Lindy AI Playbook: How to Set Up, Teach and Run Your AI Teammate

Lindy (often searched as Lindy AI) is taught in plain language, not wired step by step. You set it up by installing it in Slack and connecting email and calendar, then you teach it with four parts: memory files that hold context, skills it picks automatically when a request needs them, routines that run on a schedule or an event, and Watch and Learn, which turns a screen recording into a skill. Approvals sit over all of it: by default, anything with outside impact waits for an OK. The setup is quick. The real judgment is in what you let it do alone.

Day one: install it, link people, connect inboxes

Setup runs in a fixed order. An admin adds Lindy to Slack, each teammate links their account with one tap, then everyone connects Google or Microsoft email and calendar, and finally a phone number is added. There are no API keys and no IT ticket. One trap catches new teams: a teammate who signs up separately at lindy.ai lands in a separate workspace without the team's skills, files or meetings, so they must accept the invite instead. Keep in mind that every person who uses Lindy, including anyone whose first @mention happens in Slack, becomes a billed seat.

Slack decides what you get. New Slack installations need a paid Slack plan, and the shared teammate lives there: in a channel it works from public sources and approved team material, while each person's DMs hold a private Lindy with the full toolset of calendar, email, SMS, reminders and meeting reads. Teams that do not use Slack get the personal assistant, not the shared teammate, and Lindy's docs list Microsoft Teams as coming soon. Both Lindys draw on one credit pool, so a busy channel spends faster than one person's DMs.

Teach it once: memory, skills and Watch and Learn

Memory files

Lindy's memory, which it calls Lindy Brain, is a set of plain files you can open, edit or delete. Every save becomes a version with full history and restore, and deleting a file keeps its earlier versions ^[1]. You can drop in up to 100 files at once. Three scopes control access: Personal files are private, Team files are readable by everyone but changed only by owners and admins, and System files are built in and locked. Home also lists recent learnings, which you can add to, edit or delete. Client facts and tone rules belong in Team memory.

Skills

A skill is a job described once. Lindy picks it automatically when a request needs it, so nobody has to remember a command. Lindy ships 40+ built-in skills from a library, and you can customize them. Skills come in three kinds: personal (private), team (shared, managed by admins) and built-in. Each is stored as a SKILL.md file in a .skills folder, and a personal version overrides a team one, which overrides the built-in. One detail shapes how you write them: Lindy reads only skill names and descriptions until one looks relevant, then loads the full instructions. A vague description makes it less likely that Lindy picks the skill when you need it.

The October 2026 creative skills show how far a skill can reach: talking-head videos from a still image and a voice track, UGC-style video ads and interactive dashboards. Even for skills that large, Lindy loads the full instructions only when a request needs them. Internal tools that save data and live on their own link are a separate build path, covered in our guide to Lindy Apps.

Watch and Learn

Some tasks are easier to show than to describe. Watch and Learn takes a screen recording or written steps and turns them into a personal skill plus a proposed routine; you can then ask for the task anytime or schedule it. Recording works in Chrome, Edge or Safari on a computer. The limits are specific: video in MP4, MOV, WebM or M4V; documents in PDF, TXT or Markdown; 2 GB per file; one hour of video in total; notes up to 4,000 characters ^[2]. Recordings stop at 60 minutes. The analysis itself uses credits, so trim the recording to the task.

Put it on a clock: routines

A routine has three parts: a trigger, a plain-English prompt and a destination. Triggers are schedules (daily, weekly or monthly) or events such as an email, a Slack message, a calendar event or the start or end of a meeting. New routines default to daily at 9:00 AM. Personal routines are open to anyone. Workspace routines can only be created by admins, need the team product set up, and run for the team through Lindy in Slack. Built-in routines cover daily briefs, email drafting and labeling, urgent email alerts, follow-up bumps, meeting recording, meeting prep and meeting scheduling, each toggled per connected account.

The inbox routines get their own walkthrough in our Lindy AI assistant guide, and the scheduling ones in the scheduling assistant guide. Here the point is the controls. A dry run fires a routine on real data and shows what it would have done, without sending, posting or writing anything ^[3]. Run one before every new schedule. Run this as a script turns a personal routine into Python that can finish without waking Lindy; it does not work for workspace routines, and the script carries none of Lindy's credentials to Gmail, Slack or Calendar.

A personal routine can also have Lindy call someone, but only between 8 AM and 9 PM their time and within the phone line's limits. The phone line needs a paid plan, calls use credits from 25 a minute, and the line's number carries its own monthly fee. The credit cost of a routine depends on how often it fires and how much work each run does; a daily brief uses very few. A workspace routine spends for the whole team each time it fires, and Lindy's usage docs rate team routines Moderate, against Low for simple lookups. You can switch a routine off without deleting it. Lindy's docs also say every message in a thread re-reads everything above it, so long threads cost more credits with each turn.

Decide what it may do on its own

By default, anything with outside impact (sending an email, updating a ticket, posting to another channel, publishing a doc) waits for approval, while read-only lookups from approved sources do not. Approvals arrive as Approve and Deny buttons in Slack, which the requester or any owner or admin can press. Per integration you then choose Always allow, Require approval or Do not offer . Reads are never guarded. Lindy describes the model as trust that grows: actions wait at first, and once trusted Lindy can run a job end to end. Email follow-ups, for example, can be switched to auto-send.

Know where the fence stops. Guardrails apply in shared Slack threads, not in web chat, Slack DMs, iMessage or SMS. One guardrail covers every account under an integration, so work and personal Gmail cannot yet have different rules. MCP connections carry guardrails per person, and saved credentials sit outside guardrails entirely. Enterprise adds named approvers. Lindy's own docs warn that it can make mistakes and that users should review actions before relying on them, which is the practical case for keeping email and ticket writes on approval until dry runs and approved runs look clean.

Models and plan gates

Lindy picks a model per task, or you choose one, and on paid plans the chat menu lets you set a default. Models come from Anthropic, OpenAI, Google and other frontier providers. Two playbook moves depend on the plan. Computer use, where Lindy drives a browser like a person for tools without an integration, is only on Pro and Max, and live onboarding starts on Pro. Where none of the 3,000+ integrations fits, Lindy can also plug into a hosted MCP server or write its own integration against the tool's API.

A worked example: the project manager's week

Lindy's project manager role page shows how the parts fit together. A weekday 9:00 AM routine reads project boards, Slack and Gmail and posts a standup. An event trigger after each call extracts commitments from the transcript. A voice memo sent to Lindy is transcribed and turned into task assignments, date changes, emails and calendar events, and Lindy also chases approvals. The page lists monday.com, Asana, ClickUp, Google Calendar and Google Docs as the connected tools. In playbook terms that is two routines, a voice intake channel and a set of write actions that should start on approval.

Build it in that order. Put project names, owners and status rules into Team memory first, so the standup has context. If the standup should run for the whole team as a workspace routine, only an admin can create it. Dry-run it against a real morning, then switch it on. Leave channel posts and calendar changes on Require approval until the drafts match what you would have written. Customer stories show the pace some teams report: Ankor's CEO built ten agents in the first week and says the company now runs 20 to 30 active ones, and Interlaced built its first Lindy in under a week with hands-on onboarding. Those are the customers' own figures.

Your first routine, in four moves

Pick one recurring job that eats an hour of someone's week today. Write the context it needs into Team memory, then either describe it as a skill with a sharp name and description or record it once with Watch and Learn. Schedule it as a routine only after a dry run on real data, and keep every outside-impact action on approval for the first cycles. Then check the Credits column in Settings, under Members, to see what each person's Lindy actually draws before you add a second routine.