aiwf-deep-research

AIWF deep research workflow for source-backed, weighted, multi-domain investigation. Use when Codex must research, verify, or build a mental model across arXiv, Hugging Face, GitHub, Civitai, Reddit w

SKILL.md

GitHub

AIWF Deep Research

Overview

Use this skill to run research as an evidence pipeline, not a search session. Build a source plan first, weight every source by what it can actually prove, keep a claim ledger, and validate the receipt before using findings in an answer, plan, code change, or training data.

Use AIWF_DEEP_RESEARCH_EXTRA_URLS for required seed URLs that must appear in the source plan. The value is comma-separated. scripts/init_research_run.py also accepts repeated --extra-url flags and writes all required seed URLs to source_plan.json.

Continue an active /goal when the host exposes one, but do not claim the skill can change model, context, or tool budgets controlled by the host.

Routing

Use this skill before aiwf-agent-mok when the primary work is external research or source selection. Add aiwf-agent-mok when the research must also update plan.md, create findings datasets, or verify a multi-step project route.

Use other AIWF skills after this one when research becomes implementation:

  • Model loading, precision, quantization, LoRA compatibility: aiwf-model-loader.
  • Pipeline/runtime readiness and model-source mismatch: aiwf-ai-pipelines.
  • Training, LoRA, QLoRA, checkpoints, export: aiwf-local-ai-training.
  • Evals, benchmark receipts, promotion gates: aiwf-ai-evals.
  • Serving and endpoint checks: aiwf-inference-serving.
  • Repo edits after research: aiwf-repo-sentinel, then the focused implementation skill.

Required Workflow

  1. Restate the research question and intended decision.
  2. Read references/source-policy.md for every run.
  3. Read references/domain-targets.md before choosing sources.
  4. Read references/research-loop.md before writing the final synthesis or receipt.
  5. Read references/reddit-policy.md before using Reddit or community sources.
  6. Create a research run folder with scripts/init_research_run.py.
  7. Generate or write a source plan before searching.
  8. Search in this order: local/user evidence, primary sources, scholarly discovery indexes, implementation sources, community sources.
  9. Put important claims into claims.jsonl with source_ids, evidence_weight, and weight_reason.
  10. Preserve contradictions instead of averaging them away.
  11. Validate the run with scripts/validate_research_receipt.py.
  12. Use only validated or clearly labeled uncertain findings in the final answer.

Source Rules

Do not treat sources as interchangeable.

  • Papers, standards, official docs, official repos, and local receipts carry different claims.
  • Civitai is valid for model/resource metadata and ecosystem signals, not scientific proof.
  • Reddit is lead generation and practitioner context, not final evidence.
  • Open-source library docs prove API/runtime behavior, not scientific validity unless backed by primary research.
  • College and graduate materials are useful, but course notes, theses, dissertations, and student projects need explicit academic-level labels and weights.
  • Store citations, source metadata, summaries, and short permitted excerpts. Do not store long copyrighted source text.

Helper Scripts

Create a run:

python <this-skill>\scripts\init_research_run.py --title "wan attention research" --request "Compare WAN attention runtimes for AIWF Studio."
python <this-skill>\scripts\init_research_run.py --title "wan attention research" --request "Compare WAN attention runtimes for AIWF Studio." --extra-url "https://arxiv.org/abs/2405.00000"

Route a prompt to source targets:

python <this-skill>\scripts\source_target_router.py --prompt "Research Civitai LoRA metadata and verify against Hugging Face and GitHub."

Validate a run:

python <this-skill>\scripts\validate_research_receipt.py "research runs\20260701-120000-wan-attention-research"

Output Contract

Every non-trivial run should leave:

  • request.md
  • source_plan.json
  • queries.jsonl
  • sources.jsonl
  • claims.jsonl
  • source_weights.jsonl
  • contradictions.jsonl
  • research_brief.md
  • validation.json

The final user answer should summarize findings, cite source URLs, state evidence strength, list contradictions or gaps, and identify what should be checked next.