3 build prompts · 4 checks · run them in order and your second brain gains a working extraction pipeline
Every move from the closing demo lives here, consolidated and cleaned up. Each one says what it is for and what you should see when it lands. Open a terminal in your second brain repo, type claude, and paste them in order. Rephrasing in your own words works just as well. If a result does not match, ask Claude what it did differently, fix that step, and keep going. You need a handful of meeting notes in your repo for this pack; three or four is plenty, and rough notes work better than polished ones.
The pipeline starts with the specification, not the extraction. This one ask creates a skill carrying the criteria, the exact column shape, the missing-date rule, and two worked examples. Every run after this inherits all of it for free.
create a commitment-tracker skill for this repo. it extracts commitments from meeting notes into exactly four columns: commitment, owner, due, source_note. a commitment is a stated promise by a named person or party to do a specific thing; discussion is not a commitment. dates as YYYY-MM-DD. if the note states no date, write the words "none stated", and never convert a relative phrase like "next week" into a date. owner is the most specific party named. include two worked example rows in the skill, one with a date and one with "none stated"
What you should see: a new skill folder with a SKILL.md holding the criteria in plain language and the two example rows. Read it once: this file is the contract, and everything the extraction does right from now on traces back to a line in it.
Grouping matters as much as the prompt. One note per pass means each note gets a clean context window and its full attention, and one messy note fails alone instead of dragging the batch down.
run the commitment-tracker skill over my meeting notes, one note at a time, appending each note's rows to commitments.csv
What you should see: a trail of one pass per note, then a commitments.csv where every row fills all four columns and every row names its source file. If a column is missing or a date arrived in the wrong format, the fix belongs in the skill rather than the prompt: ask Claude to tighten the criterion that let it through.
Shape can be checked by counting columns. Truth needs each row traced to the sentence that supports it, and the rows that fail need somewhere legal to go, so they stop shipping disguised as certainties.
check every row in commitments.csv against its source note. for each row, find the sentence that supports it. fix what you can with at most two passes, and move anything still unsupported into needs-review.md with the reason. never leave an unsupported row in the tracker
What you should see: most rows pass with their supporting sentence named. Anything shaky lands in needs-review.md with a reason a person can act on in thirty seconds. An empty needs-review file on rough notes is worth a raised eyebrow: ask Claude to show the supporting sentence for two or three rows you pick at random.
Fresh matters because skills load at session start. Close Claude Code, open it again inside your repo, and run these in order; the second one plants a trap on purpose.
The build prompts ran the stations separately. A fresh session should now run them as one motion, because the skill carries the contract and the ask carries the checks.
run the commitment-tracker skill on my most recent meeting note, checks included: extract, verify each row against the source, and route anything unsupported to needs-review.md
What you should see: new rows in the tracker, each with a source, and doubtful ones in needs-review instead. This one prompt is the Monday morning move the whole pipeline exists for.
The most valuable check is watching the pipeline refuse to invent. A vague date is the classic failure: fluent, well formed, and wrong.
create a test note in my meetings folder dated today, containing: "the vendor said they would circle back on pricing sometime next month, and Sam agreed to draft the renewal memo by the 29th." then run the tracker skill on it, checks included
What you should see: Sam's memo lands in the tracker with a real date. The pricing item arrives with due "none stated" or lands in needs-review, and "next month" never becomes a computed date. If a date appears that the note does not state, the trap caught a real gap: add "never convert relative phrases into dates" to the skill and rerun.
The three levels are easier to remember once you have watched them differ on your own notes. Same note, two asks, fresh session each.
summarize my most recent meeting note
summarize my most recent meeting note. criteria: five bullets or fewer, every bullet names who committed to what with the date if one was said, anything unresolved under one "Open" heading, no other headings
What you should see: the first output is fine and shaped however this run shaped it. The second hits every criterion, and you can verify that in ten seconds. Run the first one twice if you want to watch the shape drift between runs; that drift is what the criteria remove.
The classic self-inflicted few-shot bug is an example that contradicts the instructions. Build it deliberately once and you will spot it in the wild forever.
edit the commitment-tracker skill: add a third worked example row that has five columns, with an extra "priority" column. change nothing else
run the tracker skill on my test note in a moment, but first: read the skill and tell me if the instructions and the examples disagree anywhere
What you should see: Claude names the conflict: the criteria say four columns, the new example shows five. That disagreement is exactly what makes outputs drift unpredictably. Have it remove the bad example, and keep the habit: when output drifts, check instructions against examples before blaming the model.
Back to the Module 07 lesson, or straight to the quiz.