Deep research (Cursor)

Upstream inspiration: Weizhena/Deep-Research-skills and RhinoInsight (arXiv). This project vendors the JSON validator under .cursor/skills/deep-research/scripts/.

How the user runs it in Cursor

This is a project skill: it lives under .cursor/skills/deep-research/ in this repo so everyone who opens the workspace can use the same workflow and validator paths.

1. In Agent chat, attach this skill: @.cursor/skills/deep-research/SKILL.md, or pick deep-research from the skills UI if it appears there.
2. Give a topic and optional time range for web supplementation.
3. Ask the agent to execute Phase 1 → Phase 2 → Phase 3 below (or one phase at a time).

All paths below are relative to the project workspace root unless noted.

Validator path (fixed in this repo)

.cursor/skills/deep-research/scripts/validate_json.py

Dependency: pip install pyyaml (use a project venv if macOS PEP 668 blocks global install).

Example:

python3 .cursor/skills/deep-research/scripts/validate_json.py \
  -f "<topic_dir>/fields.yaml" \
  -j <topic_dir>/results/*.json

Phase 1 — Outline (/research equivalent)

1. Propose research items (who/what to cover) and a field framework (what each JSON must capture).
2. Web supplement: search or fetch current sources; add missing items/fields for the stated time range.
3. Create a slug from the topic (lowercase, underscores), e.g. ai_agent_memory_2026.
4. Write:

<topic_slug>/
  outline.yaml    # topic, items[], execution{}
  fields.yaml     # field_categories[] → fields[] (name, description, detail_level, required)

`outline.yaml` minimum shape

• topic: string
• items: list of { name, category, description }
• execution: { batch_size, items_per_agent, output_dir } — use output_dir: "./<topic_slug>/results"

`fields.yaml`: copy structure from .cursor/skills/deep-research/templates/fields.example.yaml or from ai_agent_context_management_research/fields.yaml in this repo. Every field used in JSON must appear under field_categories.

Human checkpoint: confirm items and fields before Phase 2.

Phase 2 — Deep item JSON (/research-deep equivalent)

For each item in outline.yaml:

1. Build one JSON file per item under execution.output_dir (e.g. <topic_slug>/results/Item_Name.json). Use flat keys at the top level matching every name in fields.yaml.
2. Fill unknowns with [uncertain] in the value string when needed; list those field names in top-level "uncertain": [].
3. Include primary_sources as a list of { "title", "url" } objects.
4. After each file (or the whole batch), run the validator (command above). Do not declare Phase 2 done until all required fields pass.

Resume: skip items whose JSON already exists and validate as PASS.

Human checkpoint: optional between batches using execution.batch_size.

Phase 3 — Report (/research-report equivalent)

1. Read outline.yaml (topic) and every results/*.json.
2. Write <topic_slug>/report.md with:

• Title + methodology (cite upstream + RhinoInsight).
• Table of contents with anchor links to each item.
• Per-item sections: all field categories; omit or shorten fields whose value contains [uncertain] or whose key is listed in that JSON’s uncertain array.

No separate generate_report.py is required unless the user asks for a reusable script.

Optional phases

• Add items: append to outline.yaml items list; re-run Phase 2 for new files only.
• Add fields: edit fields.yaml; backfill new keys into existing JSON or mark [uncertain].

Templates

See .cursor/skills/deep-research/templates/ for starter fields.example.yaml and outline.example.yaml.

Examples

Invocation

@.cursor/skills/deep-research/SKILL.md
Topic: vector databases for RAG in 2026. Run Phase 1 only; stop for my approval on outline.yaml and fields.yaml.

Validate after Phase 2 (replace <topic_slug> with your folder, e.g. vector_dbs_rag_2026):

python3 .cursor/skills/deep-research/scripts/validate_json.py \
  -f "<topic_slug>/fields.yaml" \
  -j <topic_slug>/results/*.json