Claude, Escalation Flagger.
This video shows that a Claude skill named escalation-flagger flags items by checking them against numbered criteria in SKILL.md, not by any independent sense of risk.
When a Claude skill named escalation-flagger flags something for a human to look at, is that Claude exercising its own judgment about what looks risky, or is it matching input against criteria written down in a file a person could read and edit? This video answers that question directly by walking through how the skill's matching mechanism actually works.
The file is the whole program
Every Claude skill is a folder with one file, SKILL.md, written in plain sentences rather than code, and that file is the entire program. For escalation-flagger, everything the skill knows lives in that single file: a numbered list of criteria checked one after another. Claude reads the file and then follows it. The file decides what gets flagged, not some separate judgment process inside Claude.
Checked, not sensed
The video names the wrong guess most people make directly: that Claude is sensing which things feel risky, as if it had some independent intuition about danger. It isn't. Claude checks each item on the list, one criterion after another, and flags only where one actually matches. This is demonstrated by removing a criterion from the file: inputs that used to trigger a flag under that criterion stop getting flagged, and nothing hidden fills the gap. The behavior changes exactly as much as the file changes, and no more.
Consistency inside the file's scope
Run the same input through escalation-flagger twice, and it produces the same match and the same flag both times, for as long as that input matches something written in the file. This repeatability is the direct payoff of the flag being decided by a written, checkable list rather than a judgment call that could vary from one run to the next.
What happens outside the file's scope
The guarantee of consistency holds only for input the file actually describes. Send through a case the criteria list never anticipated, and nothing gets flagged. The video is precise about why: not because Claude judged the case safe, but because nothing in the file matched it. A flag's absence is not a safety verdict; it is simply the absence of a match.
The prompt to try
The video's suggested exercise is to write a SKILL.md with three short numbered criteria for flagging something a person deals with often, then hand it inputs on both sides, one that matches and one that doesn't, and watch whether the flag follows the file rather than a feeling. Watching for the flag to fire exactly where a criterion matches, and stay silent everywhere else, is the test of whether the match-not-judgment claim actually holds for a skill written by hand.
Key takeaways
- Escalation-flagger's entire behavior lives in one SKILL.md file: a numbered list of criteria checked in order.
- Claude flags input by matching it against those written criteria, not by sensing risk independently.
- Removing a criterion from the file stops the flags it used to trigger, with nothing hidden filling the gap.
- The same input run twice produces the same match and flag, as long as it matches something in the file.
- An unmatched case produces no flag, which reflects an absence of a match, not a judgment that the case is safe.
Who this is for
This is for anyone building a Claude skill meant to flag or escalate items for human review, and for anyone who wants to understand what a "flag" from a skill-based system actually represents before trusting it.
Chapters
- 0:00Claude's escalation flagger — is it flagging by judgment, or by match?
- 0:11SKILL.md is the file — the anchor: a numbered checklist of criteria
- 0:30Checked, not sensed — the wrong guess, falsified
- 0:47The anchor returns — same match twice, then a case with nothing to match
- 1:06Carry-out
- 1:17Your turn
- 1:32Outro
Full transcript(auto-generated, with timestamps)
Claude's escalation flagger — is it flagging by judgment, or by match?
[0:00]Someone assumes escalation flagger decides what's risky using its own judgement. It doesn't. It matches input against criteria written in skill. MD, so is it flagging by judgement or by match? A Claude skill is just a folder
SKILL.md is the file — the anchor: a numbered checklist of criteria
[0:12]Claude reads before it works. This one is named escalation flagger and everything it knows lives in one file skill.md plain sentences not code. Open it and you'd see something like a numbered list of criteria checked one after another. Claude reads that file then follows it. The file decides what gets flagged not Claude. It can look
Checked, not sensed — the wrong guess, falsified
[0:30]Like Claude is sensing which things feel risky. That's the guess most people make. It isn't. Claude checks each item on the list one criterion after another and flags only where one actually matches. Remove a criterion from that file and inputs that used to trigger it stop being flagged. Nothing hidden fills the gap. Go back to that same skill.md
The anchor returns — same match twice, then a case with nothing to match
[0:49]And its list of criteria. Run the same input through escalation flagger twice and you get the same match and the same flag both times for as long as the input matches something written there. Send through a case the list never anticipated and nothing gets flagged not because Claude judged it safe but because nothing in the file matched it.
Carry-out
[1:06]A flag isn't Claude's judgement call. It's a match against criteria written in a file. Same match, same flag every time. No match, no flag however it looks to a person. Your turn. Here's the
Your turn
[1:17]Prompt. Read it with me. Write me a skill.md with three short numbered criteria for flagging something you deal with often. Then hand it inputs on both sides, one that matches, one that doesn't and watch whether the flag follows the file not a feeling. Liam in for Bear. Claude escalation flagger Liam
Outro
[1:34]In for Bear.
