Recursive Research
Turns Claude Code into a self-regulated research loop that pushes any topic to PhD level, with source tiering, disk checkpoints, and auditable autonomous decisions.
Evidence shows a least-privilege-leaning design: Phase 0 interrogates the user and waits for confirmation; writes are confined to a memoria/ subdirectory of the working directory with an explanation when created; local paths are user-supplied; no exfiltration, credential access, or destructive defaults; --resume provides a recovery path. Deductions: reading arbitrary user-supplied local paths lacks explicit scoping or sensitive-file warnings; rollback beyond the checkpoint directory is not addressed despite long autonomous decision chains.
Instructions are self-consistent, phased, with designed failure feedback (--resume errors clearly if the slug is missing; explains directory creation). Deductions: static review with zero execution evidence, no test suite, no CI; quantitative gates (coverage ≥80%, saturation ≤5%, hallazgos_esperados_proxy) rest on loose, model-estimated proxies that undermine reproducibility; MCP-detection degradation paths are thin.
Clear audience and scenarios (deep cross-domain research), declared boundaries via the 5-criterion PhD gate, precise semantic triggers (/recursive-research, --resume, --list). Deductions: skill body is Spanish-only, README English, with no declared Chinese support; core retrieval paths (Firecrawl, Google Scholar, JSTOR, etc.) pose mainland-China reachability risk that the docs do not disclose.
Well-layered docs: structured SKILL.md, example dialogues, domain examples, anti-pattern list; MIT license; author and repository clear; version present (2.2.0). Deductions: body heading says v2.0 vs frontmatter 2.2.0 inconsistency; no changelog; no explicit maintenance responsibility or update path; the 'pending approval' Plugin Directory claim is unverified author assertion.
Claims structured deliverables (synthesis, actions, gaps) with marginal value over manual research (auto tiering, checkpointing, self-interrogation). Deductions: no static evidence that representative outputs are directly usable; the 'PhD level' claim depends on vague model-estimated proxies; quality and cost/benefit proportionality are essentially unverifiable from source alone.
Auditable primary material exists: a single SKILL.md, cross-checkable license and authorship, and per-criterion source-tiering rules that can be inspected. Deductions: no third-party execution evidence, no tests, no CI artifacts, no reproducible user cases; claims like pending official directory approval cannot be corroborated; only one evidence type available.
- Static review only — no code was executed; loop convergence, checkpoint resume, and PhD gating are unverified at runtime.
- The 'PhD level' claim is gated by loose, model-estimated proxies and should not be equated with genuine academic depth.
- Skill body is Spanish; core retrieval depends on overseas services (Firecrawl, Google Scholar, JSTOR), which may be unreachable from mainland-China networks.
- The skill can read arbitrary user-supplied local paths; avoid pointing it at directories containing sensitive files.
- Coverage and saturation metrics are model-estimated, so two runs on the same seed may diverge significantly.
- Publisher is not verified by the FollowSkills enterprise registry; treat identity as unknown.
What does this skill do, and when should you use it?
recursive-research is a Claude Code plugin and skill (v2.2.0, MIT) by Joseph Huayhualla. Given a research seed topic, it interrogates the user for mode (web/local/mixed), source preferences, and a cycle cap, then selects 3-5 seed threads and iterates until a five-criterion "PhD fitness function" is met. Every source is tiered (Tier 1/2/3/Rejected) into auditable files, and every cycle writes a disk checkpoint that survives context compaction via a --resume mode. All non-trivial autonomous decisions run a Weighted Decision Matrix plus Munger inversion, keeping the reasoning transparent and self-critical.
Asks six setup questions at invocation (seed, mode, local paths, priority/excluded sources, cycle cap); creates memoria/investigaciones/<slug>/ in the working directory with estado.md, hilos.md, per-tier source files, and hallazgos.md; each cycle picks the least-covered thread, scores and selects sources via WDM, fetches content through Firecrawl/Context7 MCPs or built-in WebSearch/WebFetch; tiers each consulted source with justification; writes ciclo-N.md checkpoints and updates saturation metrics; recommends a preventive pause near context limits (~150 tool calls or ~80k output tokens); and on closure produces sintesis.md, acciones.md, and gaps.md.
- A graduate student or researcher needs a systematic map of an unfamiliar field's state of the art and knowledge gaps before a proposal or literature review
- A technical writer or analyst must assemble auditable, tier-graded material for a report, paper, or study
- A business decision-maker wants to get up to speed on a new domain (e.g., SaaS pricing, AI regulation) before committing
- A user with a local library of PDFs/papers wants local files merged into one recursive research run alongside web sources
- A Charlie Munger-style thinker wants every autonomous step to carry a weighted matrix and inversion critique in writing
What are this skill's strengths and limitations?
- Explicit, auditable source tiering (Tier 1/2/3/Rejected) logged to disk
- Per-cycle disk checkpoints plus --resume/--list modes survive compaction and session loss
- Domain-agnostic, with pre-loaded seed source lists (arXiv, PubMed, JSTOR, World Bank, etc.)
- Concrete rejection criteria: no author, data-less marketing, SEO spam, unsupervised AI content
- MIT licensed, collects no data (PRIVACY.md), runs fully locally
- The five "PhD" criteria (≥80% coverage, ≥3 Tier-1 sources per thread, etc.) are self-measured and self-reported, with coverage based on proxy estimates and no external validation
- Advertised speed (10-50x faster scraping) depends on optional MCPs; fallback is slower
- No test suite or independent evaluation is shown in the repo; performance claims are author-stated
- Token-hungry: up to 20 cycles by default, with no cost estimate provided
- PDF support is limited — paywalled-PDF handling is a roadmap item, not implemented
How do you install this skill?
Option A (recommended, as plugin): inside Claude Code run /plugin marketplace add Anjos2/recursive-research, then /plugin install recursive-research. Option B: one-liner via Anthropic's central Plugin Directory (source notes approval is pending). Option C (standalone skill): git clone https://github.com/Anjos2/recursive-research.git, then copy plugins/recursive-research/skills/recursive-research/SKILL.md into ~/.claude/skills/recursive-research/ (on Windows use PowerShell New-Item and Copy-Item as shown in the README).
How do you use this skill?
Invoke as /recursive-research:recursive-research when installed as a plugin, or /recursive-research as a standalone skill. The skill asks for the research seed, mode (web/local/mixed), local paths, priority/excluded sources, and a cycle cap (default 20), then runs autonomously. Resume paused work with /recursive-research --resume <slug>; list saved research with /recursive-research --list. Installing Firecrawl and Context7 MCPs is optional but recommended for speed; without them the skill falls back to built-in WebSearch/WebFetch.
How does this skill compare with similar options?
Unlike a single-shot research answer from an LLM, this skill adds a self-regulated loop, tiered and logged sources, checkpoint/resume, and mandatory documentation of what is unknown (gaps.md). The repo names no direct competitor; it explicitly contrasts itself against anti-patterns like flat synonym-swapping searches, unvetted sources, and missing checkpoints.