Project 03 — A Week's Delegation Audit (4D)

Enforces: the 4D framework — Delegation, Description, Discernment, Diligence (from ai-fluency-framework-foundations) Surface: any (no code) · Time: ~60 min spread over a week · Difficulty: 🟢 starter

Why this project

AI fluency isn't knowing prompts — it's knowing what to hand off, how to hand it off, how to judge what comes back, and how to stay accountable for it. The 4D framework names those four skills. Running it against a real week of your own tasks turns it from a diagram you nodded at into a decision you can defend.

What you'll build

A delegation audit — a single document logging a week's worth of tasks, a delegation decision for each, three written Descriptions, a Discernment rating of the outputs, and a Diligence checklist you actually ticked.

Steps

Capture the week (Delegation) — for ~5 work days, jot every non-trivial task you do. At week's end, tag each one:
- Human — needs your judgment, relationships, or accountability; don't delegate.
- AI — well-specified, low-stakes, verifiable; hand it off.
- Together — you and Claude iterate (draft → you steer → refine). Write one sentence per task explaining the tag — that sentence is the Delegation skill.
Write three Descriptions — pick three tasks you tagged AI or Together and write a full Description for each: the product (what you want), the process (how to approach it), and the performance (role, tone, constraints, format). These are richer than one-line prompts — that's the point.
Run them and apply Discernment — execute each Description with Claude, then rate the output on three axes (1–5): accuracy (is it correct?), fit (does it match what you actually needed?), and usefulness (could you ship it?). Note one concrete flaw in each, even the good ones.
Build a Diligence checklist — for the outputs you'd actually use, tick:
- I verified every load-bearing fact against a source.
- I disclosed AI involvement where it matters (work product, attribution).
- I checked for plausible-but-wrong claims (the confident-and-incorrect failure mode).
- I own the result — if it's wrong, that's on me, not the model.
Reflect — which of the four Ds is your weakest? Most people are fine at Description and weak at Discernment (accepting confident answers) or Diligence (skipping verification). Name yours.

Acceptance criteria — you're done when

You logged a week of real tasks and tagged each Human / AI / Together with a reason.
You wrote three full Descriptions covering product, process, and performance.
You ran them and rated each output on accuracy, fit, and usefulness with one named flaw apiece.
You completed a Diligence checklist (disclosure + verification) on the outputs you'd use.
You named your weakest D and one way to strengthen it.
You journaled the audit in learning-journal-template.

Stretch goals

Re-tag one task you marked "Human" — could a Together workflow have helped? Test it.
Find one task where Claude gave a confident, wrong answer and document how Discernment caught it.
Draft a one-line personal disclosure norm ("I note AI assistance on X but not Y") and justify the boundary.

Self-assessment rubric

Level	Signal
🟢 Got it	You delegate deliberately, write rich Descriptions, and verify before trusting.
🟡 Almost	You delegate well but still accept outputs without independent verification.
🔴 Revisit	Delegation feels arbitrary; re-watch ai-fluency-framework-foundations on the four Ds.