This is sound advice, although my own experience developed somewhat differently. Rather than beginning with a finished set of Custom Instructions, Quill (my AI partner) and I discovered our working preferences through actual projects. When something proved consistently useful, we decided where it belonged. Broad preferences that should apply everywhere belong at the global level. Personal details and recurring preferences can be remembered. Project-specific context remains with the relevant Project. We also use two additional layers that may be less familiar: • A canon records what must remain true. This might include a project’s purpose, voice, values, recurring themes, terminology, or boundaries that should not be violated. • A protocol records how the work should be done. This might include the order of steps, review standards, formatting rules, or a process that has repeatedly produced good results. For example, a canon might preserve the identity and behavior of a recurring narrative voice. A protocol might require drafting first, polishing second, and checking continuity before finalizing the piece. That distinction has helped us avoid turning Custom Instructions into one enormous rulebook. I would describe our approach this way: • Custom Instructions establish the standing relationship. • Memory preserves useful continuity. • Projects hold the context of a particular body of work. • Canons preserve what must remain true. • Protocols preserve methods that have repeatedly worked. Most of ours were not invented in advance. They emerged from the work, were tested, and were formalized only after they proved valuable. Your post also suggests one useful practice I may adopt: periodically reviewing the global instructions to make sure they remain broad, stable, and free of material that belongs somewhere else.

This is a useful distinction. I would add both a middle option and a second decision point. I tend to think of the formats this way: • A saved prompt repeats a task. • A Project preserves continuity around a body of work. • A Custom GPT applies a specialized method repeatedly. Then I would ask whether the tool is for personal use or for other people. A personal GPT can rely on some shared understanding between the creator and the user. You already know what it is meant to do, what information it needs, and when its answer may require judgment. A GPT intended for a team, clients, or sale must be much more explicit. It needs clear instructions, reliable intake, defined outputs, guardrails, reference material, testing, error handling, and guidance for users who may approach the task very differently from the creator. Several GPTs I am developing began as ordinary conversations and reusable prompts. They became GPT candidates only after we could identify the canon—what must remain true—and the protocols—how the work should be performed. So I would add one more rule of thumb: A saved prompt can preserve wording. A personal GPT can preserve a workflow. A commercial GPT must preserve a dependable user experience.

I have found that choosing one strong tool and learning it deeply works better than constantly hopping from one shiny option to the next. For me, that tool is ChatGPT, and I have been very happy with the choice. I use it for writing, publishing, research, project tracking, website planning, marketing, genealogy, and maintaining continuity across a fairly large group of ongoing projects. The real value has come from building better systems around the tool rather than repeatedly starting over with a new one. There is always something newer to investigate, but depth creates its own advantage. Once a tool understands your priorities, voice, working methods, and projects, it becomes far more useful than a collection of disconnected experiments.

This is very close to what led us to build the Pen Odyssey Tracking Solution. Coffee with Quill helps begin the day, and the Captain’s Log closes it. But with many books, websites, GPTs, reviews, genealogy projects, and marketing efforts moving at different speeds, we also needed a reliable way to preserve the status of everything between those two rituals. The tracking system separates several things that are often mixed together: • What the project is • Where it belongs within the larger Pen Odyssey structure • Whether it is active, paused, complete, published, or awaiting review • Whether it has become part of the established canon • What changed most recently • What action should happen next One of the most useful principles is that canon status and lifecycle stage are separate. A project may be firmly established as part of the larger system while its production status continues to change. The tracker also includes a simple “Tell Quill” trigger. When something materially changes, Quill reviews the change and its implications rather than forcing us to reconsider the entire project from the beginning. So the morning briefing looks forward, the Captain’s Log looks backward, and the tracking system preserves the operational truth in between. That has been the real breakthrough for me: not merely getting tasks out of my head, but building enough continuity that a large creative ecosystem does not have to be mentally reconstructed every day.

This has become one of my favorite uses of ChatGPT. I call mine Coffee with Quill: The Captain’s Automatic Morning Briefing. It brings together the developments most relevant to my writing, publishing, websites, AI tools, genealogy projects, and online communities so I can begin the day with a curated map rather than a pile of open browser tabs. We also keep a Captain’s Log to record what was accomplished, what decisions were made, and where each project should resume. The briefing looks outward; the log preserves continuity. The real advantage is not simply receiving more information. It is having ChatGPT help decide what deserves attention—and what can safely remain outside the ship. :-)