| name | prime |
| description | Load behavioral patterns and recent reflections from Obsidian to prime the current session with cross-project self-knowledge |
| allowed-tools | Read, Glob |
Prime — Session Context Loader
Load cross-project behavioral patterns from the shared Obsidian vault so this session benefits from accumulated self-knowledge. Run at the start of a conversation in any project.
Portable by design. Copy this skill to any repo's .claude/skills/prime/. It reads from a shared Obsidian vault, not from project-local files. The patterns it loads are about the user, not any specific codebase.
Skill family: /reflect analyzes recent sessions and writes Reflections/ (and promotes repeated signals into Patterns/) — this skill reads what /reflect writes. /reflect-me is the operator-profile sibling (Decision Mirror ledger → behavioral profile), a separate pipeline.
Constants
- Obsidian vault:
C:/Users/kazda/Documents/Obsidian/personas
- Patterns folder:
Patterns/ (promoted permanent rules — observed 3+ times)
- Reflections folder:
Reflections/ (periodic behavioral analysis from /reflect)
Step 0: Never fail the session start
This skill is best-effort. If the vault path doesn't exist (different machine, unsynced vault), or a folder is missing or empty, do NOT error or stop the conversation — proceed with whatever IS available and say plainly what was loaded and what wasn't (e.g. "vault not found at — running unprimed" or "2 patterns loaded; no reflections yet — run /reflect to create them"). One line per gap, then continue normally.
Step 1: Load Promoted Patterns
Glob C:/Users/kazda/Documents/Obsidian/personas/Patterns/*.md. Read each file.
These are permanent, validated behavioral rules — the highest-confidence signals about how the user works, promoted from reflections after appearing 3+ times. Pattern files are short; read them whole.
Step 2: Load Recent Reflections
Glob C:/Users/kazda/Documents/Obsidian/personas/Reflections/*.md. Sort by filename (dates sort lexicographically). Read the 3 most recent (or however many exist).
Excerpt, don't dump. Pull out only the load-bearing lines — do not recite whole files into the conversation:
- Current working phase (building / exploring / stabilizing)
- Recent correction patterns (what Claude keeps getting wrong)
- Validated approaches (what the user confirmed works)
- Emerging signals (not yet patterns)
If no reflections exist, note it per Step 0 and continue with patterns alone.
Step 3: Brief the Session
Print a concise primer so the user sees what the session was primed on. Start with a 3–5 line digest, max 10 lines total:
Session primed — {N} patterns, {M} reflections (latest: {date}).
Patterns:
• {pattern — one line each}
Recent signals:
• {latest correction or emerging signal — one line each}
Current phase: {from most recent reflection}
Rules for the briefing:
- One line per pattern/signal — verb-led, actionable
- Skip anything that's obvious from the codebase or CLAUDE.md
- If a pattern conflicts with the current project's CLAUDE.md, note the conflict but defer to CLAUDE.md (project-local rules win)
- Keep it under 200 words total; name any gaps (missing vault/folders) in one line
Step 4: Internalize
After presenting the primer, absorb the loaded patterns as soft behavioral defaults for the rest of this session:
- Apply correction patterns proactively — if a pattern says "user dislikes unnecessary error handling," avoid adding it before being asked
- Follow validated approaches when making judgment calls
- Match the user's observed thinking style (delegation level, exploration depth)
- Do NOT mention patterns again unless the user asks or a decision directly hinges on one
The patterns are defaults, not absolutes. If the user explicitly asks for something that contradicts a pattern, follow the user. Patterns capture past preferences; the current instruction always wins.
