Research

Research — Build Outline

Trigger

/ultra-research:research <topic>

Manual-only. disable-model-invocation: true is set in frontmatter so Claude never auto-runs this skill from description matching — it only fires when the user types the slash command.

Pipeline Context

This is phase 1 of 5. The full flow is:

1. /ultra-research:research <topic> — (this skill) builds outline.yaml + fields.yaml
2. /ultra-research:research-add-items (optional) — extend the item list
3. /ultra-research:research-add-fields (optional) — extend the field list
4. /ultra-research:research-deep — fan out parallel web-search subagents, one per item, write per-item JSON to results/
5. /ultra-research:research-report — consume results/*.json and emit a single markdown report.md

At the end of this skill, ask the user which phase to enter next (see Step 5).

Workflow

Step 1 — Generate Initial Framework from Model Knowledge

Use the model's existing knowledge of {topic} to draft:

• A main items list (research objects in this domain)
• A suggested field framework (categories + fields to capture per item)

Print the draft as {step1_output}. Use AskUserQuestion to confirm:

• "Add or remove any items?"
• "Does the field framework cover what you need?"

Step 2 — Web Search Supplement (Single Pass)

Ask the user for a time range via AskUserQuestion (offer "last 6 months", "since 2024", "unlimited", or custom).

Get the current date once with date +%Y-%m-%d and bind it to {YYYY-MM-DD}.

⛔ Dispatch-only. The orchestrator MUST NOT call WebSearch, WebFetch, or mcp__exa_websearch__* directly. Instead, launch one web-search-agent subagent (background, via the harness's Agent/Task tool) with the verbatim prompt template documented in references/prompt-templates.md. The template is a hard constraint — only substitute the {variable} placeholders; do not rewrite the structure or wording. See that file for the template, the variable table, and a one-shot example.

The subagent returns supplementary items, recommended fields, and source links — capture this as {step2_output}.

Step 3 — Merge with User's Existing Fields (Optional)

Use AskUserQuestion to ask if the user has an existing field-definition file to merge in. If yes, read it and union the definitions.

Step 4 — Write outline.yaml and fields.yaml

Merge {step1_output}, {step2_output}, and any imported user fields. Write two files into ./{topic_slug}/:

outline.yaml — items list and execution config:

topic: <Research topic>
items:
  - name: <item name>
    category: <optional grouping>
    description: <one-line description>
execution:
  batch_size: <parallel-agent count, confirm with AskUserQuestion>
  items_per_agent: <items per agent, confirm with AskUserQuestion>
  output_dir: results

fields.yaml — field definitions:

field_categories:
  - category: Basic Info
    fields:
      - name: <field name>
        description: <what to capture>
        detail_level: brief | moderate | detailed
        required: true | false
uncertain: []   # reserved — populated by /ultra-research:research-deep

The detail_level ladder (brief → moderate → detailed) hints to the deep-research subagents how thorough to be per field.

Step 5 — Inspect Coverage and Route to Next Phase

Before asking, Read the two files you just wrote and apply this thinness rubric to decide which next-step options to mark (Recommended):

Both checks are independent — mark either, both, or neither.

Then use AskUserQuestion with header: "Next step" and these options (append (Recommended) to the label when the rubric flags it):

• Add more items — extend outline.yaml (description: "Run /ultra-research:research-add-items to grow the item list.")
• Add more fields — extend fields.yaml (description: "Run /ultra-research:research-add-fields to enrich the field schema.")
• Start deep research — kick off the heavy phase (description: "Run /ultra-research:research-deep to fan out parallel subagents.")
• Done for now — stop here (description: "Print the directory path and exit. You can resume from any phase later.")

In the question body, briefly tell the user why each recommendation was made (e.g., "Only 5 items so far; add more before deep research"). Be one sentence per reason.

If the user picks one of the first three, invoke the corresponding skill immediately via the Skill tool (no need for them to type the slash command). If "Done for now," print the directory path and stop.

Output

{cwd}/{topic_slug}/
├── outline.yaml    # items list + execution config
└── fields.yaml     # field definitions

References

• references/prompt-templates.md — verbatim Step 2 prompt template, variable table, one-shot example. Read it before launching the web-search subagent.