Compress the whole binder into one page a director can read in three minutes — then rehearse delivering it and answering the hardest questions
Day 59 of 66
A director will never read your full binder. They'll read one page, and decide in three minutes whether to fund the program and trust you to run it. Today you write that page. The binder is the proof; the one-pager is the pitch. Both have to exist — but only one gets read first.
If you can't get the program onto one page, you don't understand it well enough yet. Every cut forces a decision about what actually matters. The discipline of fitting it on one page is the same discipline that scoped the program on Day 56 — say the important thing, drop the rest, trust the binder to back you up.
A director-readable brief has exactly five parts, in this order. Anything else goes in the binder, not the page.
Problem. One sentence with a number. "Grok image prompt-adherence sits at ~72%; users get images that drop part of the request."
Plan. Two or three sentences. What data, scored how, why it'll move the metric. Name the targeted-sampling insight from your analysis.
Metric. One line. The held-out reading that proves success — gold-set accuracy plus a milestone human-eval win-rate.
Timeline. A few milestones with dates. Week 1: rubric + pipeline live. Week 3: first 1,500 pairs + interim eval. Week 6: final eval, verdict.
The Ask. The single most important line — what you need from the director to start. Headcount, annotator hours, eval budget, a decision. Be specific and bounded.
People describe the work beautifully and then never say what they need. A director's job is to allocate; if you don't name the ask, you've handed them a story instead of a decision. "I need three calibrated annotators for six weeks and a one-time human-eval budget of N comparisons" is an ask they can approve or counter. "I'd like to improve adherence" is not.
The brief is also a script. You should be able to walk it in under three minutes, out loud, without reading it. This is a Field Day — the work is performance, not writing.
Read the brief aloud, then close it and deliver it from memory. Where you stumble is where your own understanding is thin — fix the understanding, not just the sentence. Record yourself once. The first run is always too long and too hedged; the third run is the one a director would hear as "this person can run a program."
A director's value is in the hard questions. Prepare the three most likely ones with answers that point at artifacts, not at confidence.
| The tough question | Your answer (points at an artifact) |
|---|---|
| "How do I know the data is any good — not just busy?" | "Gold-set accuracy and inter-annotator agreement, audited weekly (Week 3 + Week 9). I can show you last week's kappa and the drift trend." |
| "How do you know this moved the model and not the season?" | "A pre-registered, held-out eval with position-bias controls (Week 6), measured against the current model — contamination-checked before each run." |
| "What if it doesn't hit 85%?" | "The interim eval at Week 3 is a kill/adjust gate. If the lift is short, the failure-cluster analysis (Week 7) tells us where to re-target before we spend the back half." |
Notice the pattern: every answer ends at something you built. That's the whole reason the binder exists — so that under pressure, you reach for evidence instead of adjectives.
An IC can describe their work. A lead compresses it to one page and ends on the ask — names the headcount, hours, and eval budget a director can approve or counter — then handles the three hardest questions with answers that land on an artifact, not on confidence.
Say this in an interview: "I can put a whole data program on one page — problem, plan, metric, timeline, and a specific ask — and walk it in under three minutes. When you push on it, every answer ends at something I built: the agreement audit, the held-out eval, the cluster analysis."