claude.humanitarians.ai · HAI Fellows Research Project
How do you algorithmically find the best tools for Claude in a sea of 59,000+?
This is an active research project. Humanitarians AI fellows build and test the methods — discovery pipelines, ensemble ranking signals, sandboxed audits, dedup clustering — and publish the results that hold up to scrutiny.
The catalog below is the publication surface. Every listing requires a completed audit signed with the commit SHA it ran against. The research is the process; the catalog is the evidence.
Discovery
Crawling GitHub, the claude.ai marketplace, and community indexes to find candidate tools at scale — then filtering to what is actually installable.
Ensemble Ranking
Combining multiple signals algorithmically: install success rate, risk scan results, code-to-prose ratio, commit activity, and star count — no single metric is sufficient.
Audit Pipeline
Sandboxed install checks, static risk scans (outbound network calls, exec patterns, prompt-injection strings), and triviality filters on each candidate. Each audit is signed with the commit SHA it ran against.
Dedup & Clustering
Identifying near-identical tools by code-similarity clustering and verifying whether the clusters are genuine duplicates or independent implementations of the same idea.
Fellows contribute discovery runs, audit methodology, and ranking experiments. Research happens in the open — findings ship to this catalog when they pass scrutiny.
Join as a Fellow →See listing criteriaEntries listed
16
Entries tested
16
Last audit
2026-08-03
Four coordinated skills for reading, editing, and creating Excel, Word, PDF, and PowerPoint files — including financial modeling standards and tracked-change support.
Document-skills is the most polished of the Anthropic example-skill bundles: four coordinated SKILL.md files and Python renderers covering the full Office docum…
Builds, evaluates, and iteratively improves Claude Code skills using an eval-loop: write → test → grade → rewrite → repeat.
Skill-creator closes the loop on skill development: write a draft, run eval queries through a fresh Claude subprocess, grade pass/fail with a separate grader ag…
Full read/write access to Linear via a single GraphQL shell helper — list, search, create, update, and transition issues without leaving your Claude session.
The Linear tag plugin ships a single well-structured skill and one shell helper that handles pagination, output format flags, and JQL-equivalent filtering over …
Prompt-only reference for querying Grafana dashboards, datasources, alert rules, and annotations via the REST API.
Grafana is a documentation-only skill: the SKILL.md contains structured curl recipes for dashboards, datasource queries, alert rules, silences, annotations, and…
Full read/write access to Jira Cloud via JQL search, issue transitions, comments, and sprint management with a pagination-aware shell helper.
The Jira tag plugin mirrors the Linear one in structure — one skill, one shell helper, credentials injected by the runtime — but adds something the Linear plugi…
Two lesson-design skills for K-12 educators: standards-aligned lesson planning and differentiation across math, ELA, science, and social studies.
K12-teacher-skills is the most vertically focused first-party plugin in this wave: two SKILL.md files with extensive subject-area reference libraries (math, ELA…
Eight accounting workflow skills for month-end close, reconciliation, journal entries, financial statements, variance analysis, and SOX testing.
Finance ships eight skills covering the core accounting close cycle: reconciliation, journal entry prep, close management, financial statements, audit support, …
Ten skills for in-house commercial legal work: NDA review, vendor agreement playbooks, renewal tracking, escalation routing, and SaaS MSA review.
Commercial-legal is the most thoughtfully designed first-party plugin in this wave. Ten skills coordinate a full contract workflow — intake interview, NDA revie…
Solves competition math (IMO, Putnam, USAMO, AIME) with adversarial verification: a second fresh-context agent attacks the proof before the answer ships.
Math-olympiad solves a real failure mode in AI-assisted proofs: self-verification is biased toward agreement because the verifier sees its own reasoning chain. …
Creates Claude Code hooks from natural language — analyze your conversation for past mistakes or describe a pattern you want to block, and hookify writes the rule.
Hookify turns Claude into a meta-tool for managing Claude's own behavior: run /hookify to analyze the current conversation and surface patterns worth preventing…
Drives the QuickDesign CLI to generate UGC promo videos, image edits, and product creatives via Seedance 2.0, Sora 2, Kling, Nano Banana, and GPT Image.
Quickdesign is a high-quality prompt-only plugin for AI media production: 11 cardinal rules covering model selection, reference-image labeling, voice continuity…
Connects Claude Code to the TestDino hosted MCP server for inspecting test runs, debugging failing cases, and managing manual test suites.
Testdino is a minimal MCP router: seven short skills that tell Claude which TestDino tool to call for each QA workflow — list runs, inspect failures, debug hist…
Blockchain accounting plugin for TRES Finance: 20 skills covering wallet onboarding, ledger reconciliation, cost basis, and DeFi transaction analysis.
Tres-finance is the most complex plugin in this wave: 20 skills coordinating a complete blockchain accounting pipeline — wallet import, on-chain data pull, bala…
Compresses Claude Code output using caveman-style speech — saves tokens in chat replies without touching code.
Caveman is a tightly scoped, well-documented token-compression skill with clean hooks, zero runtime network calls, and benchmark numbers that are more carefully…
Enforces YAGNI and minimal code by scanning for over-engineering, dead code, and unnecessary dependencies.
Ponytail earns its place: a well-structured minimal-code plugin with a ladder-of-minimalism prompt backed by real agentic benchmarks, not handwaved claims. The …
Audits and polishes frontend interfaces with 23 design-focused commands and 59 deterministic local detector rules.
Impeccable is the most substantial of the three: 23 commands, a 59-rule deterministic design detector that runs entirely offline without an LLM or API key, prop…