A Build is one working session with a coding agent, compressed. Not a tutorial, not a transcript. Three to twelve moments a reader can scroll in two minutes, with the raw log one layer down for anyone who wants it.

What a Build is

Header, Goal, Timeline, Verdict, Raw session. The header carries the result and two numbers: session length and how many times you intervened. The timeline is the story. The verdict is what you would do differently. Everything else is optional.

A Build with zero interventions is a Build. A Build that failed is a Build. Both get their own feed filters, because both are where readers learn the most.

The title states the result

Write what happened, with a number if there is one. Past tense. The agent name goes in the metadata, not the title.

YESMigrated 40 Laravel queue jobs to Go workers in one eveningLost the plot on a 3k-line Rust refactor. Three fails, then I gave up

NOMy experience using Claude Code for a migrationRust refactor attempt (long)

Five kinds of moment

A moment is one card on the timeline with a time offset. Three to twelve per Build; the first is a Prompt, the last is the Outcome. If you have more, you are transcribing: merge the agent stretches.

Moment	Use when	Write
Prompt	What you typed. Only the turns that changed direction.	Verbatim, trimmed. The full text sits behind “Show full prompt”.
Agent did	A stretch of agent work between your turns.	What it did, not what it printed. One or two sentences: files, approach, result.
Intervention	You stopped or corrected the agent.	What you saw, then Why. The why is required; it is the part other people reuse.
Fail	The agent was wrong and nobody caught it in time.	What broke and how you found out. A Fail without a following Intervention is allowed.
Outcome	What shipped. Exactly one per Build, always last.	Numbers over adjectives: PRs, tests, latency, lines. Link the repo.

How to write the Intervention “why”

The why is the reusable part of a Build: a rule another person can paste into their own first prompt. Say what the agent could not have known, not how you felt about it.

Good: “PHP consumers read the old payload shape for two more weeks. Schema is frozen until cutover.”
Good: “Tests were green, so the agent had no way to see the shutdown hang.”
Not useful: “It went off the rails.” “Classic AI.”

The session spectrum

The bar under every title is drawn from your timeline. Prompt, Intervention and Fail are marks of fixed width, and so is the Outcome at the end. The stretches between them are agent work, sized by minutes. A red-heavy bar means many interventions, not long ones.

Verdict

Two to four sentences. What you would put in the first prompt next time, which correction did the work, and a link to the repo or PR. The first sentence becomes the teaser in the feed, so lead with the lesson.

Export a session

Most sessions are a file on your own machine. Upload it, and we turn it into transcript turns and propose moments; you edit every one of them before anything is published. Big files are fine: screenshots and long command output are cut in your browser before the upload.

Fastest with Claude Code: the plugin. Type /coders-talk:build at the end of a session and it sends that session to a new draft, after showing you what goes. Install the plugin.

Claude Code (terminal, desktop app, VS Code, JetBrains). Open ~/.claude/projects (on Windows %USERPROFILE%\.claude\projects), then the folder named after your project path with dashes, for example C--code-shop. Take the newest .jsonl file. The folders and small .json files next to it are bookkeeping, not the conversation.
Codex. Open ~/.codex/sessions, then the year, month and day, and take the rollout-….jsonl file.
Cursor, Windsurf, Copilot and other editors. Export the chat as Markdown if the chat menu offers it, or copy the whole conversation and paste it as text.
ChatGPT, Claude.ai and other web chats. Share links need a signed-in reader, so they can't be imported. Copy the conversation from your first message to the last reply and paste it as text.
Already published somewhere. Links from gist.github.com, lore.link and dev.to work. A secret gist is a good way to pass a .jsonl without uploading it here.

Pasted text has no timestamps, so we space the turns evenly and mark the times as approximate; fix the durations while editing.

Raw session and secrets

The transcript publishes collapsed under “Raw session”, and you can hide it at any time. Before anything reaches our model or the page, we scan for keys, tokens, private keys and connection strings and replace matches with [REDACTED:TYPE]. Publishing is blocked until you have confirmed each finding. The scan is not complete: read your moments before you publish.

Community and privacy rules

Post your own sessions, or sessions you have permission to share.
No customer data, credentials, internal URLs or personal data about other people, in moments or in the raw session.
Output dumps are not moments. If a moment is mostly pasted agent text, summarize it.
Sessions need an outcome. Abandoned is an outcome; “still working on it” is not.
One run per side is not a vendor comparison. Post the run as a Build and let replicas do the comparing.
Be specific in comments and reply on the moment you mean. Editors hide spam, harassment and leaked data; three reports hide a Build until an editor looks at it.

See the content policy and the privacy policy for the full rules.