Even a US Senator Knows Vague Questions Get Vague Answers

One Letter To The NSA

On September 2, 2026, Sen. Ron Wyden sent a letter to the NSA director. He was not asking for a general briefing on VPN safety. He asked for updated configuration guidance and a specific answer on single-hop versus multi-hop tools, with a deadline of mid-October.

Notice what he did not ask for. He did not ask "is a VPN good for privacy." That question gets you a pamphlet. It gets you a paragraph everyone has already read a hundred times, technically true and useless for making a decision. Wyden asked a narrower question because narrower questions force a specific answer. Single-hop or multi-hop. By this date. From this office.

This is the exact failure mode content teams run into with AI writers every day, and most people do not connect the two.

Ask an AI tool to "write a blog post about VPN security" and you get something fluent. It reads fine sentence to sentence. It is also generic enough to apply to almost any brand, which means it applies well to none of them. The model did what you told it to do. You just did not tell it much.

Wyden's letter is a brief. It names an audience (NSA leadership), a decision that needs making (single-hop or multi-hop), and a deadline that forces specificity. That structure is what separates a usable answer from a fluent one, whether the recipient is a federal agency or a language model.

Why Fluent Isn't Usable

Most people typing a prompt into an AI writer think they are giving instructions. What they are actually giving is a vibe. Topic: VPN security. Tone: professional. Go.

The model complies. It always complies. It produces sentences that scan correctly, use the right terminology, and hit the expected structure of a blog post. Nobody reading the output would flag it as broken. That is exactly the problem. Fluent and correct are not the same thing as usable, and the gap between them is where editors spend their afternoons rewriting from scratch.

A May 2026 guide on content briefs made this point about keywords specifically: specifying search intent and reader level produces structured, useful output, while keywords alone just produce word-association filler. The AI has no reader level unless you give it one. It has no idea if the person reading is renewing a VPN subscription or deciding whether their company needs one at all, so it writes something vague enough to half-answer both.

An April 2026 workflow piece put it more bluntly. Skipping a one-page brief that covers audience, core argument, success criteria, and banned phrases is the single biggest reason AI drafts read like AI. Not the model. The brief.

The Six-Part Brief

An August 31 piece in PR Daily gave the fix a name: the six-part brief. Instead of typing a topic and a vibe, you convert your prompt into six specific parts before you ever hit generate.

Role comes first. Tell the model what seat it is sitting in. Not "write like an expert," but "you are a cybersecurity reporter briefing non-technical readers."

Audience is next, and it needs a name, not a category. Not "general readers." Try "small business owners who just heard about the Wyden letter and want to know if their VPN is fine."

Task is the actual deliverable. One paragraph explaining single-hop versus multi-hop, aimed at someone deciding whether to switch.

Tone needs concrete terms, not adjectives like "professional" or "friendly." Try "direct, no hedging, short sentences."

Constraints are your banned phrases and your boundaries. No fear-based language. No claims about specific VPN brands.

Format closes it out. Three short paragraphs, no headers, no bullet list.

A June 2026 piece out of a design school called a simplified version of this the Role-Context-Task-Tone pattern, built specifically to close off the categories where AI writing drifts generic. Same idea, fewer boxes. Either way, the brief is doing the work the prompt was supposed to do.

What Changes In The Draft

Run the six-part brief through an actual draft and watch what happens to the sentences.

Without a brief, an AI writer asked to cover the Wyden letter produces something like: "Senators are increasingly concerned about digital privacy in an evolving cybersecurity landscape." True. Says nothing. Could sit on any blog published this decade.

With the brief filled in, role set to cybersecurity reporter, audience named as small business owners deciding whether to switch VPNs, tone specified as direct with no hedging, the same model produces something closer to: "Wyden wants a specific answer by mid-October. If the NSA says multi-hop, expect single-hop VPN vendors to start losing enterprise contracts fast." Different sentence. Different verb choices. It commits to something because you told it what committing looks like.

The April 2026 workflow finding backs this up directly. Teams skipping the one-page brief kept getting drafts that read like AI, not because the model got worse, but because nothing in the prompt forced a specific claim. Banned phrases matter here too. Cut "leverage," "landscape," and "in today's world" from the constraints section and the model stops reaching for them, because it was never using those words on purpose. It was using them because nothing told it not to.