Claude-Mind: Skill for Claude Code

/research-topic — Web Research Engine

Researches topics using WebSearch/WebFetch and writes findings directly to memory tree nodes. Called by /aspirations for knowledge goals, /reflect for gap-filling, and /replay for domain transfer.

Parameters

• <topic> (required) — Topic to research
• --depth quick|standard|deep — Research intensity (default: standard)
• --target-node <key> — Tree node key to update (auto-detected if omitted)

Step 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.

Step 1: Scope

Before any external research, determine what we already know and what gaps remain.

Bash: retrieve.sh --category {topic-category} --depth shallow
# Returns tree_nodes, experiences, reasoning_bank, guardrails, etc.
# Check if tree nodes already cover this topic (avoid redundant research)
# Check if experiences show prior research attempts

1. If target-node provided: use that node
   Else: node=$(bash core/scripts/tree-find-node.sh --text "{topic}" --leaf-only --top 1)
   # Returns: {key, score, file, depth, summary, node_type}
2. If matching node found (node.score > 0):
     - Read its .md file
     - Note what's already documented, identify gaps
     - Research focus = gaps only (never re-research known content)
3. If no matching node found:
     - Identify best parent for a new node (SPROUT)
     - Compute path: bash core/scripts/tree-read.sh --child-path <parent> <topic-slug>
     - Research focus = broad (building initial understanding)

Step 1.5: Encode-Stable-Facts Gate (G17)

Before external probing, check whether the topic names a resource whose locator is already encoded in world/conventions/. This prevents redundant web discovery for stable facts (paths, endpoints, account IDs) that prior sessions already found.

Bash: bash core/scripts/encode-stable-facts-gate.sh \
  --resource-id "{topic}" \
  --probe-count 0 \
  --threshold 3

# Exit 0 → proceed to Step 2 (resource not yet at threshold, or locator already exists).
# Exit 1 → the topic names a resource that has been probed ≥3 times without encoding.
#   Read the JSON output: if locator_found is true, read locator_file and use the
#   encoded value instead of re-discovering via web search. If locator_found is false,
#   proceed to Step 2 but encode any stable locator values discovered into
#   world/conventions/ before returning (per .claude/rules/encode-stable-facts.md).
#
# Note: probe-count starts at 0 here (first invocation for this topic this session).
# Callers that invoke research-topic multiple times for the same resource should
# increment probe-count across calls. The --threshold 3 default applies; research-topic
# may use a higher threshold via --threshold if broad exploration is documented.