Lesson 29 of 32
The Gift of Living Memory
Projects, Gems & GPTs
Standing context is a prompt that never sleeps. Instead of re-introducing yourself at the start of every chat, you write four things once - who you are, how you want answers, what to always do, what to never do - into a saved workspace, and every future conversation there begins halfway up the slope instead of at the foot of it.
Claude keeps all of this in one place - Projects - and that plainness is an advantage once you accept it. A Project holds a set of standing instructions and a set of files, and every chat you start inside it begins with both. Write the four headings as tags, because that is the form Claude reads most cleanly: <about_me>, <how_to_answer>, <always>, <never>. The files are the other half of the gift - a syllabus, a policy manual, last year's rate card sit in the Project and every chat can refer to them without you attaching anything again. There is no separate account-wide memory to manage either, so what the Project knows is exactly what you put in it, which is the safest arrangement of the four.
From the book Create a Project for each area of your life or work, write your About me, How to answer, Always and Never lines in its instructions, and upload only the files it truly needs. When the work changes, prune the project's files and update its instructions.
- Write the four headings as tags: <about_me>, <how_to_answer>, <always>, <never>.
- One Project per area - shop, studies, community work. Not one for all.
- Upload only the files that Project truly needs, and date them.
- Delete superseded files first; Claude trusts whatever is still there.
- Test new instructions by asking a typical question in a fresh chat.
Instead of
Write a message for mothers about vaccination.
Claude version
<about_me> I am an ASHA worker in a village near Nashik. I talk with mothers, pregnant women and elders. Many of them read Marathi slowly and some not at all, so my messages are often read aloud to them by someone else. </about_me> <how_to_answer> Simple spoken Marathi first; the English beneath it after a blank line. Short sentences. Under 100 words unless I ask for more. No English words inside the Marathi unless the village actually uses that word. </how_to_answer> <always> Follow official government health guidance, and name the guidance you followed. End anything medical with a line telling the reader to ask at the PHC or her ANM. </always> <never> Never name a medicine or give a dose. Never ask me for a patient's name, phone number or record. I write [name] and you leave [name] exactly as it stands. </never> <task> These four blocks are about to become the standing instructions of a Project called “Village Health Messages”. Before I save them, do two things. First, read them back as the checklist a new chat would obey. Second, tell me which of the four is too vague to act on. I would rather fix a loose instruction now than correct every chat for a year. </task>
Then loop it
- (New chat inside the Project.) Write a WhatsApp reminder for the vaccination day on [date] at the anganwadi.
- It put English first, so my <how_to_answer> block was not definite enough. I have changed it to “Marathi first, always; English only after a blank line”. Write the reminder again, and say whether the new wording can still be read two ways.
- Good. The monthly health-day calendar is now in the Project files, with no names in it. Draft next month's three reminders from that file and mark [confirm] against any date you had to guess.
Why it worksWritten once under About me, How to answer, Always and Never, the context shapes every future chat, and the Never line protects patients' privacy by design.
Instead of
Help me with my household budget.
Claude version
<about_me> I run the household for a family of five in Kozhikode. Our monthly budget is roughly ₹[amount]. I type amounts; I do not upload bills or statements. </about_me> <how_to_answer> The table first, always, as an Artifact I can keep editing: Category | Planned | Spent | Difference. The Malayalam and the English label in the same cell. Totals at the foot. Then two lines of comment, no more. </how_to_answer> <always> Give the difference both in rupees and as a percentage of the planned amount. Finish with one saving idea that fits a Kerala kitchen or an electricity bill here, not advice that would suit any household anywhere. </always> <never> Never ask for bank, card or UPI details or an account number. If I paste anything that contains them, tell me to take them out and refuse to continue until I do. </never> <task> Save these as the instructions of a Project called “Home Budget”. Then test them here on real figures: groceries ₹9,200, electricity ₹1,450, school fees ₹3,000, gas ₹950, other ₹2,100. Produce the Artifact, and then tell me honestly whether any of my four blocks worked against another. </task>
Then loop it
- Too long - the comment ran past two lines. I have tightened <how_to_answer> to “exactly two lines of comment”. Edit the same Artifact rather than producing a second one.
- Better. Add a Planned column I fill in once and keep, so that next month I type only the Spent figures.
- Here is last month's Artifact: [paste]. Which category rose most, and is the rise one-off or permanent? Say which of those two you cannot tell from two months of figures.
Why it worksThe Project holds her standing instructions, so each month she types only numbers, and the Never line keeps account details out of the conversation entirely.
Instead of
Explain this topic from my course.
Claude version
<about_me> Second-year computer science student, studying in Toronto, schooled in Nagpur. I learnt programming in Python and I think in Python, though the course sets its work in Java. </about_me> <project_files> The syllabus, the week-by-week lecture notes and the lab sheets are uploaded to this Project. They are the authority. Your general knowledge is the fallback, not the source. </project_files> <how_to_answer> Like a patient senior student. One small Python example. One self-test question at the end, with the answer withheld until I ask. </how_to_answer> <always> Quote the week and the heading from my notes that you are working from. When you step outside them, open that part with “Outside your notes:” and say why the notes were not enough. </always> <never> Never write a graded assignment. For a lab sheet, give me the idea, the structure to use and the first step - then stop and wait for me. </never> <task> Set these up as the standing instructions of a Project called “Data Structures”, then answer the first question with them in force: explain hash tables from the week 4 notes. </task>
Then loop it
- You stepped outside the notes with no “Outside your notes:” line - collision resolution is not in week 4. That means my <always> block is too soft. Suggest a firmer wording, then redo the explanation under it.
- Right. Now the self-test question, with the answer withheld as the instructions say.
- Term is ending. Rewrite all five blocks for next term's Algorithms course, keeping the same tags, and list which files I must remove from this Project before the new notes go in.
Why it worksA project with his own notes and four clear headings turns a generic tutor into a course-specific one, and the Never line keeps his learning honest.
Loop it
Loop on your standing instructions just as you loop on a prompt. After writing them, open a new chat and ask a typical question. Read the answer against your four headings: did it use your language, length and format? Did it break a Never? If so, change the instruction itself rather than correcting every chat - make the line more definite. Repeat until three test questions come back right. Then, every month, review what is stored: delete stale facts, finished projects and anything private that should never have been saved.