DC CAP AI Pilot · Week 4

The Build Filter

Five questions decide whether the work in front of you deserves a Project or Skill. Score honestly. Three "yes" or more, you build. Fewer, you put it down — pick a different candidate.

1Name the candidate

From last week's Container Decision Worksheet, you have a candidate Project or Skill in mind. Describe it here in three short fields. Be specific — this is the artifact you're scoring.

The work, in one sentence
What does this Project or Skill do? Audience, output, cadence.
Container
Project (reuses context), Skill (reuses the move), or Automation (reuses the trigger).
Who inherits it
Just you? Your team? A specific role? The whole pilot cohort?

2Run the filter

Score each criterion. Yes / No / Maybe. The verdict updates as you click. The questions are deliberately blunt — vague candidates fail vague questions.

01 Frequency
Does this work happen often enough to justify the build? Count the last 90 days. If you wouldn't run this build at least three or four times a quarter, the maintenance cost outweighs the time saved.
Yes if: 3+ times per quarter, ongoing.
02 Variability
Does the move repeat with new inputs, or is each instance genuinely different? If the procedure is identical and only the inputs change, you have a Skill. If every instance demands new context but the same body of knowledge, you have a Project. If neither shape fits, the work might still be a one-shot.
Yes if: the shape of the work is stable, even if the inputs vary.
03 Inheritance
Will the team actually use this — or is it a private convenience? Private convenience is fine. The honest answer is the leadership move. If only you will run it, build a Project for yourself. Don't sell it as a team asset until a teammate has actually run it cold.
Yes if: at least one specific named person other than you will use it.
04 Stakes
What does getting this wrong cost? High stakes earn deeper customization investment, tighter governance, and a real human checkpoint. Low stakes can stay rough. If the cost of an error is "we redo a draft," you don't need much. If the cost is "we damage a partner relationship" or "we publish a wrong number," you do.
Yes if: an error reaches a stakeholder before a human catches it.
05 Substitution
Does a better tool already do this? Salesforce, the Common App, the budget system, an existing template, a colleague who already runs this manually well. If the answer is yes and the existing tool is good, don't build a Claude version of it. The right answer here is sometimes "no build."
Yes (build) if: no existing tool, or the existing tool is genuinely worse than what Claude can produce with custom context.

3Read the verdict

The score updates as you click. 3+ yes = Build. 1-2 yes = Pause and reframe. 0 yes = Pass — this isn't the right candidate. The deliberate "no" is leadership, not a failure.

Verdict 0 yes · 0 maybe · 0 no
Score the filter to see the verdict
Click yes / maybe / no on each criterion above. The verdict updates in real time. There is no minimum number of "no" answers required to pass — the goal is honesty, not consensus with yourself.

4Plan the customization

If the verdict says build, the next question is depth. A team build is only worth keeping if it carries something Claude doesn't already know. Three vectors decide depth — load at least one source against each before you ship v1.

Vector 01
Org context
DC CAP voice. Verified ground-truth metrics. Partner list. Family- and counselor-facing tone rules. Governance constraints.
Vector 02
Outside authority
Research papers. Federal datasets. Peer-org playbooks. Brand guides. Anthropic skill docs. Anything that raises the floor on what Claude can do for this task.
Vector 03
Workflow shape
When to trigger. What good looks like. What to avoid. Where the human checkpoint sits.
Data tier check
What is the highest data tier (1-4) that any reference file or context you're loading carries? Anyone with access to this build inherits access to the references.

5Plan the hand-off

A build only earns the "team-level" label after a teammate has actually run it. Decide who, when, and how you'll watch.

The teammate
Specific name. Pilot cohort preferred — they speak the same vocabulary.
What you'll watch for
Where it broke is where your customization is thin. Name the failure modes you expect before you watch.
Bring to AI Friday
One sentence on what you'd cut, sharpen, or rename in v2 — based on the hand-off, not your own taste.