Projects: Build Something With an Agent

info

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.

The track so far has taught habits: decide, brief, supply context, check, and protect. This page asks you to package those habits into one thing that keeps working after you close the tab. You will build a project with standing context, a skill for a recurring task, or a scheduled agent, use it on real work, and grade it against a rubric. The goal is not a portfolio piece for its own sake. It is a tool you will still be using next month.

Learning focus

• Choose a project track that fits a real, recurring task in your work or life
• Build a reusable project, skill, or scheduled agent using the six-field brief and the right context
• Test it on real requests, verify its output, and scope its permissions
• Grade it with a rubric and write a one-page case study of what you built and learned

Before you start

You should have worked through the six lessons from Decide to Protect, and ideally Exercises 5 to 8 on the Hands-On AI Exercises page, which practice the brief, verification, mode choice, and connectors you will use here.

Pick a task that is recurring (weekly or more often), checkable (you can tell a good result from a bad one), and yours (you have the files and the access it needs). A weekly status report, a lesson plan, client follow-ups, an expense summary, or a reading digest all qualify. "Help me with my career" does not.

Choose a track

Track	You build	Good fit when	Typical effort	Tier
1. A project	A project with standing instructions, source files, and saved briefs	You keep re-explaining the same context	2-4 hours	Any, including free (Claude's free tier allows up to five projects)
2. A skill	A packaged how-to for one recurring task, with template and checks	The task has a method you want done the same way every time	3-5 hours	Any app that supports Agent Skills
3. A scheduled agent	A task that runs on a schedule through your connectors and reports back	The task repeats on a timetable and needs current data	4-8 hours over a week	Usually a paid tier

If you are unsure, start with Track 1. Each track builds on the one before it: a good skill usually starts as a well-tuned project, and a reliable scheduled agent usually runs a skill.

Track 1: A project with standing context

The problem it solves: every new conversation starts with five minutes of explaining who you are, who the audience is, and which files matter.

Examples: a meal-planning assistant that knows your household's diets and budget; a study guide generator loaded with your course materials; a client-communications project with your style guide and past proposals; a home-renovation planner holding quotes, measurements, and constraints.

Build it:

1. Gather sources. Collect the files that genuinely matter, usually 5 to 20, and name which one is authoritative when they disagree.
2. Write standing instructions. Ten clear sentences beat two pages. Cover who you are, the audience, house style, the sources that count, what is out of scope, and the checks you always want. Cut any line whose removal would not cause a mistake.
3. Save three briefs. Write six-field briefs for the three requests you make most often, and keep them in the project.
4. Test with ten real requests. Use requests from the last month. Score each answer as usable, usable with edits, or wrong, and note why.
5. Tune and retest. Fix the instructions or sources behind each failure, then rerun the failures.

Deliverable: the working project, a one-paragraph note on what it should not be used for, and your ten-request test log.

Track 2: A skill for a recurring task

The problem it solves: a task you do often has a method, a template, and a set of checks, and you want the assistant (and your colleagues) to follow them every time without pasting instructions.

Examples: turning raw meeting notes into your team's decision log; producing a monthly customer-feedback summary from a survey export; drafting product descriptions in your shop's format; preparing a lesson plan in your school's template; analyzing a spreadsheet of expenses into a budget memo.

Build it:

1. Write down the method. Do the task once by hand and record the steps, the template, and the checks you apply at the end.
2. Create the skill. A skill is a folder with a SKILL.md file describing when to use it and how, plus any templates it needs. Agent Skills is an open standard adopted by many agent products, and the major apps include a skill-creator that interviews you and writes the file, so no code is required. The Skills, AGENTS.md, and CLAUDE.md page explains the format.
3. Write a sharp description. The assistant loads a skill when a task matches its description, so say exactly when it applies and when it does not.
4. Test triggering and output. Try five requests that should use the skill and three that should not. Then check the output of each correct trigger against your end-of-task checks.
5. Hand it to someone else. Ask a colleague or friend to run it on their own example without help from you. Every question they ask is a gap in the skill.

Deliverable: the skill folder, a before-and-after comparison on one real task, and a note of what you changed after the hand-off test.

Track 3: A scheduled agent

The problem it solves: a task that repeats on a timetable and needs current information, such as a Monday digest of your calendar and open threads, a weekly report from a shared spreadsheet, or a daily check of a few web pages for changes.

Build it:

1. Decide whether it should be an agent. If every run follows the same fixed steps and needs an audit trail, a plain workflow may be better; Workflows vs Agents has the decision rule.
2. Write the brief and run it by hand first. Get three good manual runs before you schedule anything.
3. Scope the connectors. Include only the connectors this task needs, read-only where possible. Scheduled tasks run without you watching, and an included connector's write tools can be used without asking.
4. Require approval for anything outward-facing. Have the agent save drafts rather than send, post, or delete. Ask for an action log at the end of every run.
5. Schedule it and watch three runs. For each run, check the real output (the document, the draft, the spreadsheet row), not the status message; a scheduled task can report success while the work inside it failed.

Deliverable: the scheduled task, its brief, a permissions note listing every connector and what it can do, and a log of three checked runs.

For longer-running and multi-step versions, the Build track's Loops and Harnesses section and the playbook Automate a Weekly Report with AI go further.

A four-session plan

You do not need weeks. Most people finish in four focused sessions spread over a week, which leaves time for real use between them.

1. Plan (45 minutes). Choose the task and track. Write the problem in two sentences and define what "working" means in a way you can check.
2. Build (1-3 hours). Gather sources, write the instructions, brief, or skill, and get one good result.
3. Test (1-2 hours, spread over several days). Run it on real work, log every result, and fix the causes of failures rather than the symptoms.
4. Document (45 minutes). Grade it against the rubric and write the case study.

Grade it: the rubric

Score each criterion from 1 to 3. A total of 13 or more out of 18 means you have built something worth keeping; anything scored 1 tells you what to fix first.

Criterion	1: Needs work	2: Solid	3: Excellent
Problem and value	Vague task, unclear benefit	Real recurring task, benefit described	Real task with measured time saved or quality gained
Brief and context	Instructions are long, contradictory, or missing sources	Clear instructions and the right files	Short, consistent instructions; authoritative source named; examples where format matters
Verification	No testing beyond one try	Tested on several real requests	Test log with failures fixed at the cause; output shows its evidence
Safety and permissions	Red-light data or broad write access present	Data and permissions appropriate	Least privilege, approvals on anything outward-facing, limits documented
Reusability	Only works for you, today	Works on new requests next week	Someone else can use it without your help
Reflection	No notes	Notes on what worked	Clear account of what failed, why, and what you would change

Write the case study

A one-page case study makes the project useful to someone else and to future you. Keep it to five short sections:

Problem: [the recurring task and what it cost before, in time or quality]
What I built: [track, tools, sources, and the key instructions or brief]
How I checked it: [test set, failure rate before and after, what I verify every time]
Safety: [data it touches, connectors and permissions, approvals]
What I learned: [one thing that worked, one that failed, what I would do next]

Add a screenshot of the working result and one before-and-after example. If you share it on LinkedIn, a blog, or your team's wiki, leave out any client names or data that the Protect traffic-light rule would keep yellow or red. In an interview, a case study like this demonstrates exactly what employers mean by AI fluency: judgment about what to delegate, a clear brief, and the habit of checking.

Key takeaways

• Build one reusable thing for a recurring, checkable task: a project, a skill, or a scheduled agent
• Start with Track 1 if unsure; good skills grow from good projects, and good agents run good skills
• Test on real requests, fix failures at their cause, and check real output rather than status messages
• Scope permissions narrowly and require approval for anything sent, posted, or deleted
• Grade with the rubric and write a one-page case study so the work outlasts the week

Next steps

Congratulations on finishing the AI Fluency track. Choose where to go next:

• Continue to: AI Agents Overview to build agents that do more on their own
• Set up a personal agent with standing context and a routine in the playbook Set Up Your Personal Agent
• Bring these habits to your team with AI for Teams and Leaders