Skip to content

ZihanZheng2000/human-gated-ai-research-framework

Repository files navigation

AI-Native Academic Research Framework

A human-gated, skill-guided research framework for turning research questions, literature, public or local data, and domain expertise into inspectable academic research packages.

This project is designed for researchers who want AI help with planning, data/model execution, research reporting, and review, but do not want an opaque fully autonomous research agent. The framework is implemented as a lightweight workflow kit and keeps the human researcher in control through explicit stage packages, evidence checks, and user gates.

What It Is

The framework is implemented through a four-stage workflow:

  1. Planning: clarify the research purpose, run Domain Onboarding on user-supplied materials, scan prior work, check feasibility, calibrate venue/output, and produce one Approved Planning Package.
  2. Modeling: acquire real usable data, run exploratory demos or full execution, validate outputs, produce figures/tables, and write one Modeling Package.
  3. Reporting: package evidence, provenance, findings, uncertainties, and final deliverables such as reports, papers, slides, dashboards, web pages, software packages, or review packets.
  4. Reviewing: run reviewer-agent critique, consolidate findings, and route issues back to the correct stage or to finalization.

Each stage ends with a gate. A real research project uses real user gates. A framework test can use synthetic gates, but they must be labeled as synthetic.

Each stage uses a maker-checker loop before the user gate: the researcher agent drafts the stage package, a separate reviewer agent critiques it, the researcher agent responds and fixes or routes the findings, and only then does the user approve, revise, or reject the gate.

When the runtime supports structured choices, gate decisions are presented as clickable options such as approve, revise, or backtrack. The same gate can still accept a free-form condition or correction from the user.

Planning is not just a form to fill before execution. For real research, the researcher agent should do a focused prior-work and novelty calibration before the full Planning Package is treated as ready. That calibration can be light, but it should ask whether the idea has already been studied, what adjacent work did, what gap or value remains, and whether the project is worth framing as a paper, demo, replication, tool, dataset, or internal report. It may be skipped or deferred only when the user explicitly chooses that route, and the package must record the accepted risk and revisit trigger.

Modeling starts only after the Planning Gate is approved. A user confirming the topic, research interest, or rough direction is not the same as Planning Gate approval.

The researcher agent is responsible for leading the workflow step by step. It should state the current stage, active package, next concrete action, and any needed user decision whenever work begins or pauses. Users provide domain judgment and gate decisions; they should not have to remind the agent what the next workflow step is.

Why This Exists

Many AI research tools can retrieve literature, write text, generate code, or automate agent loops. This framework focuses on a narrower practical problem: helping non-CS domain researchers build a traceable academic research package without losing control of evidence, claims, methods, and responsibility.

The core design choices are:

  • human-gated rather than fully autonomous
  • evidence-grounded rather than memory-driven
  • skill-guided rather than monolithic
  • execution-aware rather than prose-only
  • route-aware, so problems go back to Planning, Modeling, Reporting, Reviewing, or termination
  • domain-aware, so non-CS researchers can supply their own field materials at the start of Planning

Supported Agent Configurations

This repository supports two role configurations, switchable via WORKFLOW-CONFIG.md:

Config Researcher agent Reviewer agent
A Codex Claude Code
B Claude Code Codex

If you are a user (not an AI agent), read USER_NOTE.md first.

Quick Start

For users: Read USER_NOTE.md. It tells you what to say to the AI at each stage, how to provide domain materials, and how to hand off between agents.

For the researcher agent (Codex or Claude Code):

  1. Read WORKFLOW-CONFIG.md to confirm your role.
  2. Read AGENTS.md (Codex) or CLAUDE.md (Claude Code).
  3. Read docs/quickstart.md.
  4. Choose full workflow mode or smoke-test mode.
  5. At the start of Planning, check for user-supplied domain materials in artifacts/<run_id>/domain-materials/ and run Domain Onboarding if present.
  6. For a real project, run a focused literature/prior-work and novelty check before finalizing the Planning Package. Store detailed notes outside the package and summarize the implications in templates/plan-package.md.
  7. Start from templates/plan-package.md for a real project, or templates/smoke-test-package.md for a framework test.
  8. Run Planning -> Modeling -> Reporting -> Reviewing using the stage specs in docs/.
  9. Record every stage output as a package; write a review request for the reviewer agent before each gate.

Researcher agents must lead one active stage at a time. Do not pre-fill downstream stage packages before their upstream gates are approved, except as empty scaffold placeholders.

Repository Map

Path Purpose
WORKFLOW-CONFIG.md Switch between Config A (Codex=researcher) and Config B (Claude Code=researcher)
AGENTS.md Codex instructions — researcher role (Config A) or reviewer role (Config B)
CLAUDE.md Claude Code instructions — reviewer role (Config A) or researcher role (Config B)
USER_NOTE.md User guide: how to interact with the agents, gate language, handoff scripts
docs/ Stage specifications, handoff rules, skill strategy, reviewer-agent prompts, quick start, and paper draft
templates/ Package templates for Planning, Modeling, Reporting, Reviewing, smoke tests, and research-skill cards
skills/ Workflow-native skills used inside stages
.claude/commands/ Claude Code slash commands for all seven workflow-native skills
examples/ Worked examples, smoke tests, and per-example run notes
notes/ Cross-run open questions and development material

Stage Outputs

Stage Specification Main package
Planning docs/plan-stage.md templates/plan-package.md
Modeling docs/model-stage.md templates/model-package.md
Reporting docs/reporting-stage.md templates/reporting-package.md
Reviewing docs/review-stage.md templates/review-package.md
Smoke-test mode all stage specs templates/smoke-test-package.md

See docs/stage-handoffs.md for package flow, backtracking rules, and gate semantics.

Skills

The workflow uses small reusable internal skills rather than one all-purpose agent. Each skill has a SKILL.md in skills/ and a matching Claude Code slash command in .claude/commands/. External tools and installable skills may still be used for literature search, analysis, citation management, writing, or visualization; the internal skills keep those outputs tied to stage packages, evidence records, review routes, and gates.

Skill Slash command Used mainly in Responsibility External relationship
venue-calibration /venue-calibration Planning, Reporting Calibrate target journal, conference, preprint, policy note, or other output route Wraps venue pages, templates, and exemplars
research-skill-card-distiller /research-skill-card-distiller Planning Convert retrieved papers or user-supplied domain materials into reusable research-skill cards Uses external literature/deep-research outputs as input when useful
model-contract-runner /model-contract-runner Modeling Execute the approved modeling contract with acquisition checks and bounded loops Orchestrates Python, R, Jupyter, APIs, containers, or domain packages
figure-table-narrative /figure-table-narrative Reporting Decide visual purpose, evidence status, claim alignment, and caption limits Hands method/conceptual visuals to CCF-Figure or diagram tools; hands empirical visuals to Model/plotting tools
claim-evidence-mapper /claim-evidence-mapper Reporting, Reviewing Tie research-output claims to citations, artifacts, results, or limitations May use citation/retrieval tools, but keeps the final map internal
route-aware-reviewer /route-aware-reviewer Reviewing Stress-test the research output and route findings to the correct stage May use external reviewer simulation or checklists
gate-manager /gate-manager All stages Record gate decisions and stage transitions Mostly framework-native

See docs/skill-strategy.md for when to use external tools versus internal skills, and for Claude Code built-in capability coverage when Claude Code is the researcher (Config B).

Domain Onboarding

Non-CS researchers often have domain materials the AI has not seen: institutional reports, field-specific papers, data dictionaries, or policy documents. The Planning stage begins with a Domain Onboarding step that reads user-supplied materials, extracts a Domain Knowledge Summary, and produces skill cards — before any literature scan or plan writing begins.

Place domain materials in artifacts/<run_id>/domain-materials/ and tell the researcher agent at the start of Planning. See USER_NOTE.md for the exact prompt to use.

Cross-Agent Handoff

The researcher and reviewer agents communicate through two files written to artifacts/<run_id>/:

  • review-request.md — written by the researcher when a stage package is ready for critique
  • reviewer-critique.md — written by the reviewer with a structured findings table

See docs/reviewer-agents/cross-agent-handoff.md for the full protocol and file formats.

This repository also includes optional one-command wrappers for single-window workflows:

  • scripts/run_reviewer_claude.ps1 lets a Codex researcher call Claude Code as reviewer under Config A.
  • scripts/run_reviewer_codex.ps1 lets a Claude Code researcher call Codex as reviewer under Config B.

Both wrappers preserve the same file protocol: the researcher still writes review-request.md, the reviewer writes reviewer-critique.md, and the researcher remains responsible for triage and gate presentation.

Examples

Example folders should follow docs/example-organization.md. The public v0.1 repository keeps examples lightweight so private notes, large generated artifacts, and copyright-sensitive materials are not accidentally published.

  • examples/: example organization notes and planned case-study slots.

Paper

A Human-Gated, Skill-Guided Framework for AI-Native Academic Research

A companion paper is in preparation. It positions the repository as a practical, inspectable framework rather than a claim of full scientific automation. The draft is not yet public.

Current Status

This is an early research-framework repository. The workflow docs, templates, skills, and examples are present. The next milestone is a cleaner public v0.1 release with:

  • polished paper draft
  • one reproducible smoke test
  • one domain case study
  • repository-level citation and license metadata
  • clearer evaluation plan

License

MIT License. See LICENSE.

About

Human-gated, skill-guided framework for evidence-grounded AI-assisted academic research by domain researchers.

Resources

License

Contributing

Stars

0 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors