SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsdeep-research

deep-research

Writing & research

View repository

Updated 10-16 06:16

About this skill

Deep Research (Production mode)

Conduct multi-source searches, evidence collection, and adversarial verification on any topic, and output a research report with citations.

Adapted from the Feynman CLI's 7-step deep research pipeline, with 100% coverage of its core process design.

Trigger scenarios

  • The user says "Help me deeply research/study/analyze X"
  • Need to gain a systematic understanding of an unfamiliar field
  • Need to produce a research report backed by citations
  • /deep-research <topic>

Do not trigger: simple Q&A, questions answerable by a single search, restating known information.

Required outputs

Each run must leave the following files on disk:

File Path Description
Research plan 01_项目/调研/.plans/<slug>.md Core questions, search strategy, task ledger, verification log
Draft 01_项目/调研/.drafts/<slug>-draft.md Initial report draft
Cited version 01_项目/调研/.drafts/<slug>-cited.md With inline citations and post-verification references
Final report 01_项目/调研/<YYYY-MM-DD>-<slug>.md Delivery version
provenance Attached at the end of the report Source list + verification status

Slug rules: extract from the topic, lowercase with hyphens, remove function words, at most 5 words. E.g. ai-coding-tools-landscape。

Degradation rules: If any step fails, still write the final BLOCKED or partial output and provenance. Never output only chat text. When verification cannot be completed, use Verification: BLOCKED。

7-step process

Detailed execution guidance see references/pipeline.md. Agent role definitions see agents/。

Step 1: Plan

Create immediately 01_项目/调研/.plans/<slug>.md, including:

  • Core question breakdown
  • Required evidence
  • Scale decision (simple / medium / complex)
  • Task ledger (status tracking for each subtask)
  • Verification log (updated step by step)
  • Decision log (record of key judgments)

The Scale decision is completed before assigning tasks. For narrow "What is X"-type questions, you must use lead-owned direct search,do not assign a researcher subagent。

Use mcp__Nowledge_Mem__memory_add to store the plan (key: deepresearch.<slug>.plan), so that subsequent sessions can be resumed.

Gate: After outputting the plan, pause and wait for the user to confirm the direction before continuing. If the user says "continue", "go with this", or "start directly", treat it as confirmed.

Step 2: Scale

Direct searchApplies to:

  • Single fact or narrow question (including "What is X"-type explanations)
  • Can be covered in 3-10 tool calls
  • ProhibitedInflating a simple explanation into multi-agent research

Subagent Use only when decomposition is clearly helpful:

  • 2-3 direct comparisons: 2 researchers
  • Broad research or multifaceted topics: 3-4 researchers
  • Complex multi-domain research: 4-6 researchers

Step 3: Gather Evidence

Avoid PDF parsing crashes. Do not call alpha_get_paper do not fetch .pdf URLs unless the user explicitly requests it. Prefer paper metadata, abstracts, HTML pages, official documentation, and web summaries.

Direct search mode：

  • do not spawn a researcher, search yourself
  • at least 3 sets of queries from different angles (definition/history, mechanism/principle, current status/comparison)
  • record the search terms used to <slug>-research-direct.md

Subagent mode：

  • first write a per-researcher brief (e.g., 01_项目/调研/.plans/<slug>-T1.md）
  • use the Agent tool to spawn, each agent reads its own brief file
  • for the researcher role definition see agents/researcher.md
  • set failFast: false
  • the researcher writes findings to a disk file and returns a one-sentence summary

Step 4: Draft

the main agent writes the report itself, without delegating synthesis. save to 01_项目/调研/.drafts/<slug>-draft.md。

include:

  • executive summary
  • findings organized by question/topic
  • caveats and disagreements supported by evidence
  • open questions
  • do not fabricatesources, results, data, benchmarks, images, charts, tables

pre-citation cleanup：

  • every key assertion, number, chart must map to a source URL, research note, or raw artifact path
  • remove or downgrade unsupported assertions
  • label inferences as inferences

Step 5: Cite (citation verification)

Direct search mode: do the citations yourself, verify URLs are reachable, write to 01_项目/调研/.drafts/<slug>-cited.md. Do not spawn a verifier.

Subagent mode: run the verifier agent after the draft exists. This stepmustbe completed before the reviewer. The verifier and reviewer must notrun in parallel.

for the verifier role definition see agents/verifier.md。

output:01_项目/调研/.drafts/<slug>-cited.md. Verify the file exists on disk.

Step 6: Review (adversarial review)

Direct search mode: review the cited draft yourself. Write <slug>-verification.md containing FATAL/MAJOR/MINOR. Fix FATAL. Do not spawn a reviewer.

Subagent mode: run the reviewer agent only after <slug>-cited.md exists.

for the reviewer role definition see agents/reviewer.md。

FATAL handling：

  • after fixing, run an additional round of review
  • record MAJOR in open questions
  • accept MINOR

how to fix: do not make multiple replacements in one giant edit call. For small fixes (1-3 places) use a local edit. For paragraph-level rewrites or >3 substantive fixes → read the full file and then write the full text to <slug>-revised.md。

final candidate:<slug>-revised.md(if it exists), otherwise <slug>-cited.md。

Step 7: Deliver

  1. copy the final candidate to 01_项目/调研/<YYYY-MM-DD>-<slug>.md
  2. append provenance at the end of the report (for the format see references/provenance-template.md）
  3. provide 3-5 alternative titles
  4. mcp__Nowledge_Mem__memory_add store key findings (tags:research,调研）
  5. before delivery, verify that all required outputs exist on disk

Agent invocation mode

Level Researcher Verifier Reviewer
simple do not spawn do not spawn do not spawn
medium do not spawn do not spawn do not spawn
complex 2-6 in parallel 1 serial 1 serial

All subagents hand off via files and do not return the full text to the context.

Quality gate

Before delivery, follow evals/output-quality-check.md for self-check:

  • All factual assertions have citations
  • No orphan citations / no orphan sources
  • FATAL = 0
  • Provenance verification status is not empty

Key constraints

  • All outputs are written to the vault, following workspace-boundary
  • Factual assertions must have citations; assertions without sources are labeled as "subjective judgment"
  • Do not fabricate data, cases, project results, or citation sources
  • The Plan Gate is a mandatory step and cannot be skipped
  • Never output only chat text; there must be disk file outputs