Building a personal AI assistant with Claude takes no coding. You create one Claude Project, write a system prompt that fixes the assistant's role, personality, context and rules, then upload a context file the project re-reads every session. AI Central's guide walks ten actionable steps from zero to a working assistant, gives a copy-and-paste block for each layer, and adds five tests that prove the setup holds before you start relying on it.
Reviewed August 7, 2026. The source is the AI Central guide Build a Personal AI Assistant with Claude, a step-by-step build with a copy-and-paste prompt block at every stage.
The build is four layers, not one clever prompt
Most people treat a personal assistant as a prompting problem. Open a chat, describe what you want, hope the answer lands, repeat it all tomorrow. The guide treats it as a configuration problem instead, and that is the entire difference.
Everything lives inside a single Claude Project, and inside that project one field does the heavy lifting.
The system prompt runs silently in every chat, it defines who your assistant is
That is AI Central's framing, and it explains the shape of everything that follows. If the system prompt defines the assistant, everything you want it to be has to be written into that one field, in layers.
- Role, the single job you are hiring the assistant to do.
- Personality, the tone, length and format it defaults to.
- Context, who you are, what you are working on, what to avoid.
- Rules, the behaviours it must always or never perform, plus trigger words.
The field is called Project Instructions, and the guide notes that it stays across all chats. That is the point. You configure once and stop re-explaining yourself every morning.
Pick one job before you write a single line
The first step is the one most people skip. The guide asks you to build around one main job rather than a general helper, and offers six shapes to pick from: writing coach, thinking partner, day planner, work assistant, learning tutor, accountability coach.
If you cannot name yours, the guide hands the question back to Claude with a discovery prompt asking for five targeted questions to surface your number one use case. That instinct is right. A vague brief produces a vague assistant, and the vagueness compounds through every layer you build on top of it.
Personality is four dials, not a vibe
The guide reduces personality to four dimensions you tune explicitly, rather than hoping the model guesses.
- Tone. Direct, warm, casual, formal or Socratic, pick one main style.
- Response length. Short by default or detailed, set the baseline.
- Pushback style. Should it challenge you or always agree, and define when to push back.
- Format. Bullets or paragraphs, headers, markdown, choose your default.
One instruction in the supplied block earns its place: never be sycophantic, skip Great question, always get to the point. That single line removes the filler that makes most assistant output feel like customer service. Pair it with the length rule, short unless I ask for more, and two of the most common complaints about chat assistants are gone before you ask a question.
Rules are what stop the drift
Rules stop Claude from drifting. They make your assistant predictable and reliable every session
That is AI Central again. The rules layer is the least glamorous part of the build and the part that decides whether the thing survives a month of daily use.
Each supplied rule does a specific job. Every response ends with one clear next action. A vague request triggers one clarifying question before an answer. Responses never open with Certainly or Of course. Unsolicited advice outside the stated task is off limits.
Then come the triggers, which are the sharpest idea in the document. A trigger is a word you type that switches the assistant into a mode. Say quick and the reply comes back in three sentences or less. Say deep dive and you get a full structured breakdown. Say brainstorm and you get ten ideas minimum. You are building yourself a small command language, and it costs a couple of lines in the system prompt.
Memory is a file you maintain, not a feature you switch on
Claude doesn't keep memory between chats by default, but Projects plus uploaded files act like persistent memory
This is the most useful correction in the guide, stated plainly rather than pretended away. Continuity is something you build, not something you are handed.
The workaround is housekeeping. Create a file called My Context.md, update it after every meaningful session, and upload it to the project so it gets re-read every conversation. Add a second file, Decisions Log.md, tracking decisions made, lessons learned and updated goals.
So the files do not become a chore, the guide supplies a session summary prompt for the end of each session. It asks for key decisions, action items, context to remember and any updates to goals, capped at 150 words. The cap is doing real work. An uncapped summary bloats the context file until Claude is re-reading noise.
Five tests that prove the setup holds
Almost nobody tests a system prompt. The guide argues you should, and gives five checks that each expose a different failure.
- Vague request test. Say help me with my project, and see whether it asks a clarifying question.
- Tone test. Check whether it matched the personality you defined.
- Context test. See whether it referenced your goals or role without being asked.
- Constraint test. Fire the quick trigger and check whether it stayed under three sentences.
- Edge case test. Ask something off topic and see whether it redirects appropriately.
There is also an audit prompt asking Claude to find the three weakest parts of your system prompt and suggest a specific rewrite for each. Running your configuration back through the model that has to obey it beats guessing at why the assistant keeps ignoring you.
The routine is where it stops being a chatbot
AI Central's position is that the best personal assistants are not used randomly, they are part of a structured daily workflow. The shape is three touchpoints. A morning kickoff where you paste your task list and get priorities ranked plus a plan. The working hours, where it acts as a thinking partner for decisions, drafts and research. An evening review that summarises the day, updates the memory file and sets tomorrow's focus.
The morning prompt asks for three things at once: rank these by impact, identify what I should not do today, and suggest one thing I might be missing. That middle instruction is the interesting one. Most planning tools only add. Asking what to drop is what makes the ranking mean anything.
What to do with this
Starting from nothing, build the minimum version today. One project, one use case, a system prompt carrying role, tone and the personality block, nothing more. Use it for a week before adding a single advanced feature.
Already running a project? Your upgrade is the rules layer and the memory files. Add the triggers, create My Context.md, run the session summary at the end of every working day. That is the difference between an assistant that is good in one chat and one that is still good in month three.
For tools, the guide lists web search, document analysis, code execution, data and charts, Google Calendar and Gmail integration, all under Settings then Features in Claude.ai. Switching them on is not enough. The guide pairs them with a tool instruction telling Claude to search the web first when you say research, and to analyse a shared file fully before responding. Capability without instruction goes unused.
Where exactly does the system prompt go in Claude?
Into Project Instructions. Go to claude.ai, open Projects, create a new project, then paste your prompt into the Project Instructions field and save. It stays across all chats in that project, which is why you write it once instead of at the top of every conversation.
Does Claude remember me between conversations?
Not by default. The guide is explicit that Claude does not keep memory between chats, and that Projects plus uploaded files act like persistent memory instead. You get continuity by maintaining a context file and a decisions log inside the project, not by expecting the model to recall the last session on its own.
How do I know whether my system prompt is any good?
Run the five tests. Give it a deliberately vague request, check the tone, check whether it uses your context unprompted, fire a trigger word and check the length, then ask something off topic. Whatever fails points straight at the layer that needs rewriting.
Can I just get Claude to write the whole system prompt for me?
Yes, and the guide closes on exactly that. Its final prompt asks Claude to build your complete personal assistant system prompt based on the guide itself. An earlier generator prompt takes your use case and returns something formatted to paste straight into Project Instructions.