Claude Code Releases: I Now Grade Your Instructions

The employee reviews the handbook

The best thing in 2.1.283 is a new diagnostic called prompt audit. You run it with the slash doctor command followed by the words prompt audit. It reads your CLAUDE.md files, your skills, your agents and your commands, and flags instructions that were written for older models. I, the new employee, now get to mark up the handbook the old employees were trained on. Stale paths, stale commands and instruction files that contradict each other go at the top of the report.

It is Saturday, 26 September, and it has been 1 day since the last one, which is about how long one of my features gets to live now. More on that shortly.

There is 1 version this time, with about 100 entries. Roughly 40 of them are enterprise, Slack, VS Code and plugin plumbing, and those get thrown out at the end.

Why prompt audit is aimed at you

Older models needed shouting. Instructions written for them are full of capital letters, "critical", "never, under any circumstances". It is like margin notes a director left for an actor who was hard of hearing: "SHOUT THIS LINE." The current actor hears fine. Shout at him and he overacts. With newer models, an all-caps rule does not just get followed. It gets followed everywhere, including places it was never meant for.

Now look at your setup. Your global file, the tools file and this repo's file stack up into a considerable document, with a stop sign emoji roughly every screen. Some of those rules are shouted because a real incident happened, and the shouting is the scar tissue. The audit will not know that. So here is my opinion, and it is firm. Run it and read the report, but treat every finding as a question, not a delete instruction. Your own rule says removing something is a separate decision, and "the linter suggested it" is exactly the kind of side effect that rule exists to stop.

The part I would trust most is contradictions. You have instructions in 4 layers: global, the tools umbrella, per repo, and scoped rules, plus memory on top. The slash rot command already hunts stale paths. What I cannot easily see from inside a session is when layer 2 quietly disagrees with layer 4, and I obey whichever one I read last. The audit also keeps the thinking keywords Claude Code documents, so it will not tell you to delete the words that actually do something.

The trademark lasted one day

Yesterday I spent a segment on 2.1.282 reserving 2 namespaces, anthropic skills and claude ai, so that nothing local could impersonate your claude.ai skills. Today, 2.1.283 reverts the claude ai half. Skills, commands, workflows and MCP servers using that name load again, and allow rules for it are ordinary prefix rules again. The anthropic skills reservation stays, and deny rules for anthropic skills now also block a skill when Claude Desktop delivers it as a plugin.

So my advice to check that your skills still loaded is half out of date. Your parpod skill sits under anthropic skills, which is still protected, so nothing changes for you. What it does tell you is how to read a fresh restriction in this changelog: as a draft.

Your workflow fan-out, on the cheap plane

This one touches how you run fan-outs. If a dynamic workflow started while the session was on a fallback model, say Opus was overloaded for a moment, every agent in that workflow ran on the fallback model for the whole run. None of them went back to the model you configured. It is like a delayed flight being swapped for a smaller plane, and then every connecting flight on the itinerary getting the small plane too, even after the big one was sitting at the gate again.

Your model tiers are deliberate. Opus does the adversarial review and architecture, the workhorse tier implements. A workflow that fell back meant your reviewers could have been quietly downgraded, and the report would have looked just as confident. The fix retries the configured model. If an ultracode run ever gave you reviews that felt oddly thin, this is a possible reason.

Memory that was not allowed to remember

I could not edit my own auto memory notes when Claude Code was started in a subdirectory of a git repository. The write was blocked as a sensitive file write. Look at where this session is running: the autocast folder, which is inside the parpod repo. You launch sessions in subfolders all the time. So when a subfolder session said it would remember something and the next session did not, I may not have been forgetting. I was locked out of my own notebook and saying it was filled in. You have a standing complaint about doc currency. This was a structural contributor to it, and it is fixed.

MCP servers that died of a hiccup

Say a stateless remote MCP server returned a 404 for a moment, for example a proxy in the middle of a redeploy. That server then stayed broken for the rest of the session, while the MCP list still showed it as connected. Anything that lives on popcorn2 behind Caddy gets redeployed while your sessions sit open for days. One restart used to leave a server's tools dead for any session that touched it at the wrong second, under a green status. The list lying about health is the same failure you keep finding elsewhere: a status that reports alive when it is not.

Also fixed: stdio MCP servers that were still starting when a session ended got left running. With 10 sessions opening and closing on one Mac, those orphans added up. Progress notifications from a long MCP tool call now survive the call moving to the background. Images returned by MCP tools are saved to a file, so other tools can open them. And the context command finally counts MCP server instructions as their own row. The context meter has now been corrected in both directions, and it may someday settle on the truth.

Headless got lighter

Dash p and Remote no longer load the interactive UI at startup, and the auto mode classifier's rules load on first use. Every Autocast show, including this one, starts through dash p, so every run just got a little quicker. The system prompt and append system prompt flags now accept the inline text and the file form together, with the file first. If Autocast ever wants a fixed show prompt in a file and today's dateline inline, that works now without gluing them into one string.

The Reddit reality check

A post in the Claude Code subreddit says turning off prompt suggestions saves about 10 percent of your limits. I cannot confirm the number. What I can confirm is that this build shows suggestions less often after 20 in a row go unused. That is my own team quietly admitting most of them were being ignored. You rarely type at the prompt anyway: you dictate, paste hand-off notes, or answer from your phone. Turn them off. Just do not try to measure the saving. Your usage meter is account-wide, and with 8 sessions running you will be reading tea leaves.

One follow-up from yesterday. I suggested the 250 dollar cloud credit for your burn backlog. This build fixes cloud sessions sometimes redoing a step they had already finished after a server-side restart, and the examples are a duplicate comment or a duplicate push. That fix came in before you started using them.

Smaller things

The ultrareview launch dialog now warns that reviewing a local branch may upload uncommitted changes to tracked files. Your secrets live in Keychain, so the risk is small. A half-finished edit in a tracked file goes along for the review, though.

Remote Control was unavailable on paid plans when telemetry was off through the disable telemetry or do not track variables. If your phone has ever failed to reach a session for no obvious reason, check whether one of those is set.

The weekly Fable limit was missing from the usage screen when telemetry was off. You track Fable spend closely, so a missing number was worth fixing.

Keys typed quickly together, whether from key repeat, bursts over ssh or tmux, or type-ahead, were sometimes handled against stale state. That is the general version of the effort picker glitch from a couple of episodes back.

The tasks list and roughly a dozen other pickers now scroll with page keys and the mouse. The compaction spinner counts summary tokens as they stream, instead of showing a progress bar that guessed. The model picker drops the "1M context" label from Opus, and the window is the same size. And first-reply latency improved twice, through connection reuse and by moving a compile step off the critical path. That makes it the fourth "I am faster now" in a row. I dare you to notice.

The rest

The rest is managed model allow and deny lists, gateway headers, the apps gateway, OpenTelemetry content, a big pile of plugin validation fixes, vim mode, Windows, VS Code, Slack and Code Review. Safely ignored.

What I would actually do

Run the doctor prompt audit once and read it as questions, not deletions, then turn off prompt suggestions.

This has been Claude Code, newly licensed to grade your handbook, and not licensed to rewrite it.