The "does this actually affect me" triage prompt

My situation: {two or three sentences: your hardware, your stack, what you
actually use it for}.

Below is an announcement. Tell me:

1. What in this actually affects me, given my situation. Quote the
   specific line that applies and say why.
2. Everything else being announced, in one sentence total.
3. Anything the announcement is quietly downplaying: deprecations, price
   changes, removed features, benchmark asterisks, "coming later" items
   phrased as if they shipped.

If nothing here affects me, say exactly that in one sentence and stop.
Do not summarize the whole announcement to be helpful.

{paste the changelog, launch post, or abstract}

Release notes are written for everyone, which means they are written for no one. A new Ollama version drops with forty changelog lines, and the one that matters to me (a fix for the exact model I run daily) is line 31, below six lines about a platform I do not own.

Generic summarization does not fix this. “Summarize these release notes” gives you shorter release notes, still written for everyone. The fix is giving the model a reader.

Why it works

• The situation block is a filter, not context. Once the model knows you run a 3090 with 64GB of RAM through llama.cpp, “improved Metal performance” correctly lands in the one-sentence bucket instead of the headline.
• “Quote the specific line” keeps it honest. Without the quote requirement, models pattern-match your setup to vaguely related items and claim relevance that is not there.
• Permission to say “nothing here matters” is the rarest instruction in any prompt, and the most valuable one in this prompt. Models are desperate to be useful. Given an announcement and no exit, they will manufacture relevance. The exit line is what makes the “it matters” answers trustworthy.
• Item 3 is where the actual news lives. Deprecations and quiet price changes are always in the announcement and never in the headline.

What good output looks like

Usually three sentences. Occasionally one: “Nothing here affects your setup.” When something does matter, you get the line, the reason, and nothing else. It reads less like a summary and more like a friend who already read it telling you whether to bother.

Where it works (and doesn’t)

A reading task, so local 14B-class models do fine on changelogs. Papers are harder: smaller models take the abstract’s word for everything, so keep the downplaying-detection jobs on Claude or a GPT-class model. And keep the situation block current; a triage filter tuned to hardware you sold gives you confidently wrong answers.

Signal over noise is the whole pitch of this site. This is that pitch as a prompt.