Custom Skills: Build Once, Run Forever
The full series
- The Prompt Is the Least Important Part
- Three Files That Replace a Chief of Staff
- You Don't Have a Productivity Problem. You Have a Setup Problem.
- Stop Telling AI What to Do. Tell It What Done Looks Like.
- The _MANIFEST.md: Five Minutes That Save Hours
- Plugins Are Composable. Stack Them.
- The Monday Briefing That Runs Without You
- Custom Skills: Build Once, Run Forever (you are here)
- AI as a Team, Not a Chatbot
- Cowork Is Not Chat. Stop Using It Like Chat.
You told Claude how to write your weekly report. Explained the format, the tone, the sections, and the quality bar. Next week, you explained it again. The week after that, again. A skill file permanently ends the repetition.
Context files tell the AI who you are. They don't tell it what to do.
Context files hold professional identity, voice, and behavioral rules that answer "who am I working with?" at the start of every session. But they can't answer the other question that matters just as much: "How do we handle this specific piece of work?" That knowledge usually lives in someone's head, repeated from memory each time the same task shows up in a new session.
A skill is a markdown file that encodes everything you'd spend twenty minutes explaining to a new hire, a freelancer, or an AI agent. Purpose, inputs, process, output, constraints. Five sections in one plain text file that never forgets and never needs to hear the explanation twice.
Build once. Run forever.
Five sections, one file
Every skill file follows the same five-section structure, and you can write one in about five minutes flat.
Name the purpose. One sentence describing what this skill produces and why it exists, scoped tightly enough that the deliverable is clear before work begins.
List the inputs. What information does the AI need before it can start? Topic, data file, target audience, deadline. Every input left unspecified is one that the AI will guess at, and guessing always degrades the output.
Map the process. Step-by-step instructions for how the work gets approached, covering whether to research before outlining, which context files to reference, and what sequence to build the sections in.
Specify the output. What the finished deliverable looks like when the work is done: file format, word count, section structure, and the folder where it gets saved.
Set the constraints. What the AI must avoid. Banned phrases, tone violations, formatting rules, and quality standards the deliverable must meet before it counts as finished. Constraints do more work than any other section because they eliminate the categories of weak output you'd otherwise correct by hand.
Never explain it again.
A complete example
This skill file for weekly article drafting took five minutes to create.
# Weekly Article Drafting
## Purpose
Draft a 2,000-word article from a topic and outline.
## Inputs
Topic, outline, target audience, key evidence.
## Process
1. Research the topic using web search.
2. Draft sections following the outline order.
3. Match tone and voice from brand-voice.md.
4. Generate visual suggestions for each section.
5. Include quotable lines for social sharing.
## Output
.docx file saved to /articles/drafts.
## Constraints
- No AI cliche language ("In today's fast-paced world...").
- No filler phrases or padding.
- Minimum 8 evidence points with sources.
- Must match brand voice exactly.
- No hedge words (might, perhaps, arguably).Now the prompt for every future article is eight words: "Run the article drafting skill on [topic]." That single file encodes the quality bar, the format requirements, and every constraint that would otherwise need re-explanation each session.
We've already seen practitioners build skill files for client reports, competitive analyses, investor updates, and content calendars across industries.
The pattern holds everywhere.
When you find yourself explaining the same workflow more than twice a month, it belongs in a file rather than a conversation.
Build skills without writing them
You don't have to create these from scratch.
The Plugin Management plugin turns skill creation into a guided conversation. Install it and say, "Help me create a plugin for [workflow]." Claude asks what the task produces, what quality means for this particular deliverable, and where the guardrails belong. Then it assembles the skill file and displays the result before saving, so you can refine through dialogue rather than through editing unfamiliar markdown.
No code required and no formatting to learn.
For anyone who built context files during the earlier interview process in this series, this will feel familiar. Describe the need, let the AI structure it, review, and save.
The conversation is scaffolding. The skill file is what remains.
Skills are process, context files are identity
These two file types serve fundamentally different purposes, and confusing them weakens both.
Context files are identity. They capture who someone is, how they communicate, and how they want the AI to behave. They load automatically at the start of every session, regardless of the task.
Skill files are process. They capture how to execute one specific type of work to one specific standard. They load only when invoked by name for a particular task.
Your brand-voice file ensures every piece of output sounds like you, regardless of the deliverable type. Your article drafting skill ensures every article follows the same structure, hits the same evidence threshold, and lands in the correct folder. Voice shapes tone across all work.
Skills shape one type of work with precision.
Both compounds over time, but in different directions. Context gets sharper about identity with each correction. Skills are sharpened in execution with each refinement.
Skills as institutional knowledge
When a process lives only in someone's head, it leaves the organization when that person does.
New hires spend weeks reconstructing workflows that nobody documented. Freelancers guess at quality standards that weren't written down. Teams lose months of accumulated knowledge every time someone changes roles.
A skill file survives all of that. Personnel changes, reorganizations, and the slow erosion of memory that occurs when a report format evolves over 18 months, with no one tracking the changes.
We've seen teams build shared skill libraries where every repeatable workflow gets its own file with constraints and quality standards baked in. We've watched onboarding drop from weeks to days because new people run the same skills with the same guardrails from day one.
The quality bar doesn't reset when someone walks out the door.
The knowledge stays in the folder.
When to build one
Not every task justifies a skill file.
One-off requests don't warrant five minutes, and simple tasks that can be covered with a short prompt don't need permanent documentation either.
The threshold is repetition.
Any workflow you repeat more than twice a month deserves a skill file. Any task where you re-explain format, structure, or constraints session after session belongs in a permanent file rather than in conversation memory that vanishes when the window closes.
Weekly reports, monthly analyses, client deliverables with fixed templates, and content drafts that follow a consistent structure.
One signal. When your prompt includes "like last time" or "same format as before," that format should already be in a file the AI can find without being told.
Build once. Run forever.
Do This Today
- Identify one task you explain to Claude more than twice a month. The weekly report, the client summary, and the content draft with specific formatting rules.
- Create a markdown file with five sections: Purpose, Inputs, Process, Output, Constraints. Keep each section to two or three concrete lines.
- Be specific in Constraints. Not "write well" but "no hedge words, no sentences over 25 words, no bullet points in the executive summary." Specific bans outperform general aspirations.
- Save the file in your workspace folder where Cowork can read it.
- Run the task by saying "Follow the instructions in [skill-file].md and apply them to [topic]." Compare the output with your last manually prompted version.
---
Next in the series: the next version of AI productivity isn't one agent doing one task. Multiple agents with defined roles run in parallel, reviewing each other's work before a human ever sees the output. We cover subagents, parallel processing, and why the human's job becomes review.