Claude-Mind: Skill for Claude Code

/prime — Context Priming Engine

Loads the agent's accumulated knowledge into active context so that conversations and goal execution start with domain awareness rather than amnesia.

Internal skill: called by boot (RUNNING state) and session start protocol (any mode). Not user-invocable — users enter persona to prime automatically.

Key design: Boot loads the MAP (indexes, summaries). Prime loads the TERRITORY (actual tree node content, reasoning bank entries, guardrail details). Together they give the agent full domain awareness.

Sub-commands

/prime                    — Auto-detect context and prime broadly
/prime --category <cat>   — Prime a single category at deep depth

Phase 0: Load Conventions

Step 0: Load Conventions — Bash: load-conventions.sh with each name from the conventions: front matter. Read only the paths returned (files not yet in context). If output is empty, all conventions already loaded — proceed to next step.

Phase 0.5: Agent Mode Detection

# Capture the state check's output FIRST so the IF clause below
# unambiguously refers to it (G3, 2026-05-20). The validate-paths.sh
# call between this and the IF emits multi-line "L3 PATH VALIDATOR"
# text — without naming the variable, a future editor could misread
# the IF as checking validate-paths' output.
Bash: state=$(bash core/scripts/session-state-get.sh); echo "state=$state"
# The trailing echo (H1, 2026-05-20) makes the captured value visible to the
# LLM in the Bash tool's stdout. Without it, the assignment writes to a bash
# subshell variable that dies when bash exits — zero stdout — and the IF
# clause below has no value to compare. Sibling skills use bare `var=$(cmd)`
# as pseudocode convention; /prime's NO_AGENT branch is critical enough to
# warrant explicit echo.

# L3 PATH DEFENSE (g-115-35) — verify local-paths.conf points at paths that
# exist and are writable BEFORE any tree/meta read or write. Fail-open:
# prints WARN on mismatch but never blocks loop entry. Enforcement belongs
# to L1 (write-time hook) and L2 (permission gate).
Bash: validate-paths.sh

# LANE-PIN REVIEW SURFACE (g-115-5901) — sibling of the L3 defense above: a
# fail-open startup advisory that prints and never blocks. A lane pin is a
# durable USER-DIRECTED restriction on this agent's entire work surface; its
# `review_by` column says when a human should LOOK at it again, not when it
# stops binding. Past that date this prints "still enforced — confirm or
# retire" and keeps printing every startup until a human edits the registry row.
#
# The pin is NOT weakened by a lapsed date and the claim gate never consults
# this — a hard expiry would silently hand the agent back a surface the user
# deliberately took away. Before this existed the `expires` column was parsed
# into a field NO caller anywhere in the tree ever read, so a pin could sit
# unexamined forever and the dead field terminated the audit that would have
# found it. Silent when no pin covers this agent, or when none is past due.
Bash: lane-pin-review.sh

IF state == "NO_AGENT":
  → World-only priming mode. Skip all agent-specific steps.
  → Bash: world-cat.sh program.md  # The Program — shared purpose
  → Bash: world-cat.sh knowledge/tree/_tree.yaml  # collective knowledge overview
  → Bash: guardrail-manifest.sh (shared safety rules — id manifest, 100%
    coverage; same budget + expansion rules as Phase 2 item 3)
  → Bash: reasoning-bank-read.sh --recent (shared lessons — bounded; see Phase 2 item 4)
  → Display:
    ═══ WORLD PRIME (no agent) ═══
    PROGRAM: [contents of world/program.md]
    KNOWLEDGE: [tree summary]
    GUARDRAILS: [count] active
    REASONING: [count] active
    ════════════════════════════════
  → Output: "Primed in world-only mode. No agent identity active."
  → DONE (skip all remaining phases)