Before you build, decide what you're building. Four steps: describe the work, name what's reusable, pick the container, draft the artifact.
Pick one Claude move from the last two weeks. Before you can pick the right container, you have to describe the work precisely. Vague descriptions produce mis-built containers.
A specific deliverable, not a category. "Draft a partner email about the deadline shift" — not "partner communications."
Internal team, named stakeholder, partner, funder, scholar, family. The audience shapes everything downstream.
Once. Once a quarter. Once a week. Every time a partner emails. This is the question that decides one-shot vs. pattern.
The standard. What separates a draft you'd send from one you'd rewrite. Be concrete.
Re-read your Step 1 answers. The container you build depends on what — if anything — you'd reuse. Pick the one that fits best.
One or two sentences. "Same partner list and history, varied asks" → context. "Standard email shape, varied recipients" → move. If you picked "nothing," say why in one sentence — that's the discipline of the deliberate one-shot.
Each reusable thing maps to one container. Click your choice — Step 4 will show only the artifact template you need.
Fill in the panel for the container you picked. Bring the result to AI Friday.
No artifact to build — but write the one sentence that proves you considered the alternatives. This is what stops your team from rebuilding the same prompt every Tuesday.
One sentence. What makes this work different enough that turning it into a Project or Skill would be over-engineering? Name the variable that won't repeat.
Including the context you'd otherwise lose: audience, standard, stakes, timing.
A Claude Project is a workspace with a system prompt and reference files. The same context supports many different asks. Sketch it here, then build it in Claude.
Short, clear, names what it's for. "Partner Engagement," "FY26 Board Materials," "Scholar Disbursement Reporting."
Who Claude is in this Project. The voice, the standards, the recurring rules. ("You're a development associate at DC CAP. The voice is warm and direct. Never reuse phrases from the AI tell-tale list...")
Bullet list. The knowledge that makes the Project useful — and the data tier of each one. (Partner contact list — Tier 2. Prior emails — Tier 2. Voice guide — Tier 4.)
If you can't list three, it's probably a Skill, not a Project.
A Skill is a short written procedure — a SKILL.md — that anyone on the team can run against varied inputs. The move stays the same; the inputs change.
Action-shaped. "Draft a Partner Email in DC CAP Voice." "Summarize a Funder Update for the Board."
The condition that should make a teammate reach for this Skill instead of starting from scratch.
The output standard. Tone, length, structure, things it must include.
The failure modes. The AI tell-tale words. The tones that don't fit DC CAP. The shortcuts that look right but aren't.
If the partner's name, the topic, and the deadline change every time — list them here. That's how a Skill differs from a Project: nothing is preloaded.
An Automation runs on a cadence or an event. No human has to remember. Most DC CAP automations should keep a human checkpoint before anything goes out the door.
Names the cadence or trigger. "Quiet Partner Watch." "Monday Funder Pulse."
When does this run? "Every Monday at 8am." "Whenever a partner reply lands in the dev inbox."
What does the automation read each time it fires? A list, a calendar, an inbox, a folder.
What does the automation produce? A digest, a draft, a notification, a row in a sheet.
Where does a human review before anything external happens? If the answer is "nowhere," reconsider — that's an Agency-mode delegation and warrants a leadership conversation first.
Automations that touch Tier 1 or 2 data require explicit governance review before standing up. What tier is yours?
One-shot: "Draft a note to Georgetown's interim dean explaining why our timeline shifted." The dean is interim. The shift was a one-time event. Why one-shot: the variable that won't repeat is the interim dean herself.
Project — Partner Engagement: system prompt names the 13 partners and DC CAP voice; files include partner contact list (Tier 2), MOU terms (Tier 3), prior emails (Tier 2), voice guide (Tier 4). Asks: draft an email, write a meeting brief, compare three partners' responses.
Skill — Draft a Partner Email in DC CAP Voice: trigger on any partner email; good = warm, professional, names the relationship; avoid = generic openers, formal closers; inputs supplied inline = partner name, topic, deadline.
Automation — Quiet Partner Watch: trigger every Monday 8am; input = partner contact list + email log; output = list of partners untouched 30+ days with one-line suggested reach-outs; human checkpoint = development associate reviews before any reach-out goes out; tier = 2.