AI agents for operations: the skills, the job and the agent
Most operations teams use AI one question at a time: summarise this meeting, rewrite this email to a supplier, tidy up this checklist. The weekly status is still built by hand, from five boards and twenty Slack threads.
There is a ladder out of that. A skill is one task, written down in detail. A few skills in order make a job. That job running by itself, with a person approving before anything leaves the house, is an agent. Here is the ladder for operations: status, risks, meetings, suppliers and routines.
- 01
6 skills
One task each, described in detail.
- 02
Connectors
Reach into the systems the tasks need.
- 03
One job
The weekly project status
- 04
One agent
The status agent
Claude skills for operations: the 6 to write first
weekly-statusWeekly project status
Use when: someone asks for the weekly status, asks where the projects stand, or prepares the Monday meeting.
Reads the open issues and the project channels, gives every project green, amber or red by fixed rules, and lists what is late, what is blocked and what needs a decision.
Output: One page with a line per project, the blockers and at most three decisions.
risk-and-blocker-logRisks and blockers
Use when: a project is late or stuck, or a steering meeting is coming up.
Collects overdue issues, tasks without an owner and work waiting on someone else, and writes each one as a risk with impact, owner and next step.
Output: A risk list sorted by impact, every line with an owner and a date.
meeting-to-tasksMeeting notes to tasks
Use when: a meeting is over and the notes or the transcript are in.
Pulls out the decisions and the action items, takes owner and due date only from what was said, and marks every item where nobody was named.
Output: A list of decisions and a list of tasks ready for the project tool, with the gaps marked.
supplier-follow-upSupplier follow-up
Use when: a supplier is late with a delivery, a quote or an answer.
Reads the thread, states what was promised and by when, and writes a short, polite reminder with one clear ask and a new date.
Output: A draft reply per supplier, waiting in Drafts for a person to send.
process-docProcess document
Use when: someone explains how a routine is done and it should be written down, or a new colleague needs to learn it.
Turns the explanation into purpose, owner, what starts the routine, numbered steps, the systems used and what usually goes wrong, in the company’s template.
Output: A one-page process document a new colleague can follow on day one.
project-kickoffProject kickoff
Use when: a new project has been approved and needs a plan.
Asks for goal, scope, deadline, owner and who is involved, flags what is missing, and breaks the work into milestones and the first two weeks of tasks.
Output: A kickoff page with milestones, owners and the first tasks to create.
What a good skill looks like
An example written for this page, following Anthropic’s guidance for writing skills. Adapt it to how your team works.
---
name: weekly-status
description: Writes the weekly status for all active projects from Jira, Confluence and Slack, showing what is late, blocked or needs a decision. Use when someone asks for the weekly status.
---
# Weekly project status
## What you deliver
One page, four sections, in this order: Summary, Projects, Blockers, Decisions needed.
## Steps
1. Collect: every open issue in the active projects with status, priority, assignee and due date (Jira), each project's next milestone and its date from the plan page (Confluence), and the last seven days in each project channel (Slack).
2. Give every project a colour. Green: nothing overdue and no open blockers. Amber: overdue issues or an open blocker. Red: a blocker older than five working days, or a milestone date that has passed.
3. Summary: three sentences. How many projects are green, amber and red, the biggest risk, the one decision needed.
4. Projects: one table. Project, colour, owner, next milestone and date, what changed since last week.
5. Blockers: one line each. What is stuck, who is waiting for whom, since which date.
6. Decisions needed: at most three, each with who decides and by when.
## Rules
- Use only what Jira, Confluence and Slack give you. List anything missing under "Missing data".
- An issue with no assignee or no due date is a gap, not green. Write "No owner" or "No date". Never guess a name or a date.
- If there is no status from last week, write "First week, no comparison" instead of a change.
- Write for the management team: no issue keys in the summary, plain words only.
- Run risk-and-blocker-log on every red project before you hand it over.
## Example summary
Seven projects: four green, two amber, one red. Red is the warehouse move: the shelving supplier has not confirmed delivery, and 6 of 14 issues wait on it. Decision needed: keep waiting, or book the backup supplier by Wednesday.
- The description names the moment
- Until a task matches, Claude sees nothing but the name and the description. That is why the description lists the real moments: someone asks for the weekly status, asks where the projects stand, or gets ready for Monday.
- Colours with rules
- Green, amber and red are defined in the skill, so a project turns red because of a fact, such as a blocker older than five working days. Two people running it on the same data get the same colours.
- A gap stays a gap
- An issue with no owner is often the reason a project slips. The skill marks it instead of filling in a name, so the gap comes up in the meeting rather than three weeks later.
From skills to a job: The weekly project status
Today a project manager spends a few hours every week chasing updates and pasting them into a slide. With the skills and the right connectors, one person does it in half an hour, in Claude or ChatGPT.
- 01Ask the AI for the open issues in every active project from Jira, the next milestones from Confluence and this week’s blockers from Slack.
- 02Run weekly-status on them.
- 03Run risk-and-blocker-log on every project that comes out red.
- 04Read it and add what only a person knows, such as a deadline the customer moved on a call.
- 05Share it before the Monday meeting.
From job to agent: The status agent
Every Monday at 07:00.
- 01Fetches every open issue in the active projects, with status, priority, assignee and due date. via Jira MCP
- 02Reads each project’s plan page for the next milestone and its date. via Confluence MCP
- 03Reads the last seven days in the project channels its owner is a member of and picks out the blockers. via Slack MCP
- 04Runs weekly-status, then risk-and-blocker-log on every red project.
- 05Posts the draft in the operations channel for approval. via Slack MCP
The head of operations reads the draft, changes a colour where they know more, adds what was only said on the phone, and approves it. Nothing reaches the management team before that.
The status on the table for the Monday meeting, every week, in the same format.
In Claude, set it up as a scheduled task: describe the job once, choose weekly, and it runs with your skills, connectors and plugins even when the computer is asleep or the app is closed. Each run is its own session, and scheduled tasks come with every paid Claude plan. In ChatGPT Work, a weekly scheduled task does the same.
With Deployed OS, the status skill and the connectors belong to the company. Every project manager gets the same skill in Claude, ChatGPT and Copilot and connects Jira and Slack with their own login. Admins decide which groups may let the AI write, and every call the agent makes is logged.
Where it goes wrong
- Skipping the manual weeks
- Run the weekly status by hand for three or four weeks before the agent takes over. That is when you find the rules the skill is missing, such as what counts as blocked.
- Colours by feeling
- If the skill does not say what makes a project amber or red, the AI decides, and the same project can change colour between two runs. Write the rules into the skill.
- A board nobody updates
- The agent can only report what is in Jira or Asana. If half the issues have no owner or due date, the first reports will say so. Treat them as your clean-up list.
- Tasks nobody confirmed
- An agent that turns every line of the meeting notes into a task fills the board with duplicates and noise. Let it propose the tasks, and have a person confirm them before they are created.
AI agents for operations: common questions
The recurring work with clear steps: the weekly status across projects, the risk list before a steering meeting, tasks from meeting notes, reminders to late suppliers and process documents. It takes the steps in the same order every time and stops for a person before anything goes to management, customers or suppliers.
Start with the task that eats a few hours every week. For most teams that is the weekly status or turning meeting notes into tasks. Write one skill, test it on three or four real weeks, then write the next.
It can create new ones, not change existing ones. Through Deployed OS, with Edit access, the Jira connector creates issues, the Asana and ClickUp connectors create tasks and the monday.com connector creates items. None of them can change a status, reassign or close anything today. Jira issues and ClickUp tasks are created with nobody assigned, Asana tasks are assigned to the person who connected, and a monday.com item gets its name only. We recommend that a person confirms every new task, and admins decide which groups may let the AI write at all.
Either works. A skill is plain language, and through Deployed OS the same connectors to Jira, Asana, Slack and the rest work in Claude, ChatGPT and Copilot. Pick the assistant the team already uses.
It can write the reminder, not send it. The Gmail and Outlook connectors create drafts and never send, so each reminder waits in Drafts until a person has read it and pressed send. The checkpoint is built in.
It depends on the size of your team. Get in touch and we will walk you through it.
Start with one job.
Tell us which task eats your team’s week. We’ll show you the skills that do it, and the agent that runs it.
