---

name: research-synthesis description: > Evidence-first research synthesis workflow for market research, industry analysis, competitive research, policy or company information synthesis, file-only or file-augmented analysis, technical architecture reviews, risk assessment, and report-ready conclusions. Use when the user asks for comprehensive, cross-verified research or deep summarization rather than a simple lookup. keywords: [ research, synthesis, summary, summarization, report, market, industry, competitor, policy, company, evidence, sources, deep-research, cross-verification, source-backed, market-research, industry-analysis, competitive-research, policy-analysis, company-research, risk-assessment, technical-research, file-analysis, multi-agent, parallel-research, 深度总结, 总结, 研究, 行业研究, 行业分析, 市场研究, 竞品分析, 政策分析, 公司研究, 资料综述, 报告, 证据, 来源, 交叉验证, 多智能体, 并行研究, ] metadata: short-description: Evidence-first research synthesis

Research Synthesis

Use this skill for substantial research where the answer must be grounded in real files, logs, sources, or current external evidence. Do not use it for simple factual lookup, one-source Q&A, or ordinary code changes.

Hard Rules

• Inspect the real provided files, logs, repo paths, or source artifacts before explaining behavior or generating conclusions.
• Resolve one absolute {output_dir} from the current conversation before writing. If the model created a task directory, use it; otherwise use the unchanged session cwd. Save every research artifact under {output_dir}/research/ and pass absolute paths to tools and scripts.
• Reserved research artifact templates override the generic hyphen-separated filename style. Keep hyphens inside {topic}, but join every reserved suffix with the literal underscore shown in the template. For topic ai-quality-scheduling, write ai-quality-scheduling_dim01.md and ai-quality-scheduling_cross_verification.md; never write ai-quality-scheduling-dim01.md or ai-quality-scheduling-cross_verification.md. Copy the Required Outputs filenames exactly for every dimension and handoff artifact.
• Do not infer an output root from environment variables or legacy host metadata. The chosen {output_dir} is ordinary task organization and does not change the session cwd or workspace.
• Never save research artifacts directly in {output_dir}; keep them under its research/ child.
• Avoid CLI-specific assumptions. Use only tools actually available in the current host runtime.
• Use native subagents only when the user explicitly asks for multi-agent, parallel-agent, or delegation work, or when the host runtime has clearly authorized subagent use for the task.
• When subagents are authorized, use multiple subagents for independent facets or dimensions instead of defaulting to a single delegated worker.
• If subagents are not authorized or unavailable, run the same workflow locally in sequential rounds with a smaller but explicit evidence budget.
• For file-only requests, do not perform external search.
• For current or time-sensitive claims, check the current date/time before analysis and make search windows explicit.
• Match the user's search language unless the task requires a specific locale.
• Search depth comes from covering distinct evidence gaps, not from rephrasing a query that already returned usable evidence. Before every later batch, name the still-uncovered dimension or conflict; do not rerun a near-equivalent entity/fact query merely because an authority-ranked or site: query was empty.
• Treat HTTP 403/429 responses, CAPTCHA or challenge pages, robots or permission denials, repeated timeouts, and repeatedly empty or unreadable responses as blocked-source signals. For the same evidence target, allow at most one materially different recovery attempt after the first such failure. A valid recovery changes the source, exact URL, or available reading backend; a rephrased query or the same URL/backend is not a recovery.
• After two consecutive blocked-source failures for the same evidence target, stop searching that target for the current turn. Promptly tell the user what target is blocked, what was attempted, which failure signals were observed, what evidence remains usable, and which claim stays unverified. Continue with independent dimensions when possible and finish with an honest partial or framework handoff; do not silently spend the remaining search budget on near-equivalent retries or imply that retrieval is still running.
• If an exact public URL is already known, read it with an actually available exact-page tool. Prefer web_extract, or a user-selected MCP reader that accepts the exact URL and returns that page's body. In officev3, managed_browser_* is also valid for independent public-web reads. Activate managed_browser_navigate and managed_browser_snapshot together, then verify one URL at a time with navigate followed by snapshot. Navigation metadata alone is not a failed body read; the snapshot body is the evidence content bound to that URL. Use user_browser_* when the read depends on the user's current page, login state, cookies, extensions, or intranet access; this does not require a separate authorization prompt. Do not route between browser modes through a source_preference argument.
• Search-result summaries may support medium-confidence evidence when each row is bound to that result's exact URL, uses evidence_basis=search_summary, and copies an excerpt returned for that same result. Never cross-bind one result's summary to another URL or locally rewrite the excerpt. Read the exact page when stronger evidence is useful or the summary does not support the claim.
• After bounded search, verify no more than five unique exact article, report, filing, or data-page URLs. Never use an origin homepage or retry the same URL with the same page-reading backend. Do not issue parallel Playwright navigations; finish each navigate/snapshot pair before opening another URL. If two consecutive direct reads yield no usable source text, stop browsing, mark the remaining candidates unverified, and finish the handoff so downstream work can continue in partial or framework mode.
• Maintain {output_dir}/research/{topic}_evidence.json while researching. Every downstream factual claim must be bound to one target entity, source URL, source type, page excerpt, confidence, and verified / conflicting / unverified status. Use the exact schema in output_contract.md.
• Record known official domains for target entities and prefer verified first-party sources when a claim concerns that entity's own disclosure. Missing per-entity or first-party coverage is a research-quality warning, not a reason to discard other verified evidence or block downstream delivery. first_party is a claim about URL ownership, not a synonym for a high search authority score.
• When user-provided material conflicts with a source, preserve the user claim, mark user_input_alignment=conflicting, explain the conflict, and exclude that record from downstream synthesis until resolved. Never silently prefer or merge one side.
• Use standard Markdown footnotes in research artifacts: [^id] inline plus [^id]: Title. Date. URL definitions.

Start

1. Set {output_dir} to the absolute directory selected in the current conversation. Use a model-created task directory when one exists; otherwise use the unchanged session cwd. Keep this choice consistent for related follow-ups.
2. Resolve {skill_dir} to the active installed directory containing this SKILL.md. Do not assume the current working directory is the skill directory.
3. Create {output_dir}/research/.
4. Pick a stable topic slug:
   • Prefer short ASCII, lowercase, hyphenated words.
   • If the topic is mostly non-ASCII or ambiguous, use research-YYYYMMDD-HHMMSS.
5. State the selected route and one-sentence rationale before researching.
6. Follow the route details in routes.md.

Routing

Signal	Route
Files plus "only based on files", "no search", or equivalent	C: File-only
Files plus "refer to", "combine with", or no explicit restriction	D: File-augmented
No files plus broad landscape question	A: Wide search
No files plus specific bounded question	B: Focused search

When unclear, prefer Route D over Route C for file tasks, and Route A over Route B for broad multi-faceted topics. If that would conflict with an explicit user constraint, honor the constraint.

Subagent Policy

This skill supports multi-agent research but does not require it.

When subagents are authorized:

• Dispatch multiple bounded, independent research tasks using the host's native subagent tool. Prefer one subagent per facet, dimension, or validation conflict when those tasks can write separate artifacts.
• Default to up to 8 concurrent subagents per round unless the host states a different limit. For larger decompositions, launch additional bounded rounds.
• Give each agent the mission, context, allowed/disallowed sources, exact output format, and exact output path.
• Do not assume subagents inherit main-agent context; pass paths or excerpts.
• Do not launch duplicate agents on the same question.
• Do not let two subagents write the same artifact path; merge or synthesize in the main agent after their outputs are complete.

When subagents are not authorized or unavailable:

• Execute dimensions sequentially in the main agent.
• Preserve the same artifact names and citation contract.
• Reduce search counts pragmatically and record the reduced budget in {topic}_cross_verification.md.

Use prompts.md for subagent/local-round templates.

Required Outputs

All files live under {output_dir}/research/.

File	Route	Purpose
{topic}_file_analysis.md	C, D	File inventory, extraction, cross-file mapping
{topic}_wideNN.md	A	Wide exploration facets
{topic}_dimNN.md	A, B, C, D	Dimension-specific evidence
{topic}_validationNN.md	Optional	Conflict-specific validation output before merge
{topic}_cross_verification.md	A, B, C, D	Confidence tiers and conflict zones
{topic}_insight.md	A, B, C, D	Cross-dimension insights
{topic}_evidence.json	A, B, C, D	Entity-bound evidence ledger and verified handoff
{topic}_final.md	Optional	Final Markdown report when no writing skill is available

Before final handoff, run:

RESEARCH_SYNTHESIS_SKILL_DIR="${BOX_AGENT_RESEARCH_SYNTHESIS_SKILL_DIR:-${RESEARCH_SYNTHESIS_SKILL_DIR:-{skill_dir}}}"
if [ ! -f "$RESEARCH_SYNTHESIS_SKILL_DIR/scripts/validate_research_artifacts.py" ] && [ -f "$HOME/.box-agent/skills/research-synthesis/scripts/validate_research_artifacts.py" ]; then
  RESEARCH_SYNTHESIS_SKILL_DIR="$HOME/.box-agent/skills/research-synthesis"
fi
if [ ! -f "$RESEARCH_SYNTHESIS_SKILL_DIR/scripts/validate_research_artifacts.py" ] && [ -f "$HOME/.box-agent/skills/deep-research-swarm-officev3/scripts/validate_research_artifacts.py" ]; then
  RESEARCH_SYNTHESIS_SKILL_DIR="$HOME/.box-agent/skills/deep-research-swarm-officev3"
fi
if [ ! -f "$RESEARCH_SYNTHESIS_SKILL_DIR/scripts/validate_research_artifacts.py" ]; then
  echo "ERROR: validate_research_artifacts.py not found under $RESEARCH_SYNTHESIS_SKILL_DIR/scripts" >&2
  exit 1
fi
${BOX_AGENT_PYTHON:-python3} "$RESEARCH_SYNTHESIS_SKILL_DIR/scripts/validate_research_artifacts.py" --research-dir "{output_dir}/research" --topic "{topic}" --route A --report "{output_dir}/research/qa/{topic}_research_check.json"

Adjust --route for the selected route. --min-dimensions N is a recommended quality target, not a deck-delivery threshold. Record the actual reduced budget in the cross-verification file. The validator emits a producer-neutral presentation_handoff object. Downstream presentation workflows consume its delivery_mode, verified_facts, gaps, and quality_summary; they do not interpret this skill's internal QA state. full, partial, and framework remain deliverable even when research quality is incomplete. Only verified rows that passed entity, excerpt, source, first-party, and conflict checks enter verified_facts; conflicting and unverified rows stay in gaps/limitations.

Final Handoff

If a report-writing, paper-writing, document, or presentation skill is available and the user requested that output type, hand off explicit file paths and state the generic presentation_handoff.delivery_mode. For partial, downstream work uses only the verified subset and marks or omits gaps; for framework, it creates a usable placeholder structure without external factual claims. For a presentation, hand off only presentation_handoff.verified_facts[].canonical; prose artifacts are context, not an independent fact source. Otherwise, produce {topic}_final.md from the verified research artifacts.

For technical product or codebase research tasks:

• Preserve product/runtime ownership boundaries in the synthesis when relevant.
• Distinguish UI assumptions from runtime evidence.
• Call out whether a finding is source-tree evidence, runtime-bundled evidence, log evidence, user-file evidence, or external-source evidence.
• Keep recommendations narrow and reversible unless the user asked for broader strategy.