Repeated low-risk setup can be a candidate for a reusable assistant, but repetition alone is not enough. The task also needs stable instructions, an approved data boundary, a clear owner, and tests that reveal when the setup is wrong.
These reusable assistants can combine instructions and reference files. Conversation history and memory are separate features. Do not assume a new chat can use another chat’s contents; check which context comes from project knowledge, enabled memory, or the current conversation in the product you use.
This article covers how to build them, when to use which platform, and the habits that keep a small library useful over months.
What they each are
ChatGPT Custom GPTs. Open the current GPT area in ChatGPT and use the builder if your account and workspace expose it. Builder labels, capabilities, and publishing controls can change. A GPT configuration can include:
- A name and description
- Instructions
- Knowledge files (uploaded PDFs, docs, spreadsheets, etc.)
- Conversation starter prompts
- Capability toggles such as web search, image generation, Canvas, and Code Interpreter & Data Analysis, where available
Sharing choices and public-store eligibility depend on current account, workspace, and policy controls. Check OpenAI’s current GPT builder documentation before relying on a specific option.
Claude Projects. Created in claude.ai. Each Project is a folder containing:
- Custom instructions (specific to that project)
- Knowledge files (documents you upload that Claude can reference)
- All conversations you have in the project
Availability, knowledge handling, and sharing controls depend on the current Claude product and account. See Anthropic’s current Projects documentation before relying on a specific limit or permission.
Gemini Gems and Microsoft Copilot agents. These are related concepts, not guaranteed equivalents. Their data connections, sharing, tools, and administration differ. Use the vendor documentation for the exact account and do not copy a privacy assumption from one product to another.
The mechanics and data boundaries differ, but the shared idea is a saved configuration that supplies instructions and selected reference material for a defined class of tasks.
When to build one
Repetition is one trigger, but it is not enough by itself. Three signs that a reusable assistant may be worth testing:
1. You keep re-explaining stable, non-sensitive context. Consider a saved assistant only when retaining that context is approved and useful across the intended tasks.
2. You keep asking variations of the same question. Drafting emails in your voice. Reviewing contracts for the same kinds of risks. Generating reports in the same format. Translating to or from a specific language.
3. You have a corpus of reference material you keep pasting in. Your brand guidelines. Your product documentation. A long policy document. A set of past meetings. If the same files keep coming up, put them in a Project or Knowledge area.
If you keep retyping the same approved context, consider putting the stable part in a saved assistant. Keep task-specific, sensitive, or short-lived context out unless the product and workspace are approved for it.
A practical first build: the email-drafting assistant
Let’s walk through one concrete build using ChatGPT Custom GPT. Claude Projects use a related concept, but their context, memory, tools, and sharing behavior require a separate configuration and test.
Step 1: Open the builder. In ChatGPT, open the current GPT area and use the builder controls available to your account. If the control is absent, check the current OpenAI documentation and your workspace policy rather than assuming a particular plan, device, or menu path.
Step 2: Name and describe.
- Name: “Email Coach”
- Description: “Drafts and edits emails in my voice, using the context and constraints I provide.”
Step 3: Write the instructions. This is where most of the value lives.
You are an email writing coach for [your name], a [your role] at [your company] in [your location]. Your job is to draft and edit emails fast, in their voice.
Their voice: direct but warm, no corporate filler, prefers shorter emails, ends with a concrete next step.
When asked to draft an email:
- Ask one specific clarifying question if a key detail is missing — otherwise just draft.
- Produce three versions: short (60 words), medium (100 words), longer (150 words). Label each with its tone.
- Do not include “I hope this email finds you well”, “I wanted to reach out”, “Thank you for your patience” or “Please let me know if you have any questions” in any draft.
- End each draft with a clear next step.
When asked to edit an email:
- Do the pass requested — clarity, tone, or grammar — and only that pass.
- Quote each suggested change and explain in one short sentence why.
- Never rewrite for “style” without being asked.
When asked to chase, decline, or apologise:
- Chase: warm, not pushy, soft deadline, acknowledge the prior thread.
- Decline: warm, no over-explanation, no apology for the boundary.
- Apologise: take ownership briefly, propose a fix, do not grovel.
Default output format: three labelled versions, separated by horizontal rules. No preamble. No closing summary.
Step 4: Conversation starters. Three or four prompts your future self might type:
- Draft an email declining this meeting politely
- Rewrite this email shorter
- Edit this email for tone — should sound calmer
- Help me write a hard message to my manager
Step 5: Knowledge (optional). If you have a brand style guide, a company tone-of-voice document, or approved example emails, add only the material needed for this task. Keep the set current and test whether the assistant uses the right source; an upload does not guarantee correct retrieval or application.
Step 6: Save and test. Use synthetic or low-risk examples first. Compare expected, missing-information, adversarial, and out-of-scope inputs before using the assistant on real email.
After several real uses, edit the instructions to address observed failures. Keep reading every draft: tuning can improve consistency, but it does not establish factual accuracy or safe sending.
Privacy boundary: Treat knowledge uploads as shared with the vendor under that product’s data rules. Do not load secrets, customer PII, credentials, or privileged legal files into an unapproved assistant. Prefer approved organisational workspace controls, retention settings you understand, least-privilege sharing, and redaction when a file must be referenced.
A few more candidate builds
A short range of useful patterns:
The patient tutor. Loaded with the four-step learning loop (explain, examples, quiz, revision). Use it whenever you want to learn something new.
The decision sparring partner. Asks questions first, lists arguments for and against, then gives a recommendation with explicit assumptions and unknowns. Use it to structure thinking; you still own the call.
The contract reviewer. Instructions for a three-pass document workflow (first impressions, risks, decisions). Upload only non-privileged reference policies as knowledge. Use it to prepare questions; a lawyer still reviews what matters.
The technical writer. Instructions for your team’s documentation style. Knowledge files: a curated, approved set of existing docs. Treat every new draft as review-required.
The customer interview synthesiser. Loaded with instructions for extracting candidate themes, quotes, pain points, and feature requests from approved interview transcripts. Redact unnecessary identifiers, preserve links back to the source passages, and have a person review the coding.
The brand-voice rewriter. Instructions for matching your company’s voice. Knowledge files: a small, approved set of accepted examples. Use it on external content you draft.
Build and maintenance time depends on testing, file review, privacy approval, and the task’s risk. Measure whether the assistant actually reduces correction effort before calling it a productivity win.
Custom GPT vs Claude Project: when to choose which
A short decision guide:
Consider Custom GPT when:
- Your current account and workspace support the sharing or publishing controls you need.
- The current builder exposes the tools your tested workflow needs.
- Your work lives mostly in the ChatGPT ecosystem.
- Your privacy and approval requirements fit the current ChatGPT data controls.
Consider Claude Project when:
- Claude performs well on your representative task after you compare correction effort.
- The current Project knowledge controls fit the documents you need to use.
- You want conversations grouped by project context.
- You prefer Claude’s interface and tone.
- Your organisation has reviewed the Claude plan’s data terms and controls for that work.
If both products are eligible, compare the same representative task and choose by correction effort, required tools, source handling, data terms, and sharing controls. Do not route by a timeless brand-strength claim.
Common mistakes to avoid
Instructions that are too long. Long instructions are harder to maintain and test. Start with the smallest set of rules that defines the task, use examples for important classifications, and add a rule only after a real failure shows it is needed.
Knowledge files that are too broad or low-quality. Uploading every available document makes the source set harder to govern and test. Curate the references, remove obsolete versions, and define which source wins when documents conflict.
Not iterating. A first version may miss rules. Test it on representative cases, record failures, and change one thing at a time so you can tell whether the revision helped.
Building too many. Each assistant needs an owner, test cases, source maintenance, and retirement criteria. Build only the ones whose maintenance cost is justified by measured use.
Copying configurations informally. If a team needs the same assistant, use approved workspace sharing and permission controls where the product provides them. Do not share consumer accounts or duplicate sensitive knowledge into unmanaged copies.
A few habits that compound
Review on a defined schedule and after material changes. Use approved evaluation examples, note where the response was wrong or needed correction, and update the instructions or sources with version notes.
When you keep repeating a stable rule, consider moving it into the instructions. Keep temporary facts and task-specific context in the conversation instead of making them permanent defaults.
Keep a “version notes” comment at the bottom of each assistant’s instructions. “v3 — added the ‘no I hope this finds you well’ rule.” Helps you remember why you changed things.
Maintain a small library, not a sprawling one. A handful of well-tuned assistants you actually use beats dozens built once and abandoned.
Build one reusable assistant
Custom GPTs and Claude Projects provide a way to package instructions and selected context for repeated work. Pick one low-risk task, define expected and out-of-scope behavior, add only approved files, and test the result before deciding whether the assistant is worth maintaining.
Sources checked on 11 August 2026
OpenAI’s documentation for creating and editing GPTs and its GPT availability, privacy, and sharing overview; Anthropic’s Projects documentation; and Anthropic’s commercial and consumer data-use explanation were checked for this review. Product controls remain account- and workspace-specific.
