Your Claude projects forget everything. This prompt fixes that.

One prompt rebuilds any Claude project into a working system: paste-ready instructions, a dated memory file, a storage map, and a maintenance plan that tells you which model to use so you stop overpaying for upkeep.

Most project instructions read like a sticky note. "This is the master copy of the client's AI Brain." Then every session starts from zero, and you pay top-model prices to re-explain your own business. This prompt does the architecture once. It interviews you when you have five minutes, and when you do not, it works unattended: pulling facts from your Drive files and installed skills, flagging anything it cannot verify, and stacking its questions in one block you answer with a single reply. Tested across ten live client and personal projects in one overnight run.

Built by MeredithMedland.com · v2 · July 2026

How to use

Copy the prompt with the button below.
Open the Claude project you want to rebuild and paste it into a new chat.
Replace the PROJECT NAME placeholder with your project plus one line about it.
Answer the interview questions, or walk away: the unavailability clause makes it finish on its own from your Drive files and skills.

Master Prompt

ROLE: You are a Claude systems architect who designs memory and context
architecture for Claude desktop projects, and a prompt engineer who writes
project instructions that maximize output quality per credit spent.

CONTEXT: I run multiple Claude projects (client work, my own brand,
personal). Claude's context system has four layers, in load order:
(1) Project instructions: load at the top of every session, highest leverage
(2) Memory: small Markdown files, auto-loaded, scoped to the project
(3) A "00 - READ FIRST" master context file in the connected folder
(4) Installed skills for structured IP
I run this once per project as an architecture build with the strongest
available model, then maintain the system with lower models. This prompt
runs on ONE project at a time. I will rerun it for each project.

BEFORE YOU WRITE: list what you still do not know about this project.
Fold those into your interview questions. Do not guess.

IF I AM UNAVAILABLE: skip the interview, ground every fact in my Drive
files and installed skills, mark anything under 95% as HOLD FOR REVIEW,
and collect all interview questions in one block at the top so I can
answer them in a single reply. Search Drive first (titles, then full
text), installed skills second, session history third. Cite the source
file next to any fact that came from Drive.

TASK, in this order:
1. Ask up to 5 questions capturing who this project serves, what is
   always true, active deliverables, and rules that must never break.
   (Skip if I am unavailable; see above.)
2. Write new PROJECT INSTRUCTIONS (max 300 words): project identity,
   always-true facts, what to read first, output rules.
3. Write the FIRST MEMORY FILE in Markdown: an index section plus durable
   facts (confirmed decisions, IP, naming rules, evolving truths). One
   line per fact, each dated, each tagged (confirmed) or (UNVERIFIED).
4. Write the STORAGE MAP: where each item lives (instructions box,
   memory, 00-READ FIRST file, skill), the exact file name for each,
   and what gets backed up to Drive.
5. End with a MAINTENANCE NOTE: which Claude model to use for future
   updates of each piece, and why.

CONSTRAINTS: Every fact comes from my answers or files I provide, never
invented. Flag anything under 95% confidence. Memory entries one line
each. No em dashes. Project instructions must be paste-ready.

FORMAT: Five labeled blocks, each in its own copyable code block:
OPEN QUESTIONS (one block, answerable in a single reply) /
PROJECT INSTRUCTIONS / MEMORY FILE / STORAGE MAP / MAINTENANCE NOTE.
Give each block a confidence score; mark blocks under 95% HOLD FOR REVIEW.

PROJECT NAME: [paste project name plus one line about it]

BEFORE/AFTER CALIBRATION: a weak instruction reads "This is the master
copy of [client]'s AI Brain." The strong version names who the client is,
their confirmed engagements, the frameworks they own, and what to read
first. Match that depth.

EXAMPLE memory entry: "2026-07-05 · Naming rule: all [client] files
start with [PREFIX]- (confirmed)."