research
Use when the user asks to research, investigate, verify facts with sources, compare competitors, study a market or industry, review literature, prepare an evidence-backed report, decide whether a topic needs external Deep Research, or make an evidence-backed judgment or recommendation. Runs breadth-first Quick Recon, Claim Ledger gating, and an Adjudication Hypothesis Spine when needed. This generated target uses Codex native collaboration and web tools.
GiaSip · 已有 AI 审核记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add GiaSip/giasip-skills --skill research技能说明
name: research description: "Use when the user asks to research, investigate, verify facts with sources, compare competitors, study a market or industry, review literature, prepare an evidence-backed report, decide whether a topic needs external Deep Research, or make an evidence-backed judgment or recommendation. Runs breadth-first Quick Recon, Claim Ledger gating, and an Adjudication Hypothesis Spine when needed. This generated target uses Codex native collaboration and web tools."
✦ A GiaSip generated target · github.com/GiaSip/giasip-skills
GiaSip Research — Codex
Generated from the neutral canonical Research method. Do not edit this target by hand; reconcile changes into the canonical source and rebuild.
Codex Runtime Contract
Apply the compiled method directly in Codex; do not route through another Research skill or adapter to recover its core workflow.
- Inspect the callable collaboration schema before delegating and pass only supported fields.
- Use Codex's available public-web search and page-reading tools.
- Use 2 lightweight workers by default and 3 only for three genuinely orthogonal slices.
- Keep run IDs, artifact persistence, ledger mutation, synthesis, and delivery in the main task.
- If worker concurrency is unavailable, execute the selected slices sequentially and disclose the fallback.
Portable Research
This is the only human-maintained semantic source for the Research method. Claude Code, Codex, and public GiaSip packages are generated host targets; do not copy host tool names or invocation syntax back into this file.
Portability contract: “worker” means the current host’s independent task primitive. When parallel workers are unavailable, execute the same slices sequentially and disclose the fallback. Host-native wrappers define concrete tools, persistence roots, and invocation.
Core Principles
- Recon before escalation — Every research task starts with Quick Recon. 2-5 minutes of initial search helps you decide: deliver directly, or escalate to Deep Research with clear questions. Skipping Recon to submit Deep Research blindly wastes quota
- Capability fit first — When Deep Research is needed, the only criterion for platform selection is "who is best at this type of task," not cost — within your subscribed platforms
- Language determines the candidate pool — Chinese tasks prioritize domestic platforms, English tasks prioritize international platforms, mixed tasks use both
- Combination over single (high-stakes only) — Multi-platform cross-validation is only worth it for high-stakes questions (≥10pp numbers / licenses / policy-legal-financial / AI same-faction claims); for general topics (market/competitive/industry), a single platform + primary source grounding is sufficient — don't burn quota on unnecessary multi-platform runs
- Numbers and citations must be verified — All platforms can hallucinate; always remind the user to spot-check critical information
- Quota awareness — Some platforms have monthly caps (e.g., ChatGPT Plus 25/month); Recon helps you save quota for questions that genuinely need deep digging
- Verification priority invariant (core) — Primary source / locator grounding > source family convergence > heterogeneous model cross-check. First determine whether a claim has a ground-truth locator, then decide whether to spend on heterogeneous models. Heterogeneous reviewers cannot substitute for missing primary source locators (empirical: 1 model that read the primary source > 3 heterogeneous models guessing from memory). "Evidence source family" (owner/regulator/official/independent/vendor/aggregate) and "reviewer faction family" (cross-faction) are two dimensions — don't conflate them.
Core Flow
Step 0: Establish the Run Directory (persistence convention, spans the whole flow)
Any task that enters Recon, or skips Recon to escalate directly to DR, first fixes a run directory and physically persists all intermediate products — this is the prerequisite for Claim Ledger / Mini Assurance / Deep Research reflow to actually work. Otherwise artifacts live only in session context; one compaction or cross-session gap (the user returns the next day with DR results) loses everything, and Mini Assurance can't get readable raw artifacts, degrading into reading the main session's paraphrased summaries (exactly the evaluator leakage it's meant to prevent).
- Location: project research →
<project>/research/<topic>-<YYYY-MM-DD>/; no project home →~/research-runs/<topic>-<YYYY-MM-DD>/ - Structure:
<run_dir>/ manifest.md # run state anchor (cross-session recovery entry, see below) artifacts/ # each recon worker facet/gap's full raw output, one .md (incl. 00-discovery.md from Step 2A) snapshots/ # normalized main text of critical / high-risk sources, one file per claim_id (Step 2.5) quotes.tsv # externalized input for the quote gate, written before it runs (Step 2.5) ledger.md # Claim Ledger master table (Step 2.5) + Hypothesis Matrix as an independent section (Adjudication, Step 2.5) recon-report.md # final report for Recon direct delivery deep-research-prompt.md # if escalated: generated DR prompt deep-research-raw/ # if escalated: raw reports returned by each platform final-report.md # merged Recon + DR final version audit.md # Mini Assurance / fact-check audit results - The orchestrator owns path allocation and acceptance; the host runtime may write the bytes: every Round 1 / Round 2 worker's full raw output must land in
<run_dir>/artifacts/<NN>-<facet>.md, one unique file per worker, at a path the orchestrator assigns — never one the worker picks for itself, since self-named files collide and overwrite silently. If the host can direct a worker's output s