Build Your Brief and Skill Library
This content was developed with AI assistance and is regularly reviewed for accuracy.
Snapshot: October 2026. Model names, prices, and product features change monthly; What's New in AI is the maintained chronology.
A prompt library used to mean a document of templates you copied into a chat box. In 2026 your assistant can hold much of that for you: projects keep instructions and reference files, and skills package a procedure the assistant loads whenever it is relevant. This playbook helps you decide which of your repeated tasks belongs in which layer, then build one of each so recurring work comes out your way without re-explaining.
Last tested with Claude (projects and skills on the Pro plan) and ChatGPT (projects and custom instructions on Plus) in October 2026.
What you'll accomplish
- Audit your repeated tasks and sort each into the right layer
- Write a brief template for work that changes every time
- Set up a project for an ongoing body of work
- Write your first skill as a SKILL.md file
- Test, version, and share what works
Prerequisites
- The habits from the beginner playbooks, especially the six-field brief in Your First Delegated Task
- Claude or ChatGPT. Projects are available on Claude's free plan (up to five) and in ChatGPT. Skills are listed on Claude's paid plans and follow an open standard other tools also read; see Step 4.
- Two weeks of memory about what you do repeatedly
- 45-60 minutes
The three layers
Before the steps, it helps to know what each layer is for. They stack: a skill can be used inside a project, and a brief can call on both.
| Layer | What it is | Best for | Example |
|---|---|---|---|
| Brief template | A saved six-field brief with blanks | Tasks that recur but change in substance each time | A client proposal for a new prospect |
| Project | A workspace with standing instructions and reference files | An ongoing body of work with shared context | Everything for one client or one course |
| Skill | A folder with a SKILL.md file of instructions the assistant loads when relevant | A procedure you want done the same way everywhere | "How I format meeting notes" |
Memory is a fourth layer, but you do not curate it the same way; it collects what the assistant learns from your conversations. The library is the part you control on purpose.
The playbook
Step 1: Audit and sort your repeated tasks
Goal: Find the five tasks where a saved setup would save the most time.
Action: List what you did with AI, or could have, over the past two weeks. Then ask the assistant to help sort it:
Here are tasks I repeat. For each, tell me whether it fits best as a brief template (changes a lot each time), a project (ongoing work with shared files and context), or a skill (a fixed procedure I want followed every time). Explain each choice in one line, then rank the five that would save me the most time.
- Weekly status update to my manager (Fridays)
- Reviewing 2-3 resumes a week against a job description
- Monthly product announcement posts
- Agendas for weekly one-on-ones
- Summaries of long reports for my team
- Formatting meeting notes into decisions and action items
Checkpoint: Five tasks, each assigned to a layer, with a reason you agree with.
Step 2: Write a brief template
Goal: Make your most variable recurring task a two-minute fill-in.
Action: Write the six fields once, with brackets for what changes:
Objective: Draft my weekly status update to [manager].
Context: [Two lines on what changed this week.] My manager reads this on her phone between meetings.
Resources: My running notes in the Google Doc "[doc name]" and my sent email from [dates].
Constraints: Under 200 words. Lead with status (on track / at risk / blocked). No jargon. Do not invent progress that is not in the notes.
Output: Four short sections: Status, Done, Next week, Needs your help.
Verification: After the draft, list which note or email each "Done" item came from.
Store templates where you can search them: a notes app, a shared doc, or a project's files. Keep each to three to seven blanks; a template with fifteen is slower than writing from scratch.
Checkpoint: You have filled in the template twice with different weeks and both outputs needed only light edits.
Step 3: Set up a project for ongoing work
Goal: Stop re-supplying the same background for one body of work.
Action: Create a project named for the work ("Client: Northwind" or "Q3 hiring"). Add standing instructions and the reference files the assistant should always have:
This project is for hiring a senior data analyst. Always:
- Evaluate candidates against the attached job description and scoring rubric.
- Weigh relevant experience and evidence of impact above resume formatting.
- Flag anything that would need a human judgment call (career gaps, visa questions) without scoring it.
- Never infer age, gender, ethnicity, or other protected characteristics.
Attach the job description, the rubric, and one example of a review you were happy with. Examples steer format more reliably than adjectives. Every conversation inside the project now starts with this context.
If you used Custom GPTs for this kind of setup: individuals can no longer create new ones as of August 2026. On a personal ChatGPT plan, a project with instructions and files is the replacement; businesses use Workspace Agents.
Checkpoint: A new conversation in the project produces a review in your format without you restating the rubric.
Step 4: Write your first skill
Goal: Package a procedure so the assistant applies it whenever it fits, in any conversation.
Action: A skill is a folder containing a file called SKILL.md. The file starts with a short header naming the skill and describing when to use it, followed by plain instructions. Agent Skills became an open standard in December 2025, and by March 2026 OpenAI, Microsoft, Google's Gemini CLI, Cursor, and at least 25 other products had shipped support, so a skill you write once is portable.
---
name: meeting-notes
description: Use when the user shares meeting notes or a transcript and wants them cleaned up, summarized, or turned into action items.
---
# Meeting notes
1. Start with a three-bullet summary of what was decided.
2. Then an action items table: What, Owner, Due date.
3. Only list an owner or date that appears in the notes. If one is missing, write "unassigned" or "no date" instead of guessing.
4. End with open questions that were raised but not resolved.
5. Keep the whole thing under 250 words.
The description is the most important line. The assistant reads it to decide when to load the skill, so name the situations plainly. Add the folder through your assistant's skills settings; in Claude, skills added on claude.ai also sync to Claude Code.
Checkpoint: Paste some rough meeting notes into a fresh conversation without mentioning the skill. The output follows your format.
Step 5: Test, refine, and version
Goal: Make each piece reliable before you depend on it.
Action: Use each template, project, and skill on three real tasks. After each, note what you had to fix by hand. Then fix the source, not the output:
- A skill that did not trigger usually has a vague description. Add the words people actually use ("notes", "transcript", "recap").
- A project that drifts off-format needs an example file, not more adjectives.
- A brief that keeps producing invented detail needs a sharper verification field.
Put a date and a one-line change note at the bottom of each file. When a new model ships, rerun one task per item; habits that compensated for an older model sometimes become unnecessary.
Checkpoint: Each item has been used three times and updated at least once.
Step 6: Share with your team
Goal: Multiply the value without multiplying the mistakes.
Action: Share brief templates in a team doc with one line each on when to use it and what to watch for. Share skills as folders; your colleagues add them to their own assistants. If your team is on a business plan, check whether projects can be shared there too.
Treat skills from other people the way you treat software. A skill is instructions your assistant will follow, and some include scripts. Read every skill before you install it, and prefer ones from people or publishers you trust.
Checkpoint: A teammate has used one of your items and got a usable result on the first try.
Common pitfalls
- Putting everything in one giant instruction: Separate projects and skills are easier to fix than one mega-prompt.
- Writing skills for one-off tasks: If the procedure changes every time, it is a brief, not a skill.
- Vague skill descriptions: "Helps with writing" never triggers reliably. Name the situations.
- Storing the library in chat history: Chats are hard to search and easy to lose. Keep the source in files.
- Installing skills you have not read: Skills can carry instructions you would not approve. Read first.
Key takeaways
- Brief templates handle variable work, projects hold ongoing context, and skills package fixed procedures
- Projects with instructions and example files replace re-explaining (and replace personal Custom GPTs)
- A skill is a SKILL.md file; its description decides when it is used
- Fix the source after every manual correction, and date your changes
- Share what works, and read any skill before you install it
Next steps
Continue to: The AI-Powered Meeting Workflow to put your meeting-notes skill to work.
Continue to: Brief and Prompt Library for ready-made briefs to add to your collection.
Continue to: Skills, AGENTS.md, and CLAUDE.md for the full skills standard, including scripts and reference files.