deep-research

A routing-first front door for the Deep Research workflow from Weizhena/Deep-Research-skills — a two-phase, human-in-the-loop research method (outline generation, then deep investigation) consolidated into one jeo-skill with 4 reference pipelines. Each pipeline ships its own prompt templates and output contract; this skill classifies the request into the right phase, loads that pipeline, and executes it exactly.

One topic → one extensible outline → parallel per-item investigation into validated JSON → one complete markdown report. Every phase has a user checkpoint, so you keep precise control at each stage instead of handing the model a black-box "research X" prompt.

The reference pipelines hold the dismantled-and-merged upstream skill text (per-command frontmatter stripped, headings nested, prompt templates kept verbatim) — not a paraphrase — so each phase is self-contained here. Read the matching pipeline reference before executing:

• references/outline-pipeline.md — /research · /research-add-items · /research-add-fields (Phase 1: generate + extend the outline)
• references/deep-pipeline.md — /research-deep (Phase 2: parallel per-item investigation + validate_json.py coverage gate)
• references/report-pipeline.md — /research-report (Phase 3: TOC + per-field markdown report)
• references/web-search-pipeline.md — the web-search agent + 5 routed source modules (github-debug · general-web · academic-papers · chinese-tech · stackoverflow)

Plugin Installation

# This routing skill via jeo-skills (verified path)
npx skills add https://github.com/akillness/jeo-skills --skill deep-research

# Global install for one or more agents
npx skills add -g https://github.com/akillness/jeo-skills --skill deep-research -a claude-code -a codex -y

# Scripted install with knobs (Python dep + upstream slash-command skills)
WITH_DEPS=1 AGENTS="claude-code,codex" bash .agent-skills/deep-research/scripts/install.sh

The deep phase calls scripts/validate_json.py (needs pip install pyyaml). The upstream repo also ships ready-made slash commands for Claude Code, OpenCode, and Codex — WITH_UPSTREAM=1 bash scripts/install.sh clones and copies them.

When to use this skill

• The user wants to research a set of comparable things (models, papers, tools, companies, products) along consistent fields, not a single Q&A
• The task is a survey / benchmark review / literature review / competitor analysis / due diligence that benefits from a structured outline first
• The user wants parallel, source-cited investigation that lands in a shareable markdown report with a table of contents

When not to use this skill

• A one-off factual question or single-source lookup → just use web search
• Full academic research-to-publication with citation gates and reviewer rounds → use academic-research
• Multi-agent build/verify orchestration of code → use oh-my-claudecode / oh-my-codex / oh-my-agent
• Token-efficient code discovery inside a repo → use semble
• Karpathy-style autonomous ML experiment search → use autoresearch

Required intake packet

Before routing, identify:

1. Phase — outline · deep · report (which stage of the workflow)
2. Topic — the research subject (becomes {topic} and the {topic_slug}/ working dir)
3. Working dir — existing {topic_slug}/ with outline.yaml + fields.yaml, or new
4. Time range — for web-search supplementation (e.g. last 6 months, since 2024, unlimited)
5. Output target — the outline files, the per-item JSON, or the final report.md

Phase Routing Table

Instructions

Step 1: Pick the phase

Classify the request against the routing table. State the chosen phase → command explicitly before producing output (e.g. "outline → /research"). If a {topic_slug}/outline.yaml already exists in the working directory, default to the next unfinished phase (outline → deep → report) unless the user asks to extend the outline.

Step 2: Load the pipeline

Read the matching reference file and follow its workflow and prompt templates:

outline      → references/outline-pipeline.md
deep         → references/deep-pipeline.md
report       → references/report-pipeline.md
web-search   → references/web-search-pipeline.md

Every per-item / supplement search delegates to the web-search agent — always load web-search-pipeline.md and the relevant source module(s) before calling WebSearch.

Step 3: Execute with the pipeline's discipline

• Hard constraint on prompt templates: the outline and deep pipelines define prompt templates that must be reproduced verbatim — only substitute {xxx} variables; never edit structure or wording.
• Human-in-the-loop: confirm with the user at each AskUserQuestion gate (items, fields, time range, batch size, TOC fields) before moving on. Run deep research batch-by-batch with approval between batches.
• Evidence-first: every supplemented item/field and every per-item JSON must carry source links. Mark unknowns [uncertain] and list them in the uncertain array — never fabricate a value.
• Validate before done: a deep-phase item is complete only after validate_json.py passes (full required-field coverage).

Step 4: Return the phase's output packet

Close with a one-line Next step pointing to the next phase (outline → /research-deep; deep → /research-report).

Integrity principles

• No fabrication: every item, field value, and finding needs a source; unverifiable values are marked [uncertain], not invented.
• Verbatim prompts: the upstream prompt templates are a hard contract — substitute variables only.
• Human checkpoints: outline contents, time range, batch size, and TOC fields are confirmed with the user, not assumed.
• Coverage gate: deep-phase JSON must pass validate_json.py before an item counts as done.

Route-out map