AI Research Toolkit

Structured workflows for complex projects in Claude Code

What Is the AI Research Toolkit?

A collection of 13 Claude Code commands that catch errors, prevent bad plans, and keep complex projects on track. Organized into three layers: verification, workflow, and content. Validated across 40+ production sessions.

Verification
/pcv /pcv-research /coa /pace /audit
Workflow
/startup /dailysummary /weeklysummary /commit /simplify /improve
Content
/quarto /pdftotxt

What Problem It Solves

AI coding assistants are powerful but unstructured. Three recurring gaps:

Verification gap

How do you know the AI’s output is correct — especially for citations, numerical results, and analysis?

Planning gap

Complex multi-component projects need structured planning, not ad-hoc prompting.

Documentation gap

Research progress disappears when the terminal closes. Session context is ephemeral.

This toolkit addresses all three gaps with reusable commands that add structure without adding friction. Each command was built for a real research task, validated in production, and generalized for others.

Commands

Verification Layer

CommandWhat it does
/pcv Plan-Construct-Verify — structured planning with sequential clarification, adversarial review, and human approval gates before any code or document is written. Based on PCV v3.9 (Dr. Kay, NC State).
/pcv-research Parallel planning experiments — runs depth-first and breadth-first planning strategies in parallel with instrumentation for cross-instance convergence analysis.
/coa Council of Agents — spawns specialists with distinct professional perspectives (Skeptic, Economist, Practitioner, etc.) to independently analyze a question, then synthesizes convergence and divergence.
/pace Parallel Agent Consensus Engine — routes a task through two independent players with coaching review, then cross-compares for verification through redundancy.
/audit Citation & numerical audit — verifies every cited metric exists on disk and every quoted number matches the source paper. Catches misquoted figures and fabricated citations before publication.

Workflow Layer

CommandWhat it does
/startupReads recent work summaries and orients you on where you left off — across all active workstreams.
/dailysummaryCreates a dated summary of the day’s work with cross-references to decisions and open TODOs.
/weeklysummaryAggregates daily summaries into weekly workstream reports; surfaces dormant threads and stale TODOs.
/commitAnalyzes staged changes and creates logically grouped commits — prevents mixed-concern commits.
/simplifyReviews code or documents for redundancy, complexity, and performance issues via 5-lens analysis.
/improveSelf-reflective meta-agent that audits your own Claude Code infrastructure and proposes improvements.

Content Layer

CommandWhat it does
/quartoGenerates Quarto RevealJS slide decks from background documents.
/pdftotxtExtracts text from PDF, Word, and HTML files — supports single files or entire directories.

Evidence

All evidence is from actual production usage (2025–2026). The same tools work across simulation, documentation, code review, and presentation workflows.

ToolWhat it caught or produced
/audit Caught a consumer surplus figure misquoted as −35% when the source paper actually states −2.5%. Also caught benchmark numbers cited to the wrong paper and a job displacement figure with inverted framing — all on a single poster before printing.
/pace Two independent players identified circular logic in a simulation’s acceptance rate calculation that passed single-agent review. The fix changed results by several percentage points.
/coa A multi-seat council evaluated external review concerns about a simulation’s learning mechanism. The Skeptic identified the one genuine threat (signal vs. noise) that subsequent diagnostics confirmed and cleared.
/pcv-research Parallel planning for a conference poster produced convergent decisions on layout and color, but diverged on figure placement — surfacing a genuine design ambiguity the researcher had not considered.
/pcv Structured the design of a conference poster, mathematical proof verification in Lean 4, TA grading workflows, and dissertation chapter outlines — each using the same clarify-before-building discipline.
/startup Across sessions spanning multiple workstreams (simulation, poster, toolkit, coursework), session startup consistently recovered full context in under 60 seconds.

Methodology note. All agents in PCV-Research, CoA, and PACE use the same underlying Claude model. Convergence between agents indicates consistency within the model’s reasoning space, not independent validation. Cross-model validation via Gemini partially addresses this limitation.

Installation

Prerequisite: Claude Code must be installed. The toolkit runs inside Claude Code, not separately.

  1. Clone the toolkit.
    git clone https://github.com/jbenhart44/Research-Toolkit.git
cd Research-Toolkit
  2. Run the installer.
    bash install.sh
    This installs all 13 commands and the verification hooks.
  3. Configure for your project.
    Edit ~/.claude/toolkit-config.md to set your project name and workstreams.
  4. Type /pcv in any Claude Code session to verify it works.

Cite This Toolkit

If you use or build on this toolkit, please cite:

BibTeX entry 
@software{benhart_kay2026toolkit,
  author = {Benhart, Jake and Kay, Michael G.},
  title  = {AI-Assisted Research Toolkit for Claude Code},
  year   = {2026},
  url    = {https://github.com/jbenhart44/Research-Toolkit}
}

The toolkit bundles PCV v3.9 (Dr. Kay, NC State) as its planning foundation.

Contact

Jake Benhart & Dr. Michael Kay
NC State University — Operations Research
jbenhart@ncsu.edu · github.com/jbenhart44