Claude AI skills are folders of instructions and scripts that Claude Code loads when a task calls for them. The two below exist for one reason: a chat model is good at sounding right, and both skills make Claude prove it. They are part of Mr Pirate's tools, they are MIT licensed, and they ship with Skill Vision as a three-skill suite.

Research Hound: a scientific research workflow for Claude Code

Research Hound turns Claude into a methodical research partner. It runs five stages in order, and every stage after the first has scripts you can run yourself:

Stage Scripts What runs
1. Critical thinking none Evidence hierarchy, logical fallacies, statistical pitfalls and bias catalogs
2. Literature review 3 Ranking and formatting of PubMed, arXiv and Semantic Scholar results, and verify_citations.py, which checks that each DOI resolves
3. Brainstorming 3 A scored ideation matrix and session scaffolds
4. Writing 8 An IMRaD scaffold, a manuscript linter, a claim audit, reference and consistency checks, and CONSORT, PRISMA or STROBE selection
5. Evaluation 7 Rubric validation, scoring and inter-rater agreement

That is 21 scripts in all. Claude's built-in search and Deep Research find sources and summarize them in one pass. Research Hound adds what a single pass cannot do for itself: it checks that a reference resolves and traces each claim in your manuscript back to its evidence. Every check is a plain command line tool, so you can run it again tomorrow and get the same answer.

Your work stays yours. Only verify_citations.py touches the network, to ask a DOI resolver whether a reference exists. No script needs an API key, and nothing sends your prose, data or figures to a hosted model.

Ask Claude things like "search the literature on X and verify every citation", "audit the claims in my methods section against my sources" or "is this analysis p-hacked?"

Agent Oracle: decide the design before the agents run

Agent Oracle is an AI agent orchestration design skill. Before any work starts, it answers four questions out loud:

  1. Scope. Does the request belong to one segment, or span the whole pipeline?
  2. Agent count. One agent or several, and why?
  3. Loop design. Does this need a loop, and what actually stops it?
  4. Harness controls. Which surfaces are locked, and how is the budget enforced?

Each answer cites a section of one of its five guides: evaluation, multi-agent patterns, harness engineering, long-horizon prompting and self-improvement loops. The guides lean on published research. One example: a study it cites found 68 confirmed infinite-loop failures across 47 of 6,549 repositories, most of them a retry cap that never reached the loop it was meant to stop. Agent Oracle defaults to a single agent and treats extra agents as something you justify, and it treats token, tool-call, time and cost limits as a contract that ends the run when any one is breached.

Ask Claude things like "design the supervisor topology for this pipeline" or "write the launch prompt for this long-running research agent." Two standard-library scripts come with it: an evaluator that scores a JSON test set against a weighted rubric and can gate CI, and a coordination demo of supervision, handoff, consensus and retry budgets.

Install either skill

git clone https://github.com/AlveeeRahman/research-hound.git ~/.claude/skills/research-hound
git clone https://github.com/AlveeeRahman/agent-oracle.git ~/.claude/skills/agent-oracle

Both need Python 3.10 or later. Research Hound also needs requests for its one networked script. Both are scanned with NVIDIA SkillSpector on every push. To use them on claude.ai, zip each folder without .git and keep the description under the uploader's 200-character limit. Both skills already fit it.

The directory lists them as Research Hound and Agent Oracle.