Turning Repetitive Task Automation into a Specification — the loop-me Skill
機械翻訳 / Machine-translated
Original article: Turning Repetitive Task Automation into a Specification — the loop-me Skill (cccskills)
Skill page: loop-me
Key takeaways from this article
With loop-me, anyone who wants to automate recurring tasks can create workflow specifications simply by answering questions.
The result is a specification that can be handed directly to whoever handles implementation.
Trigger conditions, the moments when a human needs to review something, and the summaries shown at review time are each decided one by one through the conversation.
This article walks through the process of producing a specification, how to install the skill, example prompts, and things to check before you start.
When you invoke loop-me, the agent asks a series of questions and shapes your repetitive task into the form of a workflow specification document.
For example, you might ask:
/loop-me I manually compile last week's inquiries every Monday morning and share them with the team.
The first thing the agent checks is the NOTES.md file in your working folder.
If the file is empty or sparse, the agent starts with an intake conversation before moving on to the specification itself.
The questions cover which tools you use, how information reaches you, and what things are called internally.
After that, questions arrive one round at a time.
Each question comes with a recommended answer, so you can simply agree if it fits or describe your actual situation if it doesn't.
As each answer is settled, a specification file is created and updated inside a workflows/ folder.
What goes into the specification includes: what triggers the workflow, where a human needs to review, and what is shown at review time.
As long as even one question remains unanswered, the specification is not considered complete.
The specification is complete when an implementation agent could build it without asking a single additional question.
loop-me treats the repetitive elements of your work and daily life as "loops" and organizes them into a form that can be delegated.
It views your life as a whole through the lens of loops nested within loops.
The scope ranges from careers and weekly schedules down to morning habits and individual recurring tasks.
Seen this way, you can understand how predictable each activity is.
From there, you can identify the tasks that are worth handing off.
You don't need to have a target in mind before you start.
The skill can propose loops you may not have noticed yourself by looking at your recurring patterns.
The conversation draws on four key concepts.
These are not a checklist to fill in every time.
They are shared vocabulary that is only introduced when relevant to the workflow at hand.
| Term | Meaning |
|---|---|
| Trigger | What kicks off each run. There are "events," such as a new email or a new Issue, and "schedules," such as every morning. Event-based triggers are often more efficient. |
| Checkpoint | A point where the user reviews content or makes a judgment. Some workflows run fully automatically with no checkpoint. |
| Push right | The principle of completing all possible work before calling on a human, so that confirmation is requested only once, at the very end when everything is ready. |
| Brief | A short summary shown at review time. It covers what was produced, why those decisions were made, and links to deliverables — without exposing raw output. |
The reviewer reads the brief, not the draft.
Minimizing the time spent on review is a key priority.
loop-me does not require the use of AI, human checkpoints, or scheduled execution.
AI, checkpoints, and scheduled execution are only included in the specification when the conversation reveals they are actually needed.
Workflows that use no AI at all, and workflows that run fully automatically without any review step, can both be specified.
Two types of files are left in your working folder as outputs.
| File | Contents |
|---|---|
| workflows/*.md | One specification file per workflow. This is the authoritative document. |
| NOTES.md | Notes about the user's environment: which tools are in use, how information flows, and the agreed names for things. |
Ambiguous terms are standardized into agreed names as they come up and recorded in NOTES.md.
File creation, editing, and deletion are carried out in accordance with the finalized specification.
There are two ways to install: using a command, or asking an agent directly.
The command is shown below, and you can choose which agent to install it into — Claude Code, Codex, or others.
npx skills add https://github.com/mattpocock/skills --skill loop-me
If you want to review the contents before installing, send the following message to your agent:
次の Agent Skill をインストールしてください。
https://github.com/mattpocock/skills/tree/main/skills/in-progress/loop-me
まず SKILL.md と付属ファイルの中身を確認し、不審な処理（外部への送信、認証情報の読み取りなど）がないことを確かめてから、Claude Code なら ~/.claude/skills/loop-me/、Codex なら ~/.agents/skills/loop-me/ に配置してください。
Where you put it depends on whether you want to use it across all projects or only in a specific project.
| Agent | Use globally | Use in one project only |
|---|---|---|
| Claude Code | ~/.claude/skills/ | .claude/skills/ inside the project |
| Codex | ~/.agents/skills/ | .agents/skills/ inside the project |
The skill includes two files: SKILL.md and agents/openai.yaml.
loop-me does not start automatically — you invoke it explicitly.
In Claude Code, type /loop-me followed by a description of the workflow you want to design.
If you invoke it with nothing written, it starts by searching for candidates.
/loop-me
Please find repetitive tasks in my work that I might be able to delegate.
The agent starts by asking about the tools you use and how you receive information.
Based on that, it proposes loops from your daily and weekly routines that are candidates for delegation.
/loop-me Checking expense reimbursement requests.
I want to decide together when it kicks off and where I need to review.
The agent first asks a question about the trigger, with a recommended answer included.
The question is whether the workflow should start when a request arrives or at a fixed scheduled time.
The review step is refined with an eye toward placing it as late as possible, after as much preparation as can be done upfront.
The content of the brief shown at review time is also settled within this conversation.
/loop-me Please review the inquiry-summary specification in workflows/ and get it to a state where an implementation agent can build it as-is.
The questions raised cover the ambiguous points remaining in the existing specification.
The specification file is rewritten to match each answer, and any specifications that are no longer needed are deleted.
Once there are no more questions, the specification is ready to hand off for implementation.
loop-me only produces a specification.
The actual implementation of the automation — such as setting up email reading or scheduled execution — is a separate task.
This skill is designed to be used in conjunction with another skill, /grilling.
Before installing, confirm that /grilling is available in your environment.
The more specifically you answer each question based on your actual work, the more directly usable the specification will be.
When a recommended answer doesn't match your real situation, don't agree to it — describe your actual procedure or any exceptions instead.
Files are created inside your working folder, so starting in a dedicated folder will help keep things organized.
The source is mattpocock/skills, licensed under MIT.
This skill is located in the in-progress folder of the repository.
Before installing, read SKILL.md and the accompanying files and confirm there is nothing suspicious going on.
The same mattpocock/skills repository includes skills that are useful before and after loop-me.
ask-matt guides you on which skills to use and in what order, depending on your situation — whether you're organizing ideas, implementing something, or investigating a bug.
It's a good resource when you've created a specification with loop-me and aren't sure how to proceed next.
domain-modeling is a skill for clarifying the domain terms and conceptual boundaries used in a project by cross-referencing concrete examples with code.
Agreed definitions are recorded in a glossary.
It's well suited for when you want the names you've established with loop-me to remain consistent through the implementation phase.
Yes.
Each question comes with a recommended answer, so you can move through the process simply by saying whether it fits or explaining how your situation differs.
No programming is required.
That's fine.
If you invoke the skill without specifying a target, it will propose candidates based on the recurring patterns in your work and daily life.
At the start, you'll be asked about the tools you use and how information flows to you — just describe your usual work as-is.
No.
Whether to use AI, include human checkpoints, and use scheduled execution are each evaluated through the questions.
Procedures that use no AI, and procedures that run without any review step, can both be specified.
loop-me is a skill that drills into repetitive tasks through questions and compiles them into workflow specifications that an implementation agent can build without asking any additional questions.
Trigger conditions, the moments when a human reviews something, and the summaries shown at review time are all decided through questions with recommended answers.
What it produces is a specification; implementing the automation is done separately.
For anyone who wants to first nail down what to delegate and to what extent before moving toward automating their regular tasks, this skill is a solid first step.
See the cccskills page for more details.
This article is a reprint of the cccskills article "Turning Repetitive Task Automation into a Specification — the loop-me Skill."
Full details and installation instructions are available on the cccskills loop-me page.