Vault Scout

A Haiku-backed primitive for mechanical Obsidian-vault exploration. The main agent calls this skill when it needs a structured answer to a search/listing/frontmatter question without polluting its own context with dozens of file reads.

Why this skill exists

Vault work frequently needs the same kind of mechanical sweep:

• "Find every NPC linked to a given faction"
• "List all notes under Sessions/Session N/ and their frontmatter types"
• "Grep for [[Note Name]] across Locations/ and show which notes reference it"
• "Read these 12 location notes and tell me which ones have no discovery: field"

These tasks are pattern-match + read + tabulate. No creative judgment, no prose generation. Letting the main Opus agent run them is wasteful: the main agent pays Opus prices for Haiku-grade work, and its context fills up with file contents it doesn't need beyond the tabulation.

vault-scout dispatches a Haiku subagent with the narrow question, lets Haiku do the scanning in its own context, and receives back a clean tabular or bulleted answer.

When to use

• The task is find X / list Y / read frontmatter and tabulate / grep for pattern Z.
• The answer is a table, bulleted list, or short structured summary — not prose, not creative content.
• The work involves reading many files (≥5) or running pattern searches that would otherwise pollute main-agent context.
• The rules for the task are fully determined by the request (no judgment calls beyond "does this match?").

When NOT to use

• Creative work — NPC psychology, location design, scene architecture, prose writing. Main Opus owns these.
• Single-file reads — one file, small, go direct. Dispatch overhead (~200 Opus tokens to write the prompt + ~80 Opus tokens to read the result) exceeds the work.
• Mentions or sections of one note: vault_mentions, vault_read_note and vault_read_section return just those parts in the main context, so there is nothing to offload.
• Simple Bash ops — file renames, git mv, mkdir. Main agent Bash is faster and cheaper.
• Synthesis across sources — if the answer requires combining vault content with sourcebook content, that's a worldbuilding task. Use worldbuilding or related skills.
• Tasks needing taste/judgment — "which NPC would work best for Session N?" is creative. Main Opus.

Dispatch contract

The main agent invokes a Haiku subagent like this:

Agent(
  subagent_type="general-purpose",
  model="haiku",
  description="<short task description>",
  prompt="<self-contained task with explicit tool allowlist and output format>"
)

The dispatch prompt MUST:

1. State the task in one sentence.
2. List the paths or glob patterns to consider.
3. Name the tools the subagent may use: Glob, Grep, Read, mcp__vault-graph__vault_search_notes, mcp__vault-graph__vault_backlinks, mcp__vault-graph__vault_graph, mcp__vault-graph__vault_central_notes, mcp__vault-graph__vault_mentions, mcp__vault-graph__vault_read_note, mcp__vault-graph__vault_read_section.
4. Specify the output format (table columns, bullet structure, or short summary).
5. Cap verbosity: "Report under 200 words unless the listing itself is longer."

The dispatch prompt MUST NOT:

• Ask for interpretation, recommendations, or creative output.
• Include "also add..." side-tasks unrelated to the scan.
• Leave the output format open-ended ("describe what you find").
• Ask Haiku to determine status, canonicity, validity, or cross-reference sources — that's synthesis work for the main agent.

Anti-hallucination contract

Haiku confidently invents plausible details when source text is ambiguous. Every dispatch prompt MUST end with this suffix:

IMPORTANT — FETCH ONLY, DO NOT INTERPRET:
- Extract exact quotes from source files. Do not paraphrase.
- If a field is not explicitly stated, write "NOT STATED" — do not infer.
- Include file path and line number for every extracted value.
- Do not determine status, canonicity, or cross-reference multiple sources.
- The calling agent will interpret; you only locate and quote.

Why this matters: The main agent (Opus) owns judgment. Haiku's job is to surface raw material with provenance. When Haiku invents, the main agent has no way to verify — it trusts the subagent's "findings" and propagates hallucinations to the user.

Tool preferences (inside the subagent)

Prefer the MCP graph tools when they fit — they are faster than Glob+Grep for many questions:

• mcp__vault-graph__vault_search_notes(name_pattern="...", tag="...", note_type="...") — find notes by name, tag, or type: frontmatter. Fast, indexed.
• mcp__vault-graph__vault_backlinks(note="Name") — what links to this note. Pre-computed.
• mcp__vault-graph__vault_graph(note="Name", depth=2) — local neighborhood.
• mcp__vault-graph__vault_central_notes(tag="...") — most-connected nodes in a category. Useful for "what are the important X?".
• mcp__vault-graph__vault_mentions(note="Name"): the block around each link to the note, labelled path:line, plus a map of mentions not shown; unlinked=true for plain-text mentions. Reads the vault live.
• mcp__vault-graph__vault_read_note(note="Name"): a small note whole, a large one as an outline with line ranges.
• mcp__vault-graph__vault_read_section(note="Name", headings=["..."]): chosen sections in full.

Fall back to Glob + Grep + Read when:

• The search is text-pattern-based (e.g., "find all notes mentioning a keyword in prose").
• The graph index may be stale (recent moves/adds not yet reloaded). This applies to the graph tools only; the three tools above read the vault live.
• The task needs file-content inspection beyond the graph.

Output contract

Returns one of three shapes to the main agent:

1. Raw extraction table (preferred for multi-field queries):

   | File | Line | Field | Verbatim Quote |
   |---|---|---|---|
   | /full/path/Note.md | 42 | destination | "leads to the eastern marches" |
   | /full/path/Note.md | 45 | key | NOT STATED |

2. Bulleted list with provenance — when the task is "find all matching notes":

   - /full/path/note.md:23 — "exact quote from line 23"
   - /full/path/other.md:7 — "exact quote from line 7"

3. Short summary — when the task is a count, a boolean, or a short answer. Still cite file:line for any claim.

Key rules:

• Always absolute paths, so the main agent can Read without path juggling.
• Always include line numbers — enables verification.
• Verbatim quotes in double-quotes; NOT STATED (no quotes) when absent.
• Never summarize or paraphrase content — quote it.

Examples

Example 1 — Find NPCs by faction

Main agent invocation:

Agent(
  subagent_type="general-purpose",
  model="haiku",
  description="Find NPCs linked to [Faction Name]",
  prompt="""
  Find all NPC notes in <vault-root>/Grunvyr_Campaign/NPC/
  whose frontmatter `faction:` field contains '[Faction Name]'.

  Tools: Glob, Grep, Read.
  Output as: | File | Line | Field | Verbatim Quote |
  Extract: filename, faction: value, alignment: value (verbatim from frontmatter).

  IMPORTANT — FETCH ONLY, DO NOT INTERPRET:
  - Extract exact quotes from source files. Do not paraphrase.
  - If a field is not explicitly stated, write "NOT STATED" — do not infer.
  - Include file path and line number for every extracted value.
  - Do not determine status, canonicity, or cross-reference multiple sources.
  - The calling agent will interpret; you only locate and quote.
  """
)

Example 2 — List session scene files

Main agent invocation:

Agent(
  subagent_type="general-purpose",
  model="haiku",
  description="List Session N scene files",
  prompt="""
  List all .md files under <vault-root>/Grunvyr_Campaign/Sessions/Session N/,
  including subfolders.

  For each file, read the frontmatter and report: Path, type: field value, date: field value.

  Tools: Glob, Read.
  Output as a bulleted list, one line per file.
  """
)

Example 3 — Mentions via graph MCP

Main agent invocation:

Agent(
  subagent_type="general-purpose",
  model="haiku",
  description="Mentions of [[NPC Name]]",
  prompt="""
  Use mcp__vault-graph__vault_mentions(note="NPC Name") to get every note that links to NPC Name.md, with the block around each link.

  For each mention, report: path:line, and the sentence that mentions NPC Name, quoted from the excerpt. If the response ends with a map of mentions not shown, call the tool again with paths=[...] for those notes.

  Tools: mcp__vault-graph__vault_mentions.
  Output as a table.
  """
)

BAD vs GOOD prompt patterns

BAD — asks Haiku to interpret:

Find all portals in the vault. For each portal, determine:
- Status: discovered by PCs vs. planned/GM-only
- Whether the key has been found
- If the portal is still active

↑ "determine," "whether," "if" — these invite synthesis. Haiku will invent confident answers.

GOOD — asks Haiku to fetch with provenance:

Find all notes mentioning "Portal" or "Gate" in Locations/ and Sessions/.
For each match, extract verbatim:
- Any line containing "Destination" or "leads to"
- Any line containing "Key"

Tools: Grep, Read.
Output as: | File | Line | Field | Verbatim Quote |

IMPORTANT — FETCH ONLY, DO NOT INTERPRET:
- Extract exact quotes from source files. Do not paraphrase.
- If a field is not explicitly stated, write "NOT STATED" — do not infer.
- Include file path and line number for every extracted value.
- Do not determine status, canonicity, or cross-reference multiple sources.
- The calling agent will interpret; you only locate and quote.

↑ Main agent (Opus) receives raw quotes with line numbers, then cross-references PlayerNotes itself to determine status.

Failure modes to avoid

1. Don't dispatch for <5 file reads. The Opus-priced dispatch prompt costs more than the work.
2. Don't dispatch creative tasks. A Haiku subagent asked for "who would be a good NPC for this plot" will produce Haiku-grade NPCs. That's a loss.
3. Don't dispatch synthesis across vault + sourcebooks. That's worldbuilding territory; Haiku's cross-source reasoning is weaker.
4. Don't dispatch tasks where the main agent needs to reason about every file individually. If the main agent will Read the same files anyway to make a decision, the dispatch is pure overhead — the files will be in the main context twice.

If the main agent catches itself dispatching many sequential vault-scout calls to drill down (scan → read → scan → read), consolidate into one dispatch with a more specific prompt. Haiku is good at following narrow instructions; give it the full search-tree in one go.