AI Guides › Step-by-step guides
By Nigel Guy · 6 min read
The usual mistake is building ten half-systems in one burst of enthusiasm: a Project here, a clever prompt there, nothing you open on Monday. It feels productive because every setup looks tidy on the day. It fails because none of them has been run on real work, so you never learn which one earns its place.
The rule: build one system, run it on three real jobs, and only then build the next.
You need a Claude account and one recurring job that annoys you. Nothing here needs code.
Menus and plan limits move often. If a path below does not match your screen, search the help centre for the feature name.
| # | System | Built from | Weekly payoff |
|---|---|---|---|
| 1 | Writing-voice project | Project + sample files | Drafts that sound like you |
| 2 | Reply drafter | Project + your past replies | Faster routine email |
| 3 | Meeting-to-actions desk | Project + one template | Notes become owners and dates |
| 4 | Research brief desk | Project + source rules | Briefs with claims separated from guesses |
| 5 | Decision log | Project + log file | Past reasons you can look up |
| 6 | House-format report | Skill | Same layout every time |
| 7 | Answer bank | Project + FAQ file | Consistent customer replies |
| 8 | Morning briefing | Scheduled task + connectors | A summary waiting for you |
| 9 | Weekly review | Scheduled task | A standing Friday write-up |
| 10 | Topic watch | Scheduled task + web research | Regular news on one subject |
Which first? Pick the one you would do by hand this week anyway. For most people that is 1, 2 or 3. Leave 8 to 10 until a Project has worked, because a schedule only repeats whatever quality you give it.
Write two lines: what goes in, what must come out. For example: "Rough meeting notes in; a table of actions, owners and dates out." If you cannot write the second line, you are not ready to build.
Open Projects in Claude, create a new one and name it after the job. A Project holds its own instructions, uploaded files and chats; every chat inside it starts from that context. Paid plans use retrieval to handle larger amounts of content, per the help centre.
Put this in the Project's instructions. It is deliberately short, because instructions you cannot remember are instructions you cannot fix.
You are my working assistant for [JOB, e.g. turning meeting notes into actions].
Context: I am [ROLE] at [ORGANISATION TYPE]. The reader of your output is [AUDIENCE].
Goal: when I paste [INPUT TYPE], return [OUTPUT, e.g. a table of action, owner, due date].
A good result: [TWO OR THREE TESTS, e.g. every action has an owner; nothing invented].
Steps: 1) read the input; 2) list anything missing; 3) produce the output.
Format: [TABLE / BULLETS / SHORT EMAIL], British English, plain words.
Rules: if a name, date or fact is missing, ask me instead of guessing. Mark anything
you inferred as "assumed". Never add details that are not in my input.
Before answering, check the output against the "good result" tests and fix any miss.
Fill in everything in square brackets. Keep the "ask, don't guess" rule: it is the single most useful line.
Upload or paste examples of the finished output you would be happy with: your own past emails, a notes template, a report you liked. Two good examples beat a page of adjectives. Remove anything confidential first.
Use a real input, not a made-up one. Note every correction you make. Each correction is either a missing instruction (add it) or a one-off (ignore it). Repeat on three jobs.
Run the Monday test: on a normal working day, do the job using only the Project, without re-explaining anything. Then check three things.
If the answer to the third is no, the system is a toy. Fix the trigger (put it where you already work) or drop it.