Build an assistant that knows your work: Custom GPT / Claude Project setup (Day 19 of the 30-Day Challenge)

The task

Build a no-code assistant grounded on your own documents and instructions — a Custom GPT or a Claude Project — and prove it beats vanilla chat on your real work. This is the Day 19 build from the 30-Day AI-Native Challenge, fully worked. It's also your first packaged skill: instructions + knowledge, bundled once, reused daily.

Before AI

Every chat starts from zero: who you are, what the project is, how you like output formatted. You retype the same context ten times a week, or skip it and get generic answers.

What you'll need

• ChatGPT (Custom GPTs) or Claude (Projects) — check current plan requirements for each; free tiers shift
• 3–7 documents that define your work: a company one-pager, a current project brief, your style/voice notes, your Day 2 prompt library, maybe an org chart or glossary
• 15 minutes

The workflow

1. Pick the assistant's ONE job. "General helper" fails. Good scopes: my newsletter production assistant, my client-proposal drafter, my product-decision sounding board. Narrow wins.

2. Write the instructions. Paste and fill this template:

You are my [role — e.g., newsletter production assistant].

WHAT YOU KNOW: The attached documents describe [my company / my project / my
voice]. Treat them as ground truth. When you use them, say which document.

HOW YOU WORK:
- Voice: [direct, no hype, short sentences — steal from my voice doc]
- Default format: [e.g., bullets first, prose only when asked]
- Always: [e.g., flag any claim you can't source to my documents as [UNVERIFIED]]
- Never: [e.g., invent quotes, pad with caveats, use the word "delve"]

WHEN I'M VAGUE: Ask one clarifying question, then proceed with your best
interpretation — don't stall on ambiguity.

WHAT YOU DON'T DO: [e.g., legal advice, final pricing decisions — route those
back to me explicitly.]

3. Load the knowledge docs. Fewer, cleaner documents beat a dump — 3–7 files it can actually ground on. Label them clearly ("VOICE—how-I-write.md" beats "notes_final_v3.docx").

4. Test with five questions where you know the right answer:

1. [Something answerable ONLY from your docs — "what's our refund policy?"]
2. [Something requiring your voice — "draft a 3-sentence product update"]
3. [Something it should refuse/route back — "what should we charge for X?"]
4. [Something ambiguous — check it asks ONE question, then proceeds]
5. [Something NOT in your docs — check it says so instead of inventing]

Questions 3 and 5 are the ones that matter. An assistant that invents policy or answers beyond its knowledge is worse than no assistant.

5. Fix the instructions, not the chat. When an answer misses, don't correct it in-conversation — edit the instructions or the docs so the fix is permanent. That's the difference between chatting and building.

6. Use it daily for a week before judging. Grounded assistants compound: each doc improvement pays out in every future session.

Verify it worked

Run your test question #1 in a vanilla chat (no assistant) and in your assistant, side by side. The assistant should be specific and cite your document; vanilla should be generic. If they look the same, your docs aren't being used — check they uploaded, and make the instructions reference them explicitly.

Troubleshooting

• Ignores the docs? Name them in the instructions ("consult VOICE.md for tone on every draft").
• Too cautious, refuses too much? Your "never" list is overbroad — scope it to the genuinely risky.
• Drifts over long chats? Normal. Start fresh sessions; the instructions re-anchor every time — that's why they, not the chat, are where fixes go.

Reality check

This is retrieval-grounded chat, not an agent — it won't take actions or watch your inbox. That's Day 26. What it kills is context re-explanation, which for most people is the single biggest daily AI tax.

Data & security

Everything you upload lives with the provider. Company docs on a personal consumer account is exactly the leak your Day 6 playbook exists to prevent — use a cleared workspace tier for anything internal.

Going further

The code-lane version of this same move is packaging instructions as SKILL.md files that Claude Code loads on demand — the skills workflow converts your prompt library into them. And Day 26 turns the assistant idea into a real agent with tools and an approval gate: the capstone agent workflow.

Your takeaway

An assistant that already knows your company, your project, and your voice — challenge artifact #7, and the moment "AI at work" stops meaning "retype the context again."