A citizen developer is anyone who uses software, automation, or AI agents to improve their own work. You do not need a developer role or a coding background. A good first project starts with a task you repeat, a result you can recognize as correct, and a person who reviews what the system does.
Use this guide to try a task in TritonAI Harness, then decide how the work will change, who will rely on the result, and which supported TritonAI path fits.
How the work changesWill software repeat, improve, or replace the current steps?
02
Who depends on itWho will use it, and what happens if it fails?
03
Where to build itChoose the TritonAI path that fits the work.
Decide these three things before you build: how the work changes, who depends on the result, and which TritonAI path fits.
Hands-on with TritonAI Harness
Try a task with your connected tools
Start with one of the Harness examples and follow the exercise below. Replace the meeting, workshop, or department details with your own. Each exercise produces something you can inspect in your workspace.
Set up your workspace
Open the Harness and a project folder. If you are starting from scratch, follow access and setup to get a Gateway key and install the Harness. Use a folder where you can keep the draft and its supporting files.
Connect the sources for your task. In Harness Settings, enable Microsoft 365 or Google Workspace and sign in with your campus account. Outlook mail and calendar reads are default Microsoft 365 capabilities. Google Drive, Docs, and Sheets access is read-only. The available tools also depend on your account permissions and enabled plugin capabilities.
Name the result you want. Tell the Harness the topic, relevant dates, and output format. Ask it to identify the sources it used and any information it could not read. If a connection is unavailable, you can work from files you are approved to save in your project folder.
Plugins connect the Harness to your information. Skills guide how it carries out a task. For the onboarding exercise, an installed, enabled accessibility skill can guide the draft review. You can also practice tasks with skills below or explore the Skills Library.
Prepare a department meeting brief
Uses: Microsoft 365 mail and calendar reads. Have the meeting name and date ready.
Identify the meeting. Ask: “Find my department meeting tomorrow in Outlook. Show me the matching event and its time before preparing the brief.” If several events match, tell the Harness which one to use.
Gather the context. Continue: “Review related email conversations from the past two weeks. Draft an agenda with open questions and proposed follow-ups. Include source references and identify any gaps.”
Refine the draft. Try: “Keep this to one page. Separate decisions already recorded in the sources from items we still need to discuss. Mark proposed owners and dates as unconfirmed.”
Expected result: A meeting brief in your workspace with the event details, discussion topics, and references you can check.
Check before sharing: Confirm the meeting and time, open the cited messages, and verify that proposals are not presented as commitments. Reading the calendar and drafting the brief do not require sending invitations or email.
Summarize workshop feedback
Uses: Google Workspace spreadsheet reads and the Harness’s local file tools. Choose a workshop feedback sheet you are approved to analyze.
Find the input. Ask: “Find the feedback spreadsheet for our latest workshop in Google Drive. Show me the matching file, its tabs, and the columns you can read.” Confirm the correct workshop and tab.
Request the analysis. Continue: “Summarize the ratings and recurring themes, flag missing responses, and create a chart with a one-page summary. Suggest changes for the next session and show how you calculated the results.”
Check the calculation. Try: “State which rows you analyzed, the number of responses to each question, and how blanks were handled. Link the themes to supporting responses without including people’s names.”
Expected result: A chart and summary saved in your workspace. The connected Google Sheet stays unchanged.
Review the findings: Recalculate a sample of the totals, check the chart labels, and compare the themes with the responses. If only part of the sheet was read, narrow the task or provide an approved export before describing the results as a summary of the whole workshop.
Create an onboarding guide
Uses: Google Workspace document reads and the UCSD accessibility skill. Identify the department and the role the guide should support.
Locate the material. Ask: “Find our department’s onboarding documents in Google Drive. List the likely sources and their update dates, and identify any that you cannot read.” Select the documents that apply to the role.
Draft the guide. Continue: “Create a first-week checklist, explain how to request access to common tools, and link to the source documents. Flag conflicting instructions or missing information for me to resolve.”
Make it easier to use. Try: “Review the draft for accessibility, including headings, descriptive links, and reading order. Explain unfamiliar abbreviations and separate required steps from optional resources.”
Expected result: A draft guide in your workspace, with source links and a list of questions for the department owner.
Confirm the instructions: Ask the process owner to resolve conflicting versions, verify contacts and access-request links, and review the guide before sharing it with a new employee. Request an accessibility review of the final file format as needed.
Find information across connected sources
Uses: Microsoft 365 mail search and Google Workspace file search. Start with whatever you remember: a topic, person, phrase, or approximate date.
Describe the clues. Ask: “I’m looking for the latest instructions for requesting department software. I remember a discussion earlier this summer, but I’m not sure whether the instructions were emailed or saved in Google Drive. Search my Outlook email and Google Drive.”
Compare the matches. Continue: “Show the most relevant matches with their source, author or owner, and date when available. Explain which appears current, include source references, and flag conflicting versions.”
Narrow the search. If the matches miss the mark, add a colleague’s name, an alternate term, or a shorter date range. Ask: “Which sources did you search, what could you not access, and what should we try next?”
Expected result: A shortlist of messages and documents, with enough context to choose the source you need.
Verify the answer: Open the original material and check whether it is an approved instruction, a draft, or a discussion. A newer date alone does not establish authority. An empty result means the search found no match within the sources and limits it used.
Keep a prompt that works
Once you have checked a result, ask the Harness to save the instructions in your project folder with the meeting name, dates, or source files marked as details to change next time. Include the output format and review steps. You can rerun the task yourself, or use the project-planning questions below when you want to turn it into a recurring team workflow.
First decide how software should change your work. Any of the three patterns below can use regular code, an AI model, an AI agent, or a mix of these.
Automate
Software repeats the steps you already do, and the finished result stays the same. Start here if you know exactly what a correct result looks like and can go back to doing the task by hand if something goes wrong.
Example: Turn a monthly spreadsheet cleanup into a pipeline that shows you every change before you accept it.
Upgrade
The task keeps its purpose, but people get a better result than before. The result might be interactive, always current, searchable, or available without waiting for the next meeting.
Example: Replace a status deck you rebuild every week with a web page that always shows the latest numbers and when they were last refreshed.
Invent
Build something new that was never possible by hand. Because there is no existing process to copy, treat usefulness as a question to test before making the build permanent.
Example: Watch approved public websites for changes that matter to your unit and send the relevant ones to a person for review.
Who will use it
Add controls as more people rely on the result
The more people depend on what you build, the more care it needs: testing, access management, support, and monitoring. Even a small or temporary build must follow UC San Diego requirements for data, security, privacy, and appropriate use. The Build overview shows the hosting and accountability that come with each rung.
01
Prototype
Tests an idea with sample data you are approved to use. Keep it to yourself, write down what it showed, and shut it down or revise it when the test ends.
02
Team workflow
Does a recurring job for you or a small, known group. It needs consistent results, an approved place to run, a person who checks the output, and a way back to the manual process if it breaks.
03
Campus service
Serves people outside your own team. Before launch it needs all of the following:
a named owner
accessibility, privacy, and security review
testing under realistic load
a support path and monitoring
a process for making changes
04
Enterprise service
Serves the whole campus or the public, or hurts badly when it breaks. This requires the formal campus product and service process with ongoing ownership, and it is rarely a good first project.
Opportunity scan
Where to look in your day-to-day work
Look for tasks you repeat, explain, check, or review. These examples are starting points for a conversation with your service owner. Each project still needs the appropriate review before work begins.
Presentation work
Turn an explanation you give over and over into a walkthrough, comparison, calculator, status page, or interactive explainer.
Content work
Turn approved source material into a draft summary, update, or brief that a person reviews before it is used.
Data work
Reconcile exports, repeat an analysis you run often, flag exceptions, or replace a recurring snapshot with a dashboard that cites its sources.
Document work
Pull approved fields out of documents, fill in a template, convert files between formats, or link each finding to the exact passage a reviewer needs to check.
Inbox and intake work
Organize submissions, spot missing information, route a request to the right team, or prepare a draft response for a person to check and send.
Administrative work
Assemble a briefing, track a deadline, update a checklist, or coordinate repeatable steps across approved systems.
TritonAI ecosystem
Choose a path that fits the work
Start with the simplest supported option that meets the need. You can move a workflow to a different path later as its audience, data, or support needs change.
Work on a small task that can span files, browser work, and approved connections. The Harness asks for your approval before significant actions, so you stay in control while you learn which steps are worth repeating.
Repeatable workflow
n8n
Run a process you already understand. n8n starts the steps on a schedule, when a message or file arrives, or when another system sends a request. It fits best once you know the process and how it should handle exceptions.
Application or shared capability
APIs and skills
Add approved AI models to software you already run, or package instructions and reference material into a skill that others can review and reuse.
A good first project has a result you can verify, limited access to people and data, and a clear way back to the current process. These are the same questions the use-case conversation asks, so write down the answers before you open a builder or ask an agent to act.
Need
Who is doing the work, what keeps slowing them down, and how often does it come up?
Owner
Who decides what good looks like, and who decides whether to keep, change, support, or retire the build?
Data
What information may the workflow read, create, or change, and what must it leave alone?
Review
Where must a person check the result before it changes a record, reaches another person, or informs a decision?
Outcome
What does a correct result look like, and what evidence would show that the build helped?
Fallback
Can you stop the workflow and return to the current process without losing access or data?
Practice with skills
Try a skill on a campus task
Choose an exercise below and work through it in your Harness project folder. Each one connects a skill to a result you can inspect. Use the sample details provided or substitute your own approved material.
Before you begin: Check that the relevant skills are installed and enabled in your workspace. The Harness can select them from your request and conversation context; you usually do not need to name them. If it misses the intended skill, ask which skill it used or name the one linked in the exercise.
Ask for a review. “Review this workshop handout for accessibility. Check the headings, links, and image descriptions. Point to each issue and explain how it affects someone reading the handout.”
Improve the draft. “Create a revised copy with clearer headings and descriptive link text. Suggest image descriptions for me to verify. Keep the workshop instructions and facts intact.”
Check the result. “Summarize what changed and which checks still need a person or a different tool.” Compare the files and check the final reading order.
What you’ll get: A revised handout and a list of remaining checks. Review any image descriptions for accuracy; the exercise does not certify accessibility compliance.
Describe the plan. “We are planning a workshop signup form with these fields: name, campus email, department, and accommodation request. Explain the likely UC Protection Levels and what additional context you need.”
Refine the questions. “Which fields need closer review? Suggest questions for the responsible campus owner about what we collect, who can see it, and how long we keep it.”
Prepare for review. “Turn this into a short checklist with the proposed fields, classification reasoning, and unresolved decisions. Include the guidance you used.”
What you’ll get: An initial data-handling review to discuss with the campus owner. Confirm the classification and the approved collection and storage approach before using the form.
Prepare the folder. “Check whether this project already has memory. If it does not, set up a local memory folder. Show me where the notes will live and ask before enabling background updates or importing connected sources.”
Save a decision. “Save these sample decisions for this exercise: registration closes one week before the workshop, and the department coordinator reviews the handout. Label them as practice notes.”
Retrieve it later. Start a new conversation in the same project: “Find our saved workshop-planning decisions. When does registration close, and who reviews the handout? Cite the notes and flag anything that is missing.”
What you’ll get: Saved notes and a source-linked answer you can compare with them. Confirm that the new conversation can access the same memory folder.
Ask what is active. “Check which model this Harness session is using and which skills are available. Explain what you can confirm and what you cannot inspect.”
Clarify the evidence. “Which findings describe this running session, and which come from saved settings? Explain any difference in plain language.”
Choose a next step. “If something needed for the accessibility exercise is missing, explain how I can check or enable it. Leave the settings unchanged.”
What you’ll get: An explanation of the setup the agent can inspect, with missing information identified. Compare it with Harness Settings; a project folder alone may not reveal the running app’s configuration.
Explain the problem. “Help me draft feedback for the TritonAI team. This is a practice scenario: I asked for a meeting brief, but the task stopped before saving the file. Ask for the details needed to explain the problem. Do not send anything.”
Make it useful. “Include what I asked for, what I expected, what happened, and steps to reproduce it. Mark information I have not provided as unknown.”
Review the draft. “Show the subject and message. Keep it concise and identify any details I should check before sharing.”
What you’ll get: An email draft you can review. For a real report, confirm the facts and remove sensitive details before choosing whether to send it.
Check a spreadsheet-cleanup script
Skill:Review code changes. This exercise is for someone who already has a script or software project. Use a copy of the project with sample spreadsheet data.
Request a review. “Review the spreadsheet-cleanup script in this project. Check how it handles missing columns, blank cells, and duplicate rows. Run relevant tests with sample data and explain any problems.”
Inspect a finding. “Show a small input that demonstrates each confirmed problem, along with the expected and actual result. Separate confirmed failures from questions that need more information.”
Try a correction. “Suggest the smallest fix for the confirmed problem. After I choose a fix, apply it and repeat the relevant checks.”
What you’ll get: A review tied to examples and test results, with fixes or remaining issues identified. Compare the output with a manually checked sample before using the script on real work.
Passport Services needed a waiting-room tool that matched its check-in process at two locations. The department's technology team built a public check-in flow, a staff dashboard for each location, and operational reporting, and published the source code. Staff still manage the queue and make every service decision.
It follows the pattern on this page: a recurring need, a named owner, a bounded set of data, a person reviewing each step, and a result the team can measure.
Bring a short problem statement to a citizen-developer session or the AI Use-Case Meeting. Cover the need, the owner, the data, the review point, and the outcome you expect. That is enough to choose a pattern and test a first version.
Until you build something, it is hard to see which parts of your work have software-shaped solutions. Take the idea that seemed most interesting on this page and find out what it would take to build. Maybe it amounts to nothing and you abandon it. Either way you will know your work better than you did before.
Start with one small task and data you are approved to use, and keep the current process running while you test.