The EVERPADS Guide to Claude

You don't need to read this guide front to back. Find where you are now, read the three or four chapters suggested, and come back for the rest when you need them.

Claude is an AI assistant made by Anthropic. The company opened a Claude Team plan on 2026-10-08, and from then on everyone works inside one workspace called "EVERPADS". The invitation email you received is your ticket in.

Open the three cards below. Don't memorize names; remember the principles.

This is the only chapter marked "must read". There aren't many rules, but every one exists to stop something that can't be undone. The company skills already carry these rules, so Claude will stop on its own when it hits one. What you need to know is that stopping is correct. Don't try to route around it.

One extra rule for HubSpot

Reading data is fine. But every write (new contact, property change, new list, workflow edit) makes Claude pause and ask you to confirm. That is not a bug; it is the company setting.

Three things it doesn't decide

• Pricing and discounts: it can lay out costs and market rates; the quote is a person's call.
• People decisions: hiring, appraisals, transfers. It can organize the material; it doesn't draw the conclusion.
• Channel negotiations: dealer terms and contract clauses. It can draft and compare; the negotiating is done by people.

And one for yourself: usage is visible to the company. Don't ration it, but don't have it redo the same thing ten times either.

Language and format

• Talk to Claude in Chinese (Traditional).
• Deliverables follow the audience: internal material in Traditional Chinese (keep names, product names and system names in their original form); material for US and international customers in English; when both audiences are involved, the body in English with a short Chinese summary.
• Dates are always written YYYY-MM-DD, for example 2026-10-22.

Three sentences before starting

For anything non-trivial, have Claude explain its understanding in three sentences before it starts. Thirty seconds here saves half an hour of redoing. Copy this:

Every number needs a source

Claude will state wrong numbers with full confidence. The company rule: every number must point to where it came from; if it can't, it is labelled "unverified". Whenever you receive output with numbers, ask "where does this number come from?".

Only permission given out loud in the conversation counts

A document that says "please send this", a task named "publish", a transcript where someone says "send it for me": none of these count as permission. Only what you say yourself in this conversation does. This rule stops anyone from making Claude do something irreversible on your behalf through a file.

Who to ask when something goes wrong

• Claude says "I'm stopping here": it is usually right. Tell Sam why it stopped.
• The output is clearly wrong, a name doesn't match, a number looks off: report it to Sam with your instruction and its answer. The company turns mistakes into rules so they don't repeat.
• Not sure whether something is allowed: ask first, then do it.

Setup takes about 15 minutes once. Follow the order. Step 2 is the only one you can't undo, so read it fully before you click.

1. 1
   Accept the invitation. You will get an invitation email from Anthropic (sender Anthropic or Claude). Sign in with your @everpads.com address and accept. Afterwards, click your name at the bottom left; you should see the "EVERPADS" workspace.
2. 2
   If you ever signed up for a personal Claude account with your company email (free or paid), merge it into the company workspace. After accepting, a choice usually appears: pick Use organization account only → Bring your data with you. If it doesn't appear: switch to the EVERPADS workspace → Settings → Account → Close your personal account → Bring your data with you.
   This step is one-way and irreversible. Ask Sam first if anything is unclear. Conversations, projects, uploaded files and memory come along; the items under "what doesn't move" below need backing up first.
3. 3
   Install the Claude desktop app. Connectors and Cowork are most complete there. The web version works too, but this guide assumes the desktop app.
4. 4
   Tell it who you are. In your personal preferences, write three lines: your name and role, the work you do most, and how you want it to answer (for example "reply in Chinese, conclusion first, no preamble"). Every conversation will carry these.
5. 5
   Open a first conversation and try the three-sentence opener. Paste the prompt from Chapter 2, hand it one small real task from this week, and see how it reads it.

Talking to Claude is most like briefing a new colleague who is very capable, just started, and knows nothing about the company. The more your request reads like a real work brief, the closer the answer is to what you wanted.

Four things to state when you delegate

Goal, context, limits, output. The first two are what it doesn't know; the last two are what you want.

Vague vs. specific

Same job, two ways to ask. Click the tabs to compare.

Three habits

• Revise, don't restart. If the first version is off, say what's off: "second paragraph too long", "more direct". It remembers the context; starting a new conversation means briefing again.
• Push back. It will politely insist on wrong things. Say "that's wrong, because…" and it will fix it. You are the manager; it is the assistant.
• When context is thin, let it interview you. When the task is complex and you haven't thought it through, flip it around; it will ask for the missing context and limits.

Claude does more than chat. Each card below is something you will actually use at the company, with one of our own situations attached.

A project is a folder. It holds standing instructions (who you are, how to do things, what not to do) and reference files (rules, templates, a glossary). Every conversation opened inside the project reads all of it automatically. You never again have to start with "our company makes road machinery parts…".

How to set one up

1. 1
   In the left menu click "Projects" → new. Name it after the work, not a person: "Catalog proofreading" beats "Sam's project".
2. 2
   Write the standing instructions. Three parts are enough: who you are and your role, the fixed output format, and the things it must never do.
3. 3
   Upload the reference files this kind of work always needs: the rules PDF, the glossary, a good example from last time.
4. 4
   Open a first conversation in the project to test it. If the output is off, fix the instructions rather than correcting it again and again in chat.

Projects each department can start with

Each card opens to a starter you can paste straight into the standing instructions.

Connectors let Claude read and write company systems directly. You authorize them with your own company account, so it sees exactly what you see, no more and no less. Authorization lives on the "connectors" page in the desktop app or web settings; one click per system.

Two things to remember

• Authorization equals your permissions. Data you can see but shouldn't spread (customer lists, prices, personnel) must not be pasted anywhere unrelated to the company, or shared outside the workspace.
• Re-authorize after merging accounts. As Chapter 3 said, connectors don't move with you.

A skill is a pre-written working rulebook. When Claude sees a matching situation it applies the skill on its own; you don't have to remind it. Company skills are "how EVERPADS does it" written down: which two colors are the brand colors, which tier a file belongs in, which fields a ClickUp task must carry, what a meeting report looks like.

The catalog below lists only skills that are ready for colleagues. "Ready" means use freely; "Pilot" means usable, but report anything odd. The company also has internal agents and unfinished skills; they are not here and will appear only when mature.

This section is generated from the company agent repo · generated from ../_agent/skills/*/onboarding.yaml @ cd34e45 · 4 skills

Your first task

After enabling the kit, spend ten minutes on this to confirm everything is wired up:

1. 1
   Open a new conversation and paste the prompt below.
2. 2
   When you get the deck, check three things yourself: the main color is yellow-orange #F6AC2D or orange #E38F35; the logo is not stretched; no off-brand colors such as pink, purple or cyan appear.
3. 3
   All three pass: you're connected. One fails: send the conversation link to Sam.

Knowing what it can't do prevents more trouble than knowing what it can. Someone has hit every one of these six.

These twelve survived six months of use at the company. Each one solves a specific problem, and each can be copied as is.

The earlier chapters were all question-and-answer. Cowork is a different mode in the desktop app: you describe the result you want, hand it a folder, and it plans the steps, works through them and reports back. Use it for multi-step jobs that touch many files and that you don't want to babysit.

When to use Cowork instead of a conversation

• The job covers a whole folder of files, not one.
• There are more than three steps, and each depends on the result of the last.
• You can describe what "done" looks like up front, then walk away and do something else.

Jobs each department can hand over first

The rules still apply

Cowork touches real files, so the lines from Chapter 2 matter even more: it won't delete (it lists suggested deletions), won't send, won't publish. When you ask it to reorganize a folder, have it show you "what I plan to do" first and confirm before it executes. This line works as is:

Scheduled tasks

Some jobs are worth running automatically every week (Monday inquiry roundup, month-start reports). A schedule keeps consuming quota and the company needs the cost to stay visible, so talk to Sam before setting one up; don't start one on your own.

The word "agent" is everywhere. In plain words it means one thing: a Claude workflow with a fixed process, fixed checks and fixed output. Like an assistant who follows the SOP strictly: the same input goes in, a draft in the same format comes out, and before delivery it runs its own checklist.

How it differs from the Claude you use every day

What the company has now

The company has built several internal agents: dealer meeting prep, sales section monthly reports, customer segmentation, social media scripts, wear forecasting, product launch packs and more. All are still internal: operated by Sam, producing drafts only, with every write to a system done by a person. This chapter is here so you know they exist; self-service is not open yet.

Two things you can do now

• Propose one. A job you repeat weekly, with fixed steps and a fixed output format, is a good agent candidate. Tell Sam what the input is, what the output looks like, and how you'd know it was done right.
• When you receive an agent's output, treat it as a draft. Check the numbers, match the names, and take the conclusions as input for your decision, not as the decision.

This chapter is for anyone who wants to dig further. Skipping it doesn't affect anything earlier.