop a session going sideways.

Repos worth reading are open-source projects in the agentic-coding space, read at a specific commit, with what they get right and the parts worth borrowing. Star and issue counts carry the date they were read, because they move.

Tutorials walk through building something with an agent: the prompts to give it, the tests that decide whether the result is right, and a shipped script and test file you can run yourself. The tutorials themselves record no agent session, and each says so; the agent runs on their code are in the time-to-green chapter described below.

Measurements are numbers instead of vibes: benchmark results task by task, release timestamps and the like, taken from public data or from our own runs, with exactly how each was collected. Each ships its dataset, and every one we computed ships the script that computed it; the SWE-bench explainer collects published scores and runs nothing, and says so. Time-to-green is the one measurement made by running an agent: on 30 October 2026 we ran its ten small tasks three times each through a coding agent's headless mode, with two models, 60 attempts in all. The chapter names the agent and its version and publishes a row for every attempt, the token ledger and the redacted transcripts of all 60.

What this site refuses to do

No feature claims that were not checked against the docs or the code. No made-up transcripts. No tool versions left unstated. No "reportedly". If a tool changed after an article was written, the article gets an updated date and a note, not a quiet rewrite. No exclamation marks.

Corrections and counterexamples are welcome and acted on: contact.