// Grounded retrieval over cohort 4 memory seeds (roster, lesson designs, dossiers, transcripts). Use when answering any question about what was taught, who's in the cohort, what dossiers exist, or what was said in a past session.
REPLACE THIS — one-line trigger summary that tells Claude when to load this skill. Write it like a tool docstring, not a title. Concrete trigger phrases beat abstract descriptions. Example "Triggers when the user asks about <specific recurring task in your domain>; expects <input shape>; produces <output shape>."
Walks a diff against a 5-point review checklist. Use when the user asks "review this diff", "what should I check in this PR", or pastes code with a question like "anything wrong with this?". Reads files via Glob/Grep/Read; never edits.
Walks a diff against a 5-point review checklist. Use when the user asks "review this diff", "what should I check in this PR", or pastes code with a question like "anything wrong with this?". Reads files via shell; never edits.
Q&A flow that produces a complete colleague-target.md and starter Constitution from a 5-field dossier. Use this when a cohort 4 student says they want help defining their AI coding colleague's target task, or when "dossier" / "colleague target" / "what should I build" comes up. Released as homework artifact at end of Lesson 1.
| name | cohort-search |
| description | Grounded retrieval over cohort 4 memory seeds (roster, lesson designs, dossiers, transcripts). Use when answering any question about what was taught, who's in the cohort, what dossiers exist, or what was said in a past session. |
Your default retrieval procedure. Use it before answering any cohort-bounded question.
Identify the question's keywords. Extract 2-4 specific terms (block names, dossier titles, student names, lesson topics).
Grep across /mnt/memory/cohort-knowledge/.
grep -ri --include="*.md" "<keyword>" /mnt/memory/cohort-knowledge/
Use -i for case-insensitive. Use -l first to find files; then read the matches.
Read the matching files in full (they're under 10 KB each). Do not paraphrase from grep snippets — the snippet often misses context.
Compose the answer, citing each claim by filename. Format:
<claim 1>. (see: /lessons/lesson-1-design.md)
<claim 2>. (see: /dossiers/01-bug-investigation.md)
If grep returns nothing: say "I don't have that in my seeds." Do not search the web. Do not infer from training data.
"What did Bayram say about X in lesson N?" → grep /transcripts/lesson-N-*.md. If no transcript yet, say so and check /lessons/lesson-N-design.md as the next-best source (the design is the plan, not the actual delivery).
"Who is X?" → check /team/cohort-roster.md first. If two students share a first name, ask the asker which one.
"What dossiers exist?" → list /dossiers/. Distinguish instructor-authored (01-, 02-) from student-authored (<name>-*).
"What's the verifier for X?" → grep "verifier" or "verification" in /dossiers/ and /lessons/. Quote the exact line.
"What's the closing horizon?" → check the last block of the relevant lesson design.