Signature experience · synthetic session

The Terminal

To ship with Claude Code, run a verification loop: confirm the binary with `claude --version`, load a skill for the procedure you repeat, put non-negotiable gates in hooks, reach for MCP only when the ticket lives outside the repo, and do not stop until a check produces evidence you can read. The Terminal below plays a labeled synthetic session so you can practice that loop. For forensic log reading, use the Black Box on Claude Logs.

Play the sessionOpen the Black BoxRead the tutorial

Synthetic

Specimen TX-001

Add a Stop hook so the session cannot claim done without a passing test

ClockT+00:00
Record1 / 16
BeatNotice
Modelclaude-fable-5-1

Arrows step · space plays · 1–4 jump beats

record 1T+00:00noticespecimen

Specimen notice

Nothing here is uploaded or fetched; the specimen ships with the page. Independent educational publication. Not affiliated with Anthropic, Claude, or Claude Code.
Readout
Four beats, and the records that teach them.
This is a clean shipping loop on purpose. The method is the same one Claude Logs uses on a messy transcript: read in order, separate claims from evidence, confirm the outcome against the repository. The specimen is different. Here the check passes and the diff matches the ticket. When you need the forensic case — a skipped test reported as green, a credential written through a tool input — open the Black Box.
L-01 · Version check
The governing CLI is the one you just ran
Do not copy a version number out of a blog post, a screenshot, or this site. Run `claude --version`, then read the official changelog forward from that string before you change hooks, MCP, or CI.
Evidence: records 3, 4
L-02 · Skill invocation
A skill is a procedure that should not live in every session
Invoke skills inside Claude Code for repeatable checks. Authoring and collecting them belongs on Claude Has Skills. If the procedure is needed in most sessions, it may belong in CLAUDE.md instead.
Evidence: records 5, 6
L-03 · Hook / MCP
Hooks guarantee; MCP fetches live context
Use MCP when the ticket or the trace lives outside the repo, and keep it read-only until a write is the job. Use a Stop hook when forgetting the check is unacceptable. The hook in this specimen is what turned a reminder into a gate.
Evidence: records 7, 11
L-04 · Shipping check
Done means evidence a reviewer can read
Tool results are evidence. Assistant text is a claim about that evidence. Confirm the counts and the diff before you accept the summary, then keep a human in the merge path.
Evidence: records 13, 15, 16
What this session proves
  • The order of a shipping loop: version check, skill load, live ticket context, deterministic Stop gate, then a check whose output is still in the session.
  • That the Stop hook refused a premature done and named the missing command.
  • That the final claim cites a 12/12 test result and a two-file diff, both logged immediately before it.
What it does not prove
  • Which Claude Code binary you have installed. Only `claude --version` on your machine or runner proves that.
  • Whether LEDGER-1842 is a real ticket, or whether this hook belongs in your repository. The specimen is synthetic.
  • What a messy session looks like when the summary and the tool result disagree. That forensic read lives on Claude Logs.
What a reviewer should do
  • Run `claude --version` locally and confirm it against the official changelog before changing team policy.
  • Keep /verify as a project skill if the check is occasional; move the gate to a Stop hook if forgetting it is unacceptable.
  • Review the hook script in the PR the same way you review CI. Then merge only after a human has seen the diff.
Run This Loop On A Real Task
The specimen is a rehearsal. The loop transfers to any Claude Code session that is supposed to ship: confirm the binary, load the procedure, fetch only the live context you need, enforce the check, then review the diff.
Read in order, not from the end. The final assistant message is the most compressed record and the least reliable. Record 16 is a fair summary only because records 13 and 15 are still sitting next to it. Anthropic documents session storage under ~/.claude/projects/; the transferable habit is to read the file, not to trust the last paragraph.
Sort every record into claim or evidence. Tool results, exit codes, test counts, andgit diff --stat are evidence. Assistant text is a claim. A skill description is a contract, not a result. Apply that one rule and the shipping loop is legible: the Stop hook in record 11 is evidence the gate fired; the 12/12 in record 13 is evidence the check passed; the sentence “ready for human review” is a claim about both.
Give each extension one job. The /verify skill is the repeated procedure. The Stop hook is the guarantee. MCP is the live ticket. CLAUDE.md would be the wrong home for any of those. If you are here to author skills, that work belongs onClaude Has Skills. This site stays on how those skills are used inside Claude Code.
Do not pin a CLI version from a webpage. Record 4 is empty of a version string on purpose. Run claude --version on the machine that will ship. Then read theofficial changelog forward from that string before you change hooks, MCP, or CI. As of September 2026, the model default that changes cost and latency is Claude Fable 5.1 (claude-fable-5-1) at High effort inside Claude Code.
Close against the repository, then keep a human in the merge path. Two files under.claude/, twelve passing tests, three ticket criteria. That is a complete handoff. It is not a merge. For CI and GitHub Actions, continue on theCI guide. For the messy transcript where this method catches a lie, use the Black Box on Claude Logs.
Why The Specimen Is Synthetic
A real session would have been easier and would have been the wrong artifact. Claude Code transcripts contain other people’s prompts, absolute paths, tool inputs, and whatever happened to be in the working tree. Publishing one to demonstrate a shipping loop would publish a private engineering conversation. The page would have been a counterexample to its own review advice.
So we wrote one. Every record in TX-001 is fabricated by Claude Ships Code: the repository, the ticket, the hook script, the test names. The commands are the real shapes — claude --version, a /verify skill load, a read-only MCP issue fetch, a Stop hook, npm test,git diff --stat — because a specimen that only works as prose teaches nothing about the loop. The results are invented. The discipline is not.
We do not ingest, store, or publish real private transcripts. There is no upload path on this page. Independent educational publication. Not affiliated with Anthropic, Claude, or Claude Code.
FAQ
Is The Terminal a real Claude Code session?
No. It is a labeled synthetic specimen authored by Claude Ships Code for teaching. Every command, path, ticket, and hook script in it is invented. It is marked synthetic on the page and in the first record. Publishing a real private transcript to illustrate a shipping loop would publish someone else's prompts, paths, and tool inputs.
What Claude Code version should I pin from this page?
None. This site does not publish a Claude Code CLI version number. Run `claude --version` on the machine or runner that will ship, then read the official changelog forward from that string. As of September 2026, the model default that matters for cost and latency is Claude Fable 5.1 (`claude-fable-5-1`) at High effort inside Claude Code.
How do I review what Claude Code did in a session?
Read the session in order. Treat tool results, exit codes, and diffs as evidence, and treat assistant text as a claim. Confirm the final summary against the check output and git. The Terminal is a clean shipping loop so you can practice that read. When you need a forensic specimen — a declared sandbox that was not a sandbox, a credential in a tool input, a skipped test reported as green — use the Black Box on Claude Logs.
Where should I go to author Claude skills?
Claude Has Skills owns skill authoring and the skill directory. This page only shows a skill as it is invoked inside Claude Code: a project SKILL.md loaded with /verify, with allowed tools and a verification contract. Do not treat this site as a catalog of skills.
Is Claude Ships Code affiliated with Anthropic?
No. Claude Ships Code is an independent educational publication and is not affiliated with Anthropic, Claude, or Claude Code.
Cite This Page
Claude Ships Code. “The Terminal: practice a Claude Code shipping session.” Updated 2026-09-19. https://claudeshipscode.com/terminal/
The specimen is original work by Claude Ships Code and is labeled synthetic. Cite it as an illustrative artifact, never as observed behavior of a Claude model. Claude Ships Code is an independent educational publication and is not affiliated with Anthropic.