1 · The Project recipe
the Monday move
Take your two artifacts → create a Claude Project → name it after the pattern. The rest of this section is just the rule for what goes where inside it.
Worked example — turning the workshop output into a real Project on Monday morning. Five steps, ~10 minutes.
-
Step 1 — open Claude.ai → Projects → New Project
Name it after the pattern + slot, not the task. Good:
Drafter — User Stories. Bad:Member-Consent-PRD-stories. The pattern repeats; the task doesn't. - Step 2 — paste Context Card on top + RCTCO prompt below into custom instructions Both stay there forever. Every new chat in this Project starts with both loaded. See the "Custom instructions" tab for the exact layout.
- Step 3 — upload 3-5 reference files into knowledge Past shipped artifacts (last 3 PRDs that passed review), the domain glossary, the 5-tests rubric, any spec template. RAG-retrieved on demand — not always-loaded.
- Step 4 — start a chat, fill the Context placeholders, ship Each new task = new chat in the same Project. The role / template / output schema are already loaded; you only fill what changed.
- Step 5 — invite teammates, save the URL in your team's Notion/Confluence One pinned page: "Team Prompts." Five links. That's the entire library on day 1.
If it changes per task → knowledge file.
If it applies every chat → custom instructions.
If it applies every chat → custom instructions.
Why this rule? Custom instructions get injected into every conversation in the Project — they consume tokens always. Knowledge files are RAG-retrieved on demand. Putting variable stuff in instructions wastes tokens; putting always-needed stuff in knowledge means it sometimes doesn't surface. Source: PromptKits, 2026 + Anthropic Help — What are Projects.
- Custom instructions = "the contract" Role, output schema, constraints, voice rules, Context Card. Things you'd say at the start of every conversation if you had to. Keep it tight — every token here costs every chat.
- Knowledge files = "the library" Past artifacts, glossaries, reference specs, rubrics, examples. Things Claude pulls when the task warrants. RAG-retrieved on paid plan (10× capacity).
Layout for a Drafter Project: Context Card on top (sets the stage), RCTCO prompt below (the actual contract). Paste this into the Project's custom-instructions field, swap the bracketed bits for your team's reality.
# ===== CONTEXT CARD =====
# Always-true background about this team. Stays loaded every chat.
Team: [WebMD Health Services Vietnam — PM team]
Product: [member-facing wellbeing platform]
Audience: [US enterprise HR buyers, end-user is employee]
Voice: clear, US-English, no jargon, evidence-first.
Compliance non-negotiables:
- HIPAA-aligned PHI handling
- PDPL (VN data privacy) for cross-border
- Accessibility: WCAG 2.1 AA
Glossary: see [glossary.md] in knowledge.
Past shipped reference: [last-3-prds.md] in knowledge.
# ===== RCTCO PROMPT — DRAFTER =====
# The contract. Every new chat in this Project runs under this.
ROLE: You are a senior product manager drafting user stories
for the team above.
CONTEXT: The user will paste a feature brief or PRD section
in their first message. Treat it as the only source of truth
for the feature; do NOT invent business rules.
TASK: Generate 3-7 user stories in INVEST format covering
the feature. Group by user role.
CONSTRAINTS:
- Each story: "As a [role], I want [capability], so that [outcome]"
- Each story has 3-5 acceptance criteria in Given/When/Then
- Flag any story that touches PHI with [COMPLIANCE — HIPAA]
- Flag any cross-border data flow with [COMPLIANCE — PDPL]
- If the brief is ambiguous, ASK — don't guess
OUTPUT: Markdown. One H2 per role. Stories as H3.
End with a "Open questions" section if anything was unclear.
Knowledge files for this same Drafter Project: last-3-prds.md, glossary.md, 5-tests-rubric.md, invest-criteria-examples.md, compliance-tag-cheatsheet.md. None of those need to be in the prompt — they get pulled when relevant.
Apply the rule to 8 common items. The point isn't memorizing a list — it's training the muscle so any new item lands in the right slot.
| Item | Goes where | Why |
|---|---|---|
| Output schema (markdown headings, INVEST format) | Custom instructions | Applies every chat. The shape of the answer never varies. |
| Voice rules ("clear, US English, no jargon") | Custom instructions | Same. Voice is constant per Project. |
| Last 3 shipped PRDs (reference for style) | Knowledge file | Useful as examples but only retrieved when relevant. Don't pay tokens every chat. |
| Domain glossary (50+ terms) | Knowledge file | Big, mostly-irrelevant-per-chat. RAG pulls the few terms that match. |
| Compliance non-negotiables (HIPAA / PDPL flags) | Custom instructions | Must apply every chat or it'll be silently skipped on the chat where it matters. |
| The 5-tests rubric for "is this a good story" | Knowledge file | Reference doc. Pulled when reviewer-mode questions come up. Optional. |
| Today's PRD section being turned into stories | Paste in chat | Per-task input. Lives in the conversation, not the Project shell. |
| Team's Context Card (audience, product, team) | Custom instructions | The reason the Project exists. Always loaded. Top of the instructions. |