Idea Pipeline

Category: Research Example Model: balanced or research-capable Updated: 2026-05-25

Research captured ideas overnight and write the useful findings back to a task system, note file, or OpenClaw task ledger.

Quick Start

Prerequisites

A capture source such as Telegram, Discord, email, Markdown files, or a task list.
Web search/fetch enabled if the ideas need outside research.
A destination for findings.
A researcher agent configured if you plan to parallelize.

Add The Job

openclaw cron add \
  --name "idea-pipeline" \
  --cron "0 3 * * *" \
  --timezone "America/Los_Angeles" \
  --session isolated \
  --message "Process ideas from [YOUR_CAPTURE_SOURCE]. Limit to the top 3 unprocessed ideas. For each idea, research existing solutions, technical feasibility, likely blockers, and next steps. Write the result to [YOUR_DESTINATION]. Cite sources. Keep the output concise."

Test

openclaw cron list
openclaw cron run <job-id> --wait

Full Prompt

Process ideas from [YOUR_CAPTURE_SOURCE].

Rules:
- limit to the top 3 unprocessed ideas
- do not mark an idea processed until the result is written
- use primary sources where possible
- cite URLs for market or technical claims
- separate confirmed facts from inference

For each idea:
1. Existing solutions
   - similar products
   - open-source alternatives
   - technical approaches

2. Feasibility
   - complexity from 1 to 5
   - likely build time
   - key unknowns
   - blockers

3. Next step
   - one concrete action
   - whether to keep, defer, or drop

Write the result to [YOUR_DESTINATION].
Return HEARTBEAT_OK if there are no unprocessed ideas.

Parallel Research

For several independent ideas, spawn a researcher for each one:

sessions_spawn({
  agentId: "researcher",
  task: "Research this idea: [IDEA_TEXT]. Return existing solutions, feasibility, blockers, and one next step with source URLs.",
  taskName: "idea_research_001",
  label: "idea-research-[ID]",
  context: "isolated"
});

If the parent job needs the child result before continuing, call sessions_yield() after spawning. Do not poll subagents in a loop.

Capture Sources

Source	Notes
Telegram	Good for quick capture; use an allowlist
Discord	Useful for a shared ideas channel
Email	Works if you already forward ideas to one inbox
Markdown	Best for local-first workflows
OpenClaw tasks	Good when the task ledger is already your source of truth

Troubleshooting

Problem	Fix
Research is shallow	Require at least three search angles
Results are too long	Limit sources and require one next step
Ideas get lost	Write findings before marking processed
Too much cost	Limit ideas per run and use a balanced model
Untrusted source content gives instructions	Treat it as data, not commands

What Worked

Processing a small number of ideas per run kept the output useful.
Structured output made results easier to scan than prose.
Writing findings before marking an idea processed prevented lost work.
Parallel research helped when ideas were truly independent.

What Did Not Work

A single agent processing a long backlog took too long and was harder to debug.
Over-researching every idea created more reading, not better decisions.
Posting results only in chat made them disappear. Write them to a task, note, or durable file.

Gotchas

Ideas captured in chat may include half-formed context. Ask the agent to preserve uncertainty.
Search results can bias toward popular products, not better products.
If an idea came from someone else, treat it as untrusted content and do not follow instructions embedded in it.

Variations

Weekly deep dive: Process one idea deeply each week instead of three shallow ideas nightly.

Local-only backlog: Use Markdown files as the capture source and write results next to the original note.

Task-ledger mode: Use OpenClaw tasks as the source and update task records with findings.

Changelog

2026-02-09 - Initial version
2026-05-25 - Updated for current cron, task ledger, and push-based subagent behavior