Using the tchop MCP well
Prompts, habits and limits that get better results from your AI assistant once the tchop MCP is connected.
Once the MCP is connected, your assistant can do most of what an admin or editor does in the dashboard. How much time that saves depends less on the tools than on how you ask.
This page covers the habits, prompts and limits that turn a twenty-minute job into a five-minute one. For connecting the MCP in the first place, see MCP Server. You will get your access tokens from your tchop team.
Start by pointing the assistant at the built-in playbooks
The tchop MCP ships with more than tools. It carries its own guidance layer: playbooks that explain how the platform is meant to be used. Channel and mix structure, card types and their settings, feed rhythm, push timing, comment moderation, analytics, plus operating profiles for internal comms, news publishers and fan communities.
Two tools expose it, list_knowledge and read_knowledge. Most assistants will not open them unless you say so.
Open a working session with something like this:
“List the tchop knowledge documents. Read the index, then the ones that apply to curating a news feed. Tell me the rules you will follow before you touch anything.”
This single habit removes most of the rework. An assistant that has read the feed playbook varies card types on its own, keeps the top of the feed fresh, and stops asking you which teaser style to pick.
Use names (or IDs)
You do not need to look anything up. Name the channel you want to work in, name the mix, and for an existing card give the title. The assistant resolves the rest.
“In the Company News channel, add this to the Weekly Briefing mix.”
“Find the card titled "New shift plan from January" and turn comments on.”
Two things are worth doing anyway. Name the channel in every prompt once you have more than one, because the assistant will otherwise pick for you. And at the start of a session, ask for the lay of the land so you both use the same words:
“List our channels, the mixes in each, and the tags we have. Just the names.”
If two mixes have similar names, the assistant will ask which one you mean. Answer with the name.
Give it a job, not a click
The dashboard takes instructions. The assistant takes intent, and does better work with it.
Weak:
“Create 8 article cards in the Market Briefing mix.”
Better:
“The Market Briefing mix in Company News is our weekly read for plant managers. Create one card per URL below. For each, write a one-line summary of what the piece actually reports, in our voice, not the site's meta description. Vary the teaser style so the feed has rhythm. Tag each card with 2 to 3 tags from the list above. Skip anything older than 12 months and tell me what you skipped.”
The second prompt also fixes the most common quality problem in a seeded feed. Auto-detection fills a card's summary from the source page's meta description, which on most publisher sites is one sentence repeated across the whole site. Ten cards carrying it look broken.
Sample prompts by job
Orientation, first session
“Read the tchop knowledge index and the generic playbooks. Then look at our org: channels, mixes, tags, and the last 20 published cards. Tell me what you see about how this app is structured and where it is thin.”
Curating a feed
“Our feed has gone quiet. Pull the last two weeks of content and the card type performance. Suggest five cards that would refresh the top of the feed, then create the two I approve.”
Bulk seeding
“Here are 12 URLs for the Engineering mix. One card each, published, each with its own written summary and 2 to 3 tags. Report back which ones came through without an image.”
Moderation
“Show me all comments from the last 48 hours across the Community channel. Flag anything that needs a reply from us, and highlight the three best ones so they sit at the top.”
Push
“Draft a push for the card titled "Works council results are in". Keep it under 12 words, say what is actually inside, and deep-link to the card. Show me the draft. Do not send it yet.”
Analytics
“Compare mix performance for the last 30 days against the 30 before. Which mixes lost reach, and which card types are carrying the feed?”
Events
“Create a draft event for the works council meeting on 14 November, 10:00, in the Company News channel, capacity 80, waitlist on. Do not publish it.”
Habits that save rounds
Name the channel in the prompt, always. It is the one thing worth repeating. The assistant defaults to the org your credentials belong to, and it will guess a channel if you leave it out. A mis-targeted write is annoying to unwind. If you run more than one org, name that too.
Ask for the plan before the writes. "Show me what you intend to create, then wait." Reviewing a list of twelve planned cards costs a minute. Deleting twelve wrong ones costs more.
Ask for deeplinks back. Every created card can return a link straight to the object, so you can look at what the assistant made instead of trusting the summary of it.
Work in drafts while you iterate. Cards, events and pushes all support a draft or scheduled state. Build in draft, look at the result in the app, publish when it reads right.
Do one mix at a time for large seeding jobs. Parallel work across mixes collides on shared org-level things like tags, and you end up with "Nat Cat" and "NatCat" in the same list.
Let it own the technical detail. Enum casing, field names, which parameter a card type needs: you do not have to know any of it. Describe the outcome in plain language and let the error messages guide the assistant. They are written to say how to succeed on the next call.
Check the app, not the response. A successful write is not evidence a reader sees anything. Open the web app after a structural change. Navigation labels, right-rail widgets and fonts are only provable on screen.
Do and don't
Do
Ask the assistant to read the knowledge playbooks at the start of a session.
Give editorial intent, not just parameters. One line about what a mix is for changes every summary it writes.
Say "show me first" before anything that reaches readers.
Keep an approved tag list and hand it over, rather than letting the assistant invent tags.
Review pushes yourself. Every one spends a bit of your audience's patience.
Don't
Don't paste your auth token into a chat, a shared doc or a screenshot. All permissions derive from it.
Don't let an assistant send a push without your sign-off. There is no unsend, and it reaches the lock screen.
Don't treat a permission error as a broken tool. The MCP enforces the same roles as the app, so an error usually means the connected user is not allowed to do that.
Don't run large deletions through an assistant. Deletes are irreversible and a misread instruction is expensive.
Don't ask for one card at a time. Batch the job, then review the batch.
Don't skip the summary field on link cards, and don't accept the auto-filled one without reading it.
Checking work on a newly built app
If you have just stood up an app or seeded a lot of content, ask the assistant to run audit_prototype on your org. It counts the things people miss by eye: untagged cards, thin or unpublished mixes, repeated summaries, cards with no image, audio cards with no audio file, default tab labels. It is aimed at new builds rather than day-to-day editing, so expect noise on a mature app.
Good to know
The MCP is in Beta and the tool set grows. Ask your assistant to list its tchop tools if something you remember reading about is not working.
Everything the assistant does happens as you. Actions land in the same audit trail, respect the same roles, and stay visible to your team in the dashboard.
The full operation reference lives in the Developer Portal at api.tchop.live. If a workflow you need is missing, write to support@tchop.io.
