ena pragma
·12 min read·ai-operations · agents · ai-verification

Six small things to try this week with Claude's new features, and the trap in each

Six jobs for a business reader, each with what it needs, from Anthropic's own docs: the steps, why it works, and the trap. Checked October 25, 2026.

Contents

Anthropic's docs and changelogs show new and changed Claude features in September and October 2026. This note picks six a business reader can try this week. Each lists what it needs, the steps, why it should work, and the trap. Two, side chats and the advisor, were already in Claude Code by March and April 2026 (the oldest GitHub changelog mentions are entries 2.1.79 and 2.1.117) and changed again in that window. Item 3 is Pro and Max only and still rolling out. Five items are in Claude Code (a terminal, claude.ai/code or the desktop app's Code tab).

We read every Anthropic page named below on October 25, 2026, Central time, and ran none of the six. Plans and versions move fast, so check them first. A last-modified date in Sources is when a page changed, not when the feature launched. Two items use an analogy from another field, marked Frame. It is not evidence about Claude.

#1. Count your skills like a stockroom

Needs. Claude Code v2.1.261 or later for /skill-doctor. The skills page says 2.1.252, but the changelog lists the command as added in 2.1.261, dated September 4, 2026. It "isn't available in sessions that skip feature-flag fetching": Claude Code skips fetching the switches that turn some features on when telemetry is off or on providers such as Amazon Bedrock. /doctor prompt-audit "requires Claude Code v2.1.283 or later."

Steps.

  1. Write a keep-list first: skills you rarely use but cannot do without, such as year-end work. A usage report will make them look idle.
  2. Run /skill-doctor. It shows each skill's context cost and use count.
  3. Turn off flagged skills that are not on your keep-list, highest cost first.
  4. Run /doctor prompt-audit. Per Anthropic's memory page, "nothing in your files changes until you ask Claude to apply them," so read each edit first.

Why it should work. Every listed skill "adds to your context on every turn, whether or not Claude ever uses it." Anthropic says "The budget scales at 1% of the model's context window." Our reading: the listing has a ceiling, so the saving from turning skills off has one too.

Frame (inventory). A skill nobody calls is like dead stock in a stockroom: it takes up room and earns nothing. The comparison is ours.

Trap. The report tells you where to turn skills off. Our reading: it does not delete them. The audit is a subcommand: plain /doctor runs a setup checkup instead.

#2. Put a skeptic in the side seat

Needs. Claude Code in the terminal or the desktop app's Code tab (local, SSH and WSL sessions).

Steps. While a long task runs, open a side chat (Cmd+; on macOS or Ctrl+; on Windows in the desktop app, or type /btw) and ask three things. "In one sentence, what do you think the main task is?" "What is the riskiest thing you assumed?" "What would change your answer?" Fix the brief if they show a misunderstanding. In the terminal each side question gets one response, so type /btw per question. In the desktop app, ask questions only: the page does not say a side chat there is barred from tools.

Why it should work. The desktop page says "A side chat lets you ask Claude a question that uses your session's context but doesn't add anything back to the main conversation." In the terminal, per Anthropic, "side questions answer only from what is already in context. Claude can't read files, run commands, or search when answering a side question." So there the skeptic cannot change anything.

Trap. It reads the same transcript, and we expect the same model (the pages do not say), so it likely shares the main chat's blind spots. Our reading: it catches misunderstandings, not gaps in knowledge. The desktop page says "The desktop app doesn't save side chats to disk, so you can't return to one after you close the app," so copy out any answer you want to keep.

#3. Run the month-end close as a project, with a handoff note

Needs. Claude Code Projects, launched in a September 17, 2026 article. The docs page says "Projects are in public beta on Pro and Max plans and rolling out gradually," and "Projects aren't offered on Team or Enterprise plans yet." The rollout "started with accounts that have used cloud sessions and don't have existing projects in claude.ai chat or Cowork." If Projects is not in your sidebar or under More, it has not reached you.

Steps.

  1. Decide what you may upload (see the data line), then create a project and upload last month's reports. Anthropic says "Upload the documents instead of adding a repository."
  2. Tell it: "Start a thread each for receivables, payables and payroll tie-out. Every thread ends with a five-line note: status, numbers changed, open items, needs a person, next step."
  3. Read the notes, then tie every number in them to your ledger or bank before anything is posted.

Why it should work. "You tell it what needs doing and it starts a thread for each task."

Frame (handoffs). A 2025 review of hospital handoffs found "moderate certainty evidence that the I-PASS tool" (a bundle of a fixed format and training) "reduces medical errors and adverse events." It says nothing about five-line notes, and we found no set report-back format on Anthropic's page. The note is our borrowing.

Data line. Bank statements and payroll are sensitive, and each thread is usually a cloud session, "Claude Code running in the cloud rather than on your machine." Check your plan's data terms and your organization's settings, and strip account numbers, names and salaries before uploading. Every cloud thread can use all your claude.ai connectors (apps linked to your Claude account, such as email) "with no per-project setup," so turn off every connector the job does not need (email, files, accounting, payments): open a thread and choose Connectors from the + menu beside its message box. That also saves it as your account default, so new threads and your claude.ai chats start without it until you turn it back on.

Trap. "Cloud threads don't pick up anything from the Claude Code setup on your own machine," so local skills and checks will not run there. And "A project draws on the same plan limits as your other Claude Code sessions and uses them faster."

#4. Make a dashboard your bookkeeper can audit

Needs. Claude Dashboards. Anthropic's Help Center page says it is "available in beta on Pro, Max, Team, and Enterprise plans," and on Enterprise it is "off by default until an owner turns it on."

Steps. Connect a data warehouse (Amazon Redshift, BigQuery, ClickHouse, Databricks or Snowflake) or an app already connected to Claude, such as Salesforce. If you have neither, skip this item. Ask Anthropic's example, "Build a dashboard of this quarter's revenue by region", open a chart's query, and have your bookkeeper tie the totals to the ledger.

Why it should work. Anthropic's launch article says "Click any number to see the query behind it, or ask Claude to explain it."

Trap. Anthropic says Claude "keeps it current as the data changes," but not how often. Read each chart's last-refreshed stamp before you trust a number. Our reading: a chart's query can show table and column names. Anthropic says "Dashboards start private to you," so share one only with people who should see it. Sharing can reach beyond your company. The sharing page, modified September 23 (before Dashboards launched), offers Pro and Max "Anyone with the link," leaves sharing outside the organization to an owner on Team and Enterprise, and says artifacts that "connect to your apps" cannot use that link. The launch article says dashboards can be shared "with anyone who has the link, if your admin allows it." We did not resolve that.

#5. Cheap hands, a consultant on call, and a person at the door

Needs. Claude Code. The advisor is experimental and needs a direct Anthropic connection. Its page lists Claude Platform on AWS among the unavailable platforms, yet the 2.1.300 changelog entry, dated October 14, 2026, says "Added /advisor on Claude Platform on AWS", so check yours. It stays off in a session where a variable that turns off flag fetching (Claude Code's fetch of Anthropic's feature switches) is set, "such as DISABLE_TELEMETRY." A mod is a small add-on, written in code, that changes how Claude Code behaves. Admins can limit or block mods.

Steps.

  1. For a long, routine job such as sorting messages, run claude --model sonnet --advisor opus: the docs' claude --advisor opus plus a main model, and the pairing table lists Sonnet 5.5 with Opus 5 or later. The flag covers that session only. /advisor opus is saved in your user settings, where it "persists across sessions." After launch you should see "Advisor Tool (experimental) is on". If not, no advisor is running: an advisor that ranks below your main model starts the session without one, and some pairings exit with an error. A pairing the API refuses after launch is dropped with no error, so after the end-of-job consult look for an "Advising" line in the transcript. The flag is not listed in claude --help.
  2. Add to your instructions: "Before you say a whole job is done, consult the advisor and list the three riskiest items." Ask for one consult at the end, then check /usage.
  3. To make Claude ask before it touches your email or messaging connector, start with a permission rule, which takes no code: ask Claude to add an ask rule for that connector's tools, then try one harmless call. Use a mod only if a rule cannot do the job, and have Claude write one that requires a person's approval, not an automatic one. Read its code first.

Why it should work. The advisor lets Claude "consult a second, typically stronger model at key moments during a task." Anthropic's events page says "For a fixed command or path, use a permission rule." A mod's tool.check hook fires before "the permission prompt, auto mode's classifier, or a PermissionRequest hook," which is why a mod can insist on a person. Per the 2.1.304 changelog entry, decider: "person" means only "the permission dialog or a host's callback can approve the ask, and it is refused with the mod's reason where nobody can be asked," so an unattended job gets that call refused.

Traps.

  • Cost. Each consult bills "at the advisor model's rates in addition to your main model's usage" (Opus, with --advisor opus) and counts toward plan limits, and "There is no setting to cap or force advisor calls." "Each advisor call processes the full transcript anew, with no reuse between calls." Our reading: long jobs cost most.
  • Anthropic lists "tasks you want independently checked before Claude declares them done" as a use, but the advisor reads the same conversation and Claude decides when to call it. It is a consultant, not a gate.
  • "The hook throws or times out: the call also gets the decision it would get without the mod." Ask for the .catch handler Anthropic documents, which returns a deny. If the handler itself fails, the call is decided as if the mod were not there.
  • "A mod is code that runs with your permissions," and after three worker crashes "Claude Code unloads every mod that isn't built in." Never make a mod your only protection.

#6. Give Claude a job title in ten lines

Needs. Claude Code in the terminal, v2.1.269 or later for /output-style, per Anthropic's output styles page.

Steps.

  1. Ask Claude to write a style file at ~/.claude/output-styles/front-desk.md: plain words, answers under 80 words, no jargon, end with the next step for the customer. Leave out the keep-coding-instructions field: "Leave it out if Claude won't be doing software engineering."
  2. Restart Claude Code. It "reads style files when it starts."
  3. Run /output-style front-desk, ask a question, then run /output-style concise and ask the same one. Keep the answer you would send to a customer. Run /output-style default to go back.

Why it should work. "A custom style can also turn Claude into something other than a software engineer, such as a writing assistant or a data analyst." The built-in Concise style makes "Responses lead with the result and leave out preamble, narration, and recaps."

Trap. The command has a hyphen: /output-style. And "An output style gives Claude instructions to follow. It doesn't guarantee that something always happens or never happens." A style sets tone, not facts: check any price, policy or date before a customer sees it. For a rule that must hold every time, Anthropic points to a hook.

#Where to start

Pick the item that shows you something you cannot see today: a cost report, a restated task, a five-line note, a query, a click. Related field notes: Claude Dashboards and Motion launched in beta covers item 4, Anthropic's new project coordinator is real checks Projects, and Installing an agent skill is running untrusted code covers what to check before you add a skill.

#What we did not check

  • Whether the desktop app saves a side chat or lets one message the main chat. The desktop page says neither, but a 2026-10-15 Desktop release note describes a new side-chat pane, and we did not run the app.
  • Whether claude --model sonnet --advisor opus pairs on your install, or whether a rule or mod can hold calls to your own email connector. Anthropic's example is GitHub.
  • Your plan's data terms. We read none.

#Sources

All read October 25, 2026, Central time. Page dates are UTC.

Keep reading