name: com.example/research-desk
version: 2.1.0
title: Research desk
description: >-
Researches a market question in parallel, cross-checks sources and writes a cited report.
icon: assets/icon.svg
authors:
- name: Jane Doe
email: jane@example.com
url: https://example.com/jane
role: maintainer
- name: Example Labs
url: https://example.com
links:
homepage: https://example.com/teams/research-desk
repository: https://github.com/example/teams
issues: https://github.com/example/teams/issues
docs: https://example.com/docs/research-desk
license: Apache-2.0
when:
- research a market question
- write a cited competitive report
derived_from:
name: com.example/research-desk
version: 2.0.3
project: required
requires:
harnesses: [claude, codex]
inputs:
question: The market question to research
depth:
description: How deep to go
required: false
default: standard
choices: [quick, standard, deep]
permissions:
network: [api.example.com, "*.wikipedia.org"]
repo_write: true
shell: true
browser: false
secrets:
- name: SEARCH_API_KEY
purpose: Web search for the researcher role
required: true
- name: SLACK_WEBHOOK
purpose: Optional digest to a channel
required: false
files:
- path: assets/icon.svg
description: Package icon
content: |
- path: essences/how-we-research.md
description: Our research method; read before planning
content: |
# How we research
Primary sources first. Every claim gets a link and a date.
- path: templates/report.md
description: Shape of the final report
content: |
# {Question}
## Findings
## Sources
- path: data/competitors.csv
description: Known competitors with segments
content: |
name,segment
Example Corp,enterprise
Sample Systems,mid-market
Demo Analytics,small business
- path: assets/logo.svg
description: Logo for the report cover
content: |
memory:
- title: Prefer filings over press releases
kind: rule
body: When a filing and a press release disagree, trust the filing and note the conflict.
tags: [sources]
- title: Market sizes older than two years mislead
kind: learning
body: Past reports cited stale sizes; always state the year and flag figures older than two years.
skills:
- name: source-check
description: Verify a claim against its cited source.
content: |
---
name: source-check
description: Verify a claim against its cited source.
---
1. Open the source. 2. Quote the exact sentence. 3. Mark verified, partial or wrong.
files:
- path: scripts/fetch.py
content: |
import sys, urllib.request
print(urllib.request.urlopen(sys.argv[1]).read()[:2000])
- name: market-sizing
description: Estimate a market size from cited figures.
content: |
---
name: market-sizing
description: Estimate a market size from cited figures.
---
1. Collect sizes from primary sources and record each figure's year.
2. Flag any figure older than two years.
3. Give a range, not a point, and cite every input.
automation:
- name: digest
description: Posts a progress digest every hour while research runs
content: |
trigger: { every: 1h, while: research_running }
action: { notify: orchestrator, digest: progress }
- name: stale-task-nudge
description: Nudges the lead when a task has been idle for two hours
content: |
trigger: { idle_task: 2h }
action: { notify: orchestrator }
managers:
orchestrator:
name: Lead
start: with_team
model: [claude:opus, codex:gpt-6.1-sol]
reasoning: high
skills: [source-check]
tools: [mcp:io.github.example/search]
files: [essences/how-we-research.md, templates/report.md]
instructions: |
How this team works: you frame the question with the owner, Atlas splits
it into one task per angle, researchers work in parallel and the checker
verifies every claim against its source.
Agree the question with the owner before planning. Merge findings into
templates/report.md and cite every claim.
Never publish outside the project without the owner's approval.
planner:
name: Atlas
start: with_team
instructions: |
Split the question into one task per angle.
Start with the three biggest competitors, then widen.
cartographer:
start: on_demand
model: [claude:sonnet]
curator:
start:
schedule: 45m..2h
model:
same_as: orchestrator
workers:
researcher:
description: Researches one angle of the question.
model: [claude:sonnet, codex:gpt-6.1-sol]
reasoning: medium
skills: [market-sizing]
tools: [mcp:io.github.example/search]
files: [data/competitors.csv]
instructions: |
Research exactly one angle. Every claim gets a source link and a date.
checker:
description: Verifies every claim in a researcher's report.
model: [codex:gpt-6.1-sol]
reasoning: high
skills: [source-check]
instructions: |
Verify every claim in one researcher's report against its cited source.
Use the source-check skill: open the source, quote the exact sentence and mark
each claim verified, partial or wrong.
Return the claims with their marks and list every claim without a source.
qa:
description: Checks the final report renders and every link resolves.
model: [claude:haiku]
instructions: |
Open the report, follow every link, list broken ones.
writer:
description: Turns merged findings into the final report.
start: on_demand
instructions: |
Use templates/report.md. Plain language, no hype.
x-domios:
min_version: "1.20"