Claude Desktop Research — Self-Contained Prompt Dispatch

Write deep-research prompts for Claude Desktop (macOS .app) — also covers Claude Web (claude.ai projects) as a subset. Both surfaces support Drive grounding and Deep Research mode. Only Claude Desktop has local MCP connections — and only in REGULAR CHAT, not in Research mode (empirically verified 2026-06-10, see below). The differentiator is Desktop's chat-mode MCP access.

🔑 DRIVE-WORKSPACE RULE (Etan, standing — never violate): Etan's research surface has his Google Drive workspace CONNECTED. Prompts must REFERENCE the relevant Google Doc(s) by NAME so the agent pulls them itself via the Drive connector. NEVER instruct Etan to attach/upload files — he should never have to attach anything, ever. Reference doc names, not attachments.

🔑 OUTPUT-PLACEMENT CONTRACT (Etan, standing — 2026-06-10): Every prompt this skill emits MUST end with a placement block stating where the output lives next: the owning-context folder (large-plan / design-doc / sprint — one folder per research, placement is dynamic), the filename convention, and whether Brain Drive archival applies. Claude-side outputs land on ~/Desktop or ~/Downloads (native exports: compass_artifact_<UID>*.md; Etan also hand-pastes .rtf/.rtfd) and MUST be deliberately moved to the owning folder — the moving agent verifies the destination is correct before moving (no blind automation for local files). The evidence of why we did what we did lives with the work it justified.

Why this skill is named "desktop"

Both Claude Web and Claude Desktop now expose:

• Deep Research mode (5–60 min web sweeps with citations)
• Google Drive grounding — Etan's Drive workspace is already connected. The prompt REFERENCES the relevant Doc(s) by NAME and the agent reads them itself via the connector. (Never an "attach a folder" step — see the DRIVE-WORKSPACE RULE above and the emit-gate below.)

But only Claude Desktop has local MCP connections (wired via claude_desktop_config.json + .mcpb extensions) — and they are exposed in REGULAR CHAT ONLY.

⚠️ VERIFIED 2026-06-10 (empirical, gen-15 weave): Claude Desktop Research mode CANNOT call local MCP tools. A Research-mode run instructed to call brain_search before any web work returned the mandated fallback "TOOL UNAVAILABLE IN RESEARCH MODE" and the UI logged Failed to call tool "brain_search" — Research routes everything through its own pipeline (launch_extended_search_task); local MCPs are unreachable mid-research. Any prompt that assumes otherwise hits a silent miss.

So the skill has two dispatch modes:

• Research mode — web sweeps + Drive connector grounding only. No local MCPs. Use for citation-heavy external research.
• Chat-bench mode — a benchmark-style prompt in a REGULAR Desktop chat: full local MCP access (brain_search, filesystem, voicelayer, …). This is the proven pattern for MCP-grounded deep work (2026-06-09 2nd-pass runs: context-hygiene, examplechannel-audio, m1-sync — all called brain_search + filesystem from chat).

If Gemini Desktop ever ships an MCP connections file, that gap closes — for now (2026-04-30) only Claude Desktop has it.

Pre-flight: Apply the CHECK-FIRST + GROUND + emit-only-if-pass gate before writing any paste-ready deep-research prompt.

When to use this vs. Gemini vs. quick lookup

Pre-flight: which MCPs are actually wired?

Before writing a prompt that assumes desktop-mode MCP access, verify the connection list. Claude Desktop wires MCP servers via TWO paths — both must be checked.

Path 1: Legacy claude_desktop_config.json

python3 -c "
import json
with open('$HOME/Library/Application Support/Claude/claude_desktop_config.json') as f:
    d = json.load(f)
servers = d.get('mcpServers', {})
print('Legacy MCPs (claude_desktop_config.json):')
for name in sorted(servers.keys()):
    print(f'  - {name}')
print(f'Total legacy: {len(servers)}')
"

Path 2: Extensions (.mcpb)

EXT_DIR="$HOME/Library/Application Support/Claude/Claude Extensions Settings"
echo "Installed .mcpb extensions:"
ls "$EXT_DIR"/*.json 2>/dev/null | xargs -n1 basename | sed 's/\.json$//' | sed 's/^/  - /'
[ -f "$EXT_DIR/extensions-installations.json" ] && python3 -c "
import json, sys
with open('$EXT_DIR/extensions-installations.json') as f:
    d = json.load(f)
print('extensions-installations.json:')
print(json.dumps(d, indent=2)[:1500])
"

The Extensions path is where modern MCPs live (e.g., local.mcpb.etan-heyman.brainlayer v1.0.0, installed 2026-03-28 — provides brain_search, brain_store, etc., but NOT visible in claude_desktop_config.json).

If a prompt instructs Claude Desktop to call a tool, the tool must appear in EITHER the legacy config OR the Extensions list. Always run BOTH commands above before writing a desktop-mode prompt — checking only one path can produce a false-negative ("MCP not wired" when it actually is, just under the other path).

.mcpb is the canonical wiring strategy going forward

The legacy claude_desktop_config.json is being phased out — entries currently in it (e.g., exa, whatsapp, whatsapp-business, voicelayer) will migrate to .mcpb extensions in a separate consolidation sprint. Until that sprint lands, BOTH paths must be checked. This skill should be revisited once the migration is complete to drop the legacy enumeration step.

Out-of-the-box behavior

If the relevant MCP is missing from BOTH paths, the prompt will hit a silent miss. Either (a) tell the user to install the missing .mcpb (preferred) or add to claude_desktop_config.json (legacy), then restart Claude Desktop; or (b) flag the gap and rewrite the prompt to web-mode (no MCP).

File Locations

Primary (Drive — works for BOTH desktop AND web)

Brain Drive/03_RESEARCH/Active/<project>/
  description.md
  instructions.md
  context/
    00-code-map.md
    01-*.txt
    40-*.txt
  prompts/
    R{NN}-{topic}.md
  results/
    R{NN}-claude-desktop-result.md     # was: R{NN}-claude-web-result.md (legacy alias retained)

Local cache (optional, not authoritative)

$ORCHESTRATOR_ROOT/docs.local/claude-desktop/projects/<project>/
  description.md
  instructions.md
  research-context/
  prompts/
  results/

Known projects

• batch-brainlayer → BrainLayer (memory layer, enrichment, search quality)
• batch-cmux → cmux (terminal multiplexer, agent orchestration)
• batch-orchestrator → Orchestrator (cross-repo coordination, agent design)

Workflow

Research Gates (mandatory)

1. Plan-author gate: When Claude Desktop/Web returns a generated research PLAN before execution, send that PLAN back to the prompt author and wait for an explicit START / EDIT verdict. Do not fire the research run just because the plan looks plausible.
2. Harvest own-brief gate: At harvest, any report claim that cites one of YOUR OWN briefs/prompts/plans must be cross-checked against the actual brief text before you accept, store, or relay it. Do not let a report cite a fabricated premise from your own setup. This gate exists because two 2026-06-05 catches were fabricated-premise failures: a hallucinated Drive doc in R01's plan and a nonexistent delta-brief mandate in R02's report.
3. Dispatch harvest contract (required — see /research-lifecycle): At dispatch time record harvest_owner, harvest_trigger, and export_route (e.g. $ORCHESTRATOR_ROOT/docs.local/research/R{NN}-result.md → brain_store). Applies to batch and one-off dispatches. Research without a harvest path = parked, not done. Export route must name how results return to the fleet.

Step 0 — Verify active account + MCP wiring

# Expects RESEARCH_ACCOUNT, else the RESEARCH_ACCOUNT= line in ~/.config/research-lifecycle/research.env.
bash skills/golem-powers/_shared/research/verify-account.sh

Then run the MCP pre-flight (above). If desktop-mode MCP is required and the relevant server is NOT wired, STOP and tell the user — do not write a prompt that will silently miss its tool.

Step 1 — Determine project and next RNN number

1. Ensure Drive folders exist:

   python3 skills/golem-powers/_shared/research/drive-paths.py ensure-project-folders <project>

2. List existing prompts from Drive:

   python3 skills/golem-powers/_shared/research/drive-paths.py list-prompts <project>

3. The next prompt gets R{NN+1}.

Step 2 — Write the research prompt

File: R{NN}-{topic-slug}.md

Template:

# R{NN}: {Title — What You're Researching}

## Mode

[desktop / web]   ← desktop if MCP needed, web if not

## Context

{2–3 paragraphs explaining WHY this research is needed.}

## Grounding (Drive workspace — CONNECTED; REQUIRED block)

{Ground this research in the following docs from my Google Drive — pull them
yourself via the connected Drive workspace; do NOT ask me to attach anything:
- "<Exact Google Doc name 1>"
- "<Exact Google Doc name 2>"
If no Drive doc applies, write "None — self-contained / web-only" so the
omission is deliberate. NEVER replace this block with an attach/upload ask.}

## Our Architecture

{Include INLINE — Claude Desktop in WEB MODE has zero codebase access; in DESKTOP MODE it can call MCP but still benefits from inline architecture for non-MCP-wired details. Include:
- Code snippets of relevant implementations
- Data shapes (SQL schemas, JSON structures)
- Exact numbers (chunk counts, tag counts, performance metrics)
- Tool/library versions}

## Available MCP tools (DESKTOP MODE ONLY — verify before relying)

{Skip this section for web-mode prompts. For desktop-mode, list the MCPs the prompt assumes:
- `mcp__brainlayer__brain_search` — semantic + keyword recall
- `mcp__voicelayer__*` — TTS / STT / pill-state
- `mcp__exa__web_search_exa` — web search
- ...}

## Research Questions

### 1. [Topic Area]
- Specific question 1
- Specific question 2

### 2. [Topic Area]
- ...

## Deliverable

Structured report with:
1. {Section 1 title} — {what it should contain}
2. ...

CRITICAL RULES for the prompt:

• SELF-CONTAINED for web mode: Zero codebase access. Every architecture detail / code snippet / number INLINE.
• MCP-AWARE for desktop mode: Name the exact MCP tool (mcp__brainlayer__brain_search) and what to do with the result. Don't say "use BrainLayer" — say "call brain_search('<query>') and report top 5 chunks."
• SPECIFIC: "How does X handle Y at Z scale?" not "Tell me about X."
• EXACT NUMBERS: "223K chunks", "22 domain tags", "p50=1480ms" — not "many chunks."
• DELIVERABLE SPEC: Tell Claude Desktop exactly what sections to return.
• REFERENCE VERSIONS: "Gemini 2.5 Flash-Lite", "sqlite-vec 0.1.6", "Next.js 16."
• GROUND BY NAME, NEVER BY ATTACHMENT: The ## Grounding block names the Drive Doc(s); Claude reads them via the connected workspace. The prompt must NEVER tell Etan to attach/upload/drag a file. (Gate-enforced — see Step 2.5.)

Step 2.5 — 🔒 EMIT-GATE: Drive-Workspace compliance (BLOCKING)

Before the prompt leaves this skill — and before the Step 5 user-facing message — run the mechanical gate. Do not eyeball it:

# Exit 1 = FAIL. Scan the prompt file you just wrote AND the dispatch message text.
bash skills/golem-powers/_shared/research/drive-grounding-gate.sh R{NN}-{topic}.md
# or pipe text:  printf '%s' "$MSG" | bash skills/golem-powers/_shared/research/drive-grounding-gate.sh -

• FAIL (gate finds attach / upload / drag / "add the file" aimed at Etan) → do NOT emit. Rewrite the offending lines to REFERENCE the Doc(s) by NAME in the ## Grounding block, re-run, repeat until PASS.
• PASS → safe to emit.

Step 3 — Update project context files in Drive

If the project's context has changed since the last prompt:

1. description.md — One paragraph. Current state, recent changes, what's live.
2. instructions.md — Full architecture reference. Tool status, DB state, constraints.
3. context/*.md — Add/update domain-specific files as needed.
4. Keep numbering aligned with skills/golem-powers/_shared/research/context-numbering.md.

Step 4 — Upload prompt + context to Drive (and optional local cache)

1. Upload R{NN}-{topic}.md to Brain Drive/03_RESEARCH/Active/<project>/prompts/ with mcp__google-drive__uploadFile(..., mimeType="text/markdown"). Set mimeType from the extension on every upload (.md → text/markdown, .txt → text/plain, .json → application/json): a prompt starting with <!-- is otherwise stored as text/html. createTextFile takes no mimeType; don't use it for these files.
2. Keep description.md, instructions.md, and context/ current in Brain Drive/03_RESEARCH/Active/<project>/context/
3. Optionally mirror to $ORCHESTRATOR_ROOT/docs.local/claude-desktop/projects/<project>/ for local review

Step 5 — Tell the user

Research prompt ready: R{NN}-{topic}.md
Mode: [desktop / web]   ← desktop if MCP needed, web if not
Primary: Brain Drive/03_RESEARCH/Active/<project>/prompts/
Local cache: docs.local/claude-desktop/projects/<project>/ (optional)

Open Claude Desktop and paste the prompt. Your Drive workspace is connected, so
Claude pulls the named Doc(s) itself — nothing to attach or upload.
[If desktop-mode] Confirm these MCPs are wired before submitting: <list>.
When Claude returns its research PLAN, send that PLAN back here for START/EDIT
before clicking Start Research.

Before sending the message above, it must PASS Step 2.5's gate too — a dispatch message that says "attach"/"upload" is the same violation as a prompt that does.

Step 6 — After results arrive

When the user downloads compass_artifact or copies the result:

1. Save to Brain Drive/03_RESEARCH/Active/<project>/results/R{NN}-claude-desktop-result.md
2. Optionally mirror to $ORCHESTRATOR_ROOT/docs.local/claude-desktop/projects/<project>/results/
3. brain_digest the raw content (extracts entities/relations)
4. brain_store conclusions with tags ["research", "conclusions", "{project}", "{topic}"]
5. Keep the Drive result file authoritative
6. Run the Harvest own-brief gate above for any report claim that cites your own prompt/brief/plan text
7. Apply findings to the codebase / sprint plan

Creating a new project batch

python3 skills/golem-powers/_shared/research/drive-paths.py ensure-project-folders newproject
# Then upload description.md, instructions.md, context files, and the first prompt.

Migrating legacy Obsidian batches

bash skills/golem-powers/claude-desktop-research/scripts/migrate-obsidian-to-drive.sh batch-cmux --dry-run
bash skills/golem-powers/claude-desktop-research/scripts/migrate-obsidian-to-drive.sh batch-cmux

Procedure: skills/golem-powers/claude-desktop-research/workflows/migrate.md

Quality checklist

Before telling the user the prompt is ready:

• drive-grounding-gate.sh PASSES on the prompt AND the dispatch message (zero attach/upload/drag — Step 2.5)
• ## Grounding block present, naming the Drive Doc(s) by exact name (or "None — self-contained")
• Mode declared (desktop / web) — both prompt header AND user-facing message
• If desktop mode: relevant MCPs verified in claude_desktop_config.json
• Prompt is fully self-contained (no "see the codebase" references for web-mode prompts)
• Architecture includes code snippets, not just descriptions
• All numbers are exact (not "many", "large", "several")
• Research questions are specific (not "tell me about X")
• Deliverable section specifies exact output structure
• Library/tool versions included
• Prompt author will review the generated research PLAN before execution
• Harvest includes own-brief cross-check for any self-cited setup claim
• Context files are up to date with current state
• Prompt uploaded to Drive prompts/
• Context lives in Drive context/
• Optional local cache updated if needed
• Next RNN number is correct (no collisions)

Renamed from /claude-web-research (2026-04-30)

User correction: BOTH Claude Web AND Claude Desktop have deep research + Drive grounding. The differentiator is MCP — only Desktop has local connections. The skill formerly known as /claude-web-research is now /claude-desktop-research (desktop is the umbrella; web is the MCP-less subset).

The legacy /claude-web-research slash trigger retired on 2026-05-30. Use /claude-desktop-research for Claude Desktop/Web research prompts.