parallel-research（平行研究）

DEPRECATED -- Use /deep-research instead

This skill uses claude -p for parallel research agents, which resulted

in an account ban. The replacement /deep-research skill uses Codex

(codex exec --full-auto) for web research and Claude as in-context

orchestrator -- zero ban risk.

/parallel-research -- Parallel Web Research via tmux + claude -p

Decompose a research question into MECE domains, spawn parallel Claude agents in tmux panes (up to 12), each running claude -p with pre-approved web tools. Collect findings into structured output files.

Argument

A research topic or question to investigate broadly. Example: "state of agentic coding patterns Dec 2025 - Feb 2026"

Procedure

Phase 1 -- Understand

Clarify the research question. Identify:

• The core topic
• The timeframe
• How many dimensions / angles the user wants covered
• Whether the user wants adversarial / contrarian coverage

If unclear, ask the user before proceeding.

Phase 2 -- Decompose into MECE Domains

Split the research into 3-6 independent domains. For each domain, create 2-3 search variants (a/b/c) that approach the same domain from different angles. This multiplies coverage without redundancy.

Decomposition rules:

• Divide by research domain, NOT by source type. Each agent owns a vertical slice of a topic (e.g., "enterprise adoption", "failure stories"), not a horizontal layer (e.g., "find all tweets", "find all blogs"). Horizontal decomposition causes massive redundancy.
• Each variant within a domain should use different search terms and angles to maximize coverage.
• All domains together should be collectively exhaustive -- no gaps.

Present the decomposition table and wait for user approval:

| Domain | Agent A (Variant) | Agent B (Variant) | Model |
|--------|-------------------|-------------------|-------|
| Voices | Social media power users | Industry leaders & engineers | sonnet |
| Blogs  | Workflow pattern posts | Code quality problem posts | sonnet |
| Critics | Failure postmortems | Skeptics & contrarians | opus |

The Model column controls which Claude model each domain's agents use. Default is claude-sonnet-4-6. Use opus for domains that require deep analysis or nuanced reasoning; use haiku for simple URL-scraping or aggregation agents. This maps to the second argument of run-agent.sh.

Do NOT proceed until the user approves this decomposition.

Phase 3 -- Create Agent Directories and Prompts

Create the directory structure:

mkdir -p research-agents/{agent-name-a,agent-name-b,...}
mkdir -p research-agents/outputs

For each agent, write research-agents/{agent-name}/prompt.md with this structure:

You are a research agent. DO NOT modify any files in any codebase.
Your ONLY job is to search the web and compile findings.

# Mission: [Specific research goal for this agent]

## Search Strategy
[10+ specific search queries, tailored to this agent's angle]

## What to Capture
[Structured template: what fields to record per finding]

## Output Format
[How to organize the output -- categories, ranking, etc.]

Prompt design rules:

• Always start with "DO NOT modify any files" -- prevents agents from touching code
• 10+ search queries per agent -- ensures breadth
• Structured output format -- makes aggregation possible later
• Variance between a/b agents -- different search terms for the same domain to catch what the other misses

Phase 4 -- Create Launcher Script

Write research-agents/run-agent.sh:

#!/bin/bash
# Usage: bash run-agent.sh /path/to/agent-dir [model]
# Pipes the agent prompt to claude -p with pre-approved web tools.
# Writes output to output.md and prints a COMPLETE marker when done.
cd "$1"
MODEL="${2:-claude-sonnet-4-6}"
echo "=== Starting research agent: $(basename $1) with model $MODEL ==="
echo "=== $(date) ==="
echo ""
cat prompt.md | claude -p \
  --model "$MODEL" \
  --allowedTools "WebSearch,WebFetch,Bash,Read,Write,Grep,Glob" \
  2>&1 | tee output.md
echo ""
echo "=== Agent $(basename $1) COMPLETE at $(date) ==="

Make it executable: chmod +x research-agents/run-agent.sh

CRITICAL details in this script:

• Pipe prompt via stdin (cat prompt.md | claude -p). NEVER pass the prompt as a positional argument after --allowedTools -- the variadic flag will consume it as a tool name.
• Comma-separated tools ("WebSearch,WebFetch,..."). NEVER use spaces -- they cause the same variadic consumption problem.
• --allowedTools is mandatory. claude -p runs non-interactively and cannot prompt for permission approval. Without this flag, every WebSearch call gets denied silently and the agent exits with "I'm blocked".
• tee output.md streams to both the terminal and a file.

Phase 5 -- Build tmux Session

Calculate layout: ceil(total_agents / 4) windows, 4 panes each.

Create the session one window at a time. Validate each window before creating the next.

# Create first window
tmux new-session -d -s research -n "{window-name}" -x 220 -y 55
tmux split-window -t research:{window-name}
tmux split-window -t research:{window-name}
tmux split-window -t research:{window-name}
tmux select-layout -t research:{window-name} tiled

# VALIDATE before continuing
tmux list-windows -t research -F "#{window_name}: #{window_panes} panes"

Repeat for additional windows with tmux new-window.

CRITICAL: Always validate pane count after creating each window. Never proceed to launching agents without confirming the panes exist. Sending tmux send-keys to non-existent panes can crash the tmux server -- killing ALL tmux sessions on the machine, not just yours.

Phase 6 -- Launch Agents

Send the launcher command to each pane, one window at a time:

BASE="$(pwd)/research-agents"
L="$BASE/run-agent.sh"

# Pass model as second arg (from decomposition table's Model column)
tmux send-keys -t research:{window}.0 "bash $L $BASE/{agent-a} claude-sonnet-4-6" Enter
tmux send-keys -t research:{window}.1 "bash $L $BASE/{agent-b} claude-sonnet-4-6" Enter
tmux send-keys -t research:{window}.2 "bash $L $BASE/{agent-c} claude-opus-4-6" Enter
tmux send-keys -t research:{window}.3 "bash $L $BASE/{agent-d} claude-sonnet-4-6" Enter

Match the model argument to the Model column from the approved decomposition table. Omitting it defaults to claude-sonnet-4-6.

After launching all agents, verify the processes are running:

ps aux | grep "claude -p" | grep -v grep | wc -l

The count should match the number of agents launched.

Print monitoring instructions for the user:

## Parallel Research Agents Running

Attach to session:     tmux attach -t research
Switch windows:        Ctrl-b n (next) / Ctrl-b p (previous)
Switch panes:          Ctrl-b + arrow keys
Zoom a pane:           Ctrl-b z (toggle)
Kill session:          tmux kill-session -t research

Check progress:
  ls -la research-agents/*/output.md

Phase 7 -- Monitor and Collect

Automatic Completion Detection

Use the wait-for-text.sh helper from the /tmux skill to poll each pane for the COMPLETE marker that run-agent.sh prints on exit.

Build a polling loop that checks all panes:

WAIT="$HOME/Developer/personal_projects/agent-skills/skills/tmux/scripts/wait-for-text.sh"
TIMEOUT=600  # 10 minutes per agent; adjust as needed

all_done=true
for pane_id in {list of session:window.pane targets}; do
  agent_name="{name for this pane}"
  if bash "$WAIT" -t "$pane_id" -p "COMPLETE" -T "$TIMEOUT" -i 5; then
    echo "$agent_name: DONE"
  else
    echo "$agent_name: TIMED OUT or FAILED -- check pane manually"
    all_done=false
  fi
done

if $all_done; then
  echo "All agents complete."
else
  echo "Some agents did not finish. Check timed-out panes."
fi

Run this loop SEQUENTIALLY (one wait-for-text.sh call per agent) so that each agent gets the full timeout. The loop will block until each agent finishes or times out.

Manual Progress Check

If you prefer a quick status check instead of blocking:

for agent in {list-of-agents}; do
  dir="research-agents/$agent"
  if [ -s "$dir/output.md" ]; then
    bytes=$(wc -c < "$dir/output.md")
    echo "$agent: ${bytes} bytes"
  else
    echo "$agent: still running..."
  fi
done

Note: tee buffers lazily. Empty output files for 30-60 seconds is normal while agents are doing web searches. Verify processes are alive with ps aux | grep "claude -p" rather than relying on file sizes.

When all agents have completed, report the final output sizes and proceed to synthesis.

Phase 8 -- Synthesize

Read every research-agents/*/output.md file. Cross-reference the findings and produce research-agents/outputs/synthesis.md with this structure:

# Research Synthesis: {topic}

## Executive Summary
[3-5 sentences capturing the most important findings across all agents]

## Key Findings by Domain
### {Domain 1}
- [Bullet points synthesizing findings from agents in this domain]
- [Include source attribution: "Agent {name} found..."]

### {Domain 2}
...

## Cross-Domain Patterns
[Themes, trends, or insights that appear across multiple domains.
These are the highest-value findings -- they show convergence.]

## Contradictions and Tensions
[Where agents found conflicting information or opposing viewpoints.
Note which sources support each side.]

## Gaps and Open Questions
[What the research did NOT cover well. Topics that need deeper
investigation. Questions raised by the findings.]

## Raw Source Index
| Agent | Domain | Output File | Key Sources |
|-------|--------|-------------|-------------|
| {name} | {domain} | {path} | [list of URLs/papers cited] |

Rules for synthesis:

• Cross-reference aggressively. The value is in connections between domains, not summaries of individual agents.
• Preserve attribution. Every claim traces back to a specific agent and source.
• Flag contradictions explicitly. Don't silently pick a side.
• Keep the raw outputs. Synthesis is additive, not a replacement.

Present the synthesis for user review before finalizing.

Troubleshooting

Check agent status Run the health-check recipe from /tmux to classify every pane:

DIAG="$HOME/Developer/personal_projects/agent-skills/skills/tmux/scripts/diagnose-agents.sh"
bash "$DIAG" research "research-agents/*/output.md"

Each pane is classified as DONE, WORKING, FAILED, or UNKNOWN. See the /tmux skill's "Health-Check Running Agents" recipe for details on what each state means and how it's detected.

FAILED state -- what to do Inspect the pane for the root cause:

tmux capture-pane -t research:{window}.{pane} -p -S - | tail -30

Common causes: rate limit hit, permission denial (missing --allowedTools), traceback, or OOM. Fix the cause and re-launch the agent (see below).

DONE but no output.md This is a tee buffering issue. The agent finished (COMPLETE marker present in the pane) but the file wasn't flushed. Capture the pane scrollback directly:

tmux capture-pane -t research:{window}.{pane} -p -S - > research-agents/{agent}/output.md

Re-running a single failed agent Send the launcher command to the same pane (or a new one):

tmux send-keys -t research:{window}.{pane} "bash $L $BASE/{agent-name} {model}" Enter

The previous output.md will be overwritten by tee.

tmux session died Re-create the session and windows (Phase 5). Run diagnose-agents.sh or check which agents have output.md with content. Only re-launch agents that didn't complete.

Rate limits across many agents If multiple agents hit rate limits simultaneously, reduce the number of concurrent agents or stagger launches with sleep 30 between send-keys commands.

Important

• Always decompose by research domain, never by source type.
• Always wait for user approval of the decomposition before creating directories.
• NEVER pass the prompt as a positional argument to claude -p after --allowedTools. Always pipe via stdin.
• NEVER use space-separated tool names in --allowedTools. Always use commas.
• ALWAYS validate tmux panes exist before sending commands to them. Validate after each window creation, not at the end.
• ALWAYS launch one window of agents at a time. Verify that window's panes are alive before moving to the next window.
• The tmux session name is research. Do not reuse a session name that already exists -- check with tmux has-session -t research first and pick a different name if it exists.
• Do NOT modify any files in the codebase. This skill is research-only.