To turn weekly sales exports into a report, Claude needs the files, instructions for calculating the totals, and an output you can review. Each feature handles a different part of that job. Choose the feature by what is missing.
Nate Herk’s Every Claude 2.0 Concept Explained for Non-Coders groups Claude’s features by models, apps, outputs, saved context, and independent work. “Claude 2.0” is his nickname for recent changes, not a model name. The distinctions below follow that grouping, with product details checked on October 30, 2026.
Austin Marchese’s 7 NEW Claude Features To Get You Ahead of 99% of People Using AI adds practical uses for Google Workspace, dashboards, mods, and remote sessions. Those additions were checked against Anthropic’s documentation on November 5, 2026.
01Separate the settings
Model choice and effort answer different questions. Haiku, Sonnet, Opus, and Fable are model names; effort controls how much work the chosen model spends on a response. Anthropic’s model settings let you change them separately.
| Term | What it controls | What to check |
|---|---|---|
| Model | Which model performs the task | Whether it can produce an acceptable result on your actual work |
| Effort | How much work that model spends responding | Whether more effort improves the result enough to justify the time and usage |
| Token | A unit of text processed by the model | Input as well as output contributes to usage |
| Context window | How much information a model can work with at once | Files, instructions, conversation, and room for the response |
| Usage limit | How much work your plan permits over time | Your current allowance and reset time |
Herk’s seven-workflow comparison favored Sonnet for clearly specified tasks and Opus for vague planning work. Treat that as a reason to compare models on your own tasks, rather than assuming a more expensive model will win.
Herk also reports a comparison on one task where extra-high effort beat max: roughly 90 minutes and $26 versus 150 minutes and $50, with bugs remaining at max. Those are his reported results for that task. They do not establish the best setting for yours.
Anthropic’s effort guidance starts at the model’s default and adjusts based on results. Ultracode also coordinates work across agents; it is a separate setting, not another effort level.
For a hard decision within a larger task, Claude Code’s advisor tool lets the main model consult a stronger model. Recheck your model choices each release covers comparing settings on work you can judge.
Try a cheaper model on a copy of work you have already checked. Marchese suggests testing Haiku 5.5 on your own task. For the sales report, compare its totals, missing-data warnings, and drafting time with your current model. Count retries and corrections when comparing cost.
As of November 5, 2026, Haiku 5.5’s standard API rates are $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Above that threshold, the rates are $0.50 and $2.50. These rates apply when software sends requests through Claude’s API. Your Claude app subscription has a separate allowance.
A large context window does not give unlimited usage. Anthropic’s limits documentation distinguishes context capacity from shared usage across Claude products. Most paid plans have five-hour and weekly limits; verified Ultra accounts are exempt from the five-hour limit. Context capacity varies by model and product.
Check Settings > Usage before buying more capacity. Usage credits pay for work beyond included limits. The monthly API credits included with Max and Ultra are separate and do not pay for extra usage inside Claude or Claude Code. Stop hitting your token limit covers reducing unnecessary input.
02Choose where the work runs
| Place | Use it for | Check first |
|---|---|---|
| Claude chat | Questions, research, and documents using supplied or connected material | Whether the needed files and tools are available |
| Claude Code | Work in a project folder, including reading files, editing them, and running commands | Which folder and actions the session can access |
| Cowork | Delegated work across files and tools | Your current interface: Cowork and chat are merging |
| Duet | Ongoing help through connected apps, such as an email and calendar brief | Waitlist access, connected accounts, and background usage |
Anthropic announced the Cowork and chat merger on September 16, 2026. Herk’s separate categories may not match the tabs on your screen. Choose by available capabilities. A local Code session is useful for the sales report when the exports already live in a folder on your computer.
A session is one conversation and its working context. Several sessions can work in parallel. Give them distinct outputs or coordinate edits so two sessions do not overwrite the same file. The desktop documentation covers parallel sessions and communication between them.
Google Workspace. When the report already lives in Google Docs, Sheets, or Slides, use Claude in the file. The Workspace add-on can work on selected text, cells, or slides and use your Claude skills and connectors. Sheets supports native formulas and charts; Slides can use the deck’s layouts and themes. This is the practical benefit behind Marchese’s Workspace example.
Install the add-on from Google Workspace Marketplace, then open a file and choose Extensions > Claude > Open Claude, following the setup guide. It is a paid-plan beta, and your organization may restrict installation. Start with Ask before edits so you can inspect proposed changes.
The separate Google Docs, Sheets, and Slides connectors let you create or edit Google files from Claude. Choose that route when starting a new report or combining several files. Choose the add-on when reviewing a file already open in Google Workspace.
Remote Control. To continue a local Code session from your phone, enter /remote-control in that session and open its link in Claude on the other device. Remote Control needs an eligible paid subscription and may require organization approval. The files, tools, and running process stay on your computer.
Marchese pairs phone access with keeping the computer awake. Current desktop guidance places Keep computer awake under Settings > Desktop app or Settings > This computer > System. Keep the app running and the computer online; closing a laptop lid still puts it to sleep. Before leaving a long task, send a harmless message from your phone and confirm that the local session receives it.
A side chat asks about the current session without adding its question and answer to the main conversation. In the Windows desktop app, use Ctrl+; or /btw. Side chats are not saved to disk, so move any decision the main task needs into the main conversation. Improve your Claude Code workflow covers when a side question belongs there.
Duet is a US research preview accessed through a waitlist on Pro, Max, and Ultra. It works on its own cloud computer and is designed to ask before consequential actions. Its background work consumes your plan’s usage. Review drafts and approval requests; access to an app does not prove the result is correct.
03Choose the output
An artifact is an output you can open and work on beside a conversation. It is different from the model or app that made it.
| Output | Choose it when you need |
|---|---|
| Docs | A written report that people can read and edit |
| Slides | A presentation you can deliver or export |
| Design | A visual layout, screen, or mockup |
| Dashboards | Charts connected to data that changes |
| Motion | An animated explanation you can export as video |
Anthropic’s October 8 announcement puts Docs, Slides, and Design on every plan, including Free. Dashboards are a paid-plan beta; Motion is a Team and Enterprise beta. Organization settings can affect access.
For a weekly sales report, choose a document if someone needs to read commentary and a dashboard if they need to explore refreshed figures. A dashboard’s query is the instruction that selects and totals its source data. Marchese’s dashboard demonstration shows why access to that calculation helps you check the result.
Click a dashboard number to inspect its query, or ask Claude to explain it. Check:
- Which source rows, dates, and refund rules produced the total.
- Whether your own calculation from those rows matches it.
- When the chart last refreshed, and whether newer records are missing.
The query can still use the wrong definition of revenue. Agree on that definition before treating the chart as a business report.
For written commentary, ask for a Docs artifact and select the passage you want changed. Marchese uses selected-text edits for longer documents. Check the revised passage and its surrounding text before accepting it. If collaborators already work in a Google Doc, use the Workspace route above.
For a deck, give Claude the claim the audience needs to understand and any brand files or reference slides. Marchese recommends an original idea, design references, and several versions. Ask for five opening-slide directions, choose one, then build the deck from it. Review a visual plan before you build covers comparing directions before completing the work.
Check delivery before spending time polishing a deck. Current sharing rules depend on the artifact and access method: Docs and email invitations require a Claude account; most artifacts shared with Anyone with the link allow viewing without one. Organization rules can restrict sharing, and connected data needs its own access. Test the intended recipient’s access, or export to PowerPoint, PDF, or editable Google Slides and inspect the exported file.
Motion animates text, charts, shapes, and images with code and exports MP4; it does not generate video footage.
Design now works inside Claude conversations. The standalone Design site is due to close on December 14, 2026. Follow Anthropic’s Move from standalone Claude Design to Claude for existing work. The older Move design work into Claude Design guide describes the earlier interface and the design workflow.
04Save the right thing
Instructions, procedures, and access solve different problems. Herk’s skills explanation starts with a repeated request that becomes a saved procedure.
“A skill is a saved set of instructions for one specific job”
Nate Herk, “Every Claude 2.0 Concept Explained for Non-Coders” · 13:44| Feature | What it saves or adds | Sales-report example |
|---|---|---|
| Auto memory | Notes Claude writes for later sessions | A correction about how your team names a region |
CLAUDE.md | Standing project instructions | Where exports live and which definition of revenue to use |
| Skill | A reusable procedure in SKILL.md | Calculate the weekly totals, compare periods, and draft commentary |
| Connector | Tools for a connected service, often using MCP, the Model Context Protocol | Read the approved sales data |
| Plugin | A package of extensions | Distribute the reporting skill and related tools to a team |
| Mod | Code that changes Claude Code’s tools or interface | Add a panel showing report progress |
Claude Code memory carries notes across sessions by reading saved files. It does not retrain the model. Keep CLAUDE.md short and put detailed procedures in skills, whose full instructions load when needed. Review saved notes when a correction stops applying. Audit what your agent saved to memory covers that check.
A connector provides access; a skill explains how to use that access for your task. A plugin packages extensions for reuse. Anthropic’s extension overview explains those boundaries. Write smaller Claude skills that work covers saving repeated instructions, and One skills library for your whole team covers sharing them.
Install a mod with the same care as other software. It runs with your account’s permissions and can read files, start programs, and make network requests. A friendly panel does not limit what its code can do. Use an existing skill or connector when that already meets the need.
Choose a mod for a repeated problem you can name. Marchese’s examples are a usage display and a reminder to save corrections. A usage display should help you decide when to shorten a session; compare it with Settings > Usage before treating it as a measure of your remaining allowance. A correction reminder should show the proposed rule for review so a one-off request does not become a permanent instruction. Turn repeated fixes into rules covers deciding what deserves to persist.
If you keep a mod, install it as a plugin and confirm that it loads in a fresh session. Its availability can differ between the terminal, desktop, and remote interface.
05Control independent work
Permission modes determine how Claude Code approves actions. Current documentation lists these modes:
| Mode | Approval behavior |
|---|---|
| Manual | Prompts for actions that need approval |
| Accept edits | Allows file edits and common file commands in permitted directories |
| Plan | Explores and proposes work without editing source files |
| Auto | Uses a background classifier for actions that would otherwise need routine approval |
| Don’t ask | Denies actions that would require a prompt, subject to documented exceptions |
| Bypass permissions | Skips most permission prompts |
Auto is not a guarantee that every action is safe. Use permission rules to control access, and inspect the work. Writing “never send email” in CLAUDE.md does not remove a sending tool. Assume your agent uses every tool it has explains that distinction.
The features that organize independent work also differ:
| Feature | Job | Limit to remember |
|---|---|---|
| Subagent | Handle a bounded task in separate context and return results | Its requests still consume usage |
| Code project | Coordinate ongoing work through separate task threads | Cloud threads need their own available files and tools |
/goal | Keep a session working toward a completion condition | Completion depends on the evidence in the conversation |
| Scheduled task | Start work at a later time or on a supported trigger | Local and cloud runs have different access and runtime requirements |
Subagents keep research out of the main context; they do not make it free. Give a report-checking subagent the draft report and the totals you calculated from the exports. Ask it to identify any differences. Delegate to your agent in stages covers reviewing work before expanding responsibility.
Code projects coordinate a stream of tasks, usually through cloud sessions. That is different from the classic chat project that groups conversations with shared instructions and files. A cloud project does not automatically inherit your local skills, settings, or tools. Check its configuration before moving a working process there.
Coordinating projects are a staged public beta on Pro, Max, and Ultra; they are not yet offered on Team or Enterprise. Use a coordinating project when work keeps arriving for the same purpose. For reporting, one thread could investigate a data discrepancy while another drafts commentary from totals you have approved.
Marchese’s project example describes a coordinator that starts task threads and reports which need your input. Check the results each thread returns before combining them into the report.
With /goal, specify a result and the evidence for it. The evaluator reads what the session reports; it does not independently inspect your files. A goal can end as impossible or stop on an error requiring your action. Give your coding agent a finish line covers defining the evidence Claude must report before calling a task complete.
Scheduled tasks start a new run. Local tasks need the computer awake and the app open. Cloud scheduled tasks, formerly called routines, can run while your computer is off. They need access to their inputs in the cloud. Test the scheduled run itself, as described in Keep Claude automations working.
06Try one complete report
Herk’s starting advice is a project instruction file, one useful skill, and a clear goal. Start with sample exports and a report you can check before connecting live accounts or adding a schedule. Set up Claude Code in seven steps covers the broader setup.
Use this original prompt in a local Code session:
Read the sample sales exports in this folder. Ask me how revenue, refunds, and the reporting week should be defined if the files do not make those rules clear.
Create a draft weekly report showing the source files, reporting dates, revenue total, and any missing or conflicting data. Show the calculation behind the total. Keep the source files unchanged. Do not connect accounts, send the report, or create a schedule.
Finish when the draft and calculation are saved and you have compared the reported total with the source rows. If a missing definition prevents a reliable total, stop and ask me.