convez.ai
InicialPluginScore · de 0 a 100

Claude Code skill — multi-agent research pipelines. Scout, fan-out, pressure-test, synthesize. One file per run.

Instalação

$npx convez-ai mreza0100/rr

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
1
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 1 mês
Versão
—
Issues abertas
1
Contribuidores
2

README

A tradução para o português deste README ainda está na fila.

RR — Research & Report

Version: 1.3.1 · License: MIT · Repo: github.com/mreza0100/rr

A Claude Code skill for structured, multi-agent research pipelines. Instead of a single search and a one-paragraph answer, RR runs a Workflow pipeline that builds knowledge in batches — each batch shaped by what the previous batch found — and finishes with a synthesized, confidence-rated report and an actionable plan.

Why this exists

When you ask an LLM to "research X," it runs one search, reads the top results, and gives you a summary. That's fine for simple lookups. But for real research — comparing options, auditing a codebase, investigating regulations, evaluating trade-offs — you need a pipeline that:

  1. Maps the landscape first (what are the sub-questions?)
  2. Researches in judge-steered rounds — after every research wave, a judge reads ALL findings so far and decides the next wave's questions; research → judgment → research, never a plan fixed up front
  3. Refuses to stop early — the judge may only declare saturation after a pressure-test round has hunted counter-evidence
  4. Adversarially verifies the load-bearing claims (tries to refute each, flags single-source claims, re-rates confidence)
  5. Synthesizes everything into a verdict + plan (not just raw findings)

RR does this. It delegates the work to a background Workflow so the main conversation stays clean, persists the full research record to a file, and delivers a terse executive summary.

Two modes

ModeTriggerWhat happens
RRRR <topic>Runs the committed Workflow engine (scout → judge-steered rounds → verify → synthesize), delivers a report
RRPRRP <topic>Writes a self-contained prompt you can run in another chat

Three research surfaces

SurfaceTools usedWhen to pick
internetWebSearch, WebFetchExternal topics — libraries, regulations, market research
codebaseRead, Grep, Glob, BashInternal topics — "how is auth wired", "audit our data flow"
bothAll of the aboveMixed — "best practice X and how we currently do it"

RR infers the surface from the topic. If ambiguous, it asks.

Installation

As a Claude Code skill

# From your project root
mkdir -p .claude/skills/rr
cp SKILL.md workflow.js .claude/skills/rr/

Then use it in Claude Code:

RR <topic>              — run a full research pipeline
RRP <topic>             — write a portable research prompt
research <topic>        — alias for RR
look into <topic>       — alias for RR

Pipeline architecture

User: "RR <topic>"
  │
  ├─ Step 1: Refine the goal (what do they actually want?)
  ├─ Step 2: Determine storage path
  ├─ Step 3: Launch the committed engine — Workflow({scriptPath: '.claude/skills/rr/workflow.js'})
  │    ├─ phase Scout:    one agent maps the landscape → 2-6 sub-questions     (opus)
  │    ├─ repeat until saturated or round cap (default 6):
  │    │    ├─ phase Research: wave of researchers (sonnet) ∥ decided-retrieval
  │    │    │                  collectors (haiku) — barrier at end of wave
  │    │    └─ phase Judge:   reads ALL findings, steers the next wave,        (opus)
  │    │                      calls saturation only after a pressure-test round
  │    ├─ phase Verify:   one adversarial skeptic per load-bearing claim       (sonnet)
  │    └─ phase Synthesize: writes the ONE report file, returns exec summary   (opus)
  ├─ Step 4: Verify the file landed (backfill from the structured return if not)
  └─ Step 5: Deliver terse summary to user
       └─ Verdict + key findings + plan + file path

Key design decisions

Delegate to a Workflow, don't inline. Research generates a lot of tool noise. RR runs a background Workflow so the main conversation stays clean and focused.

A committed engine, not an improvised script. workflow.js ships with the skill and is invoked via scriptPath — the loop is enforced by a scheduler, never re-authored per run. A prose protocol asks the model to iterate; a scheduler makes the judge's output literally become the next round's input.

Judge-steered rounds, not one fan-out. Researchers within a wave run in parallel on different sub-questions; the dynamism lives between waves — the judge reads everything found so far, flags contradictions, and writes the next wave's questions. Decided retrievals (exact URL / grep / file) go to cheap collectors. The judge may only saturate after a pressure-test round has hunted counter-evidence.

Tiered models. Judgment (scout, judge, synthesizer) stays on opus; researchers and verifiers ride sonnet; decided-retrieval collectors ride haiku. An optional sacred: true lifts every stage to opus for compliance-grade topics. Collectors return raw excerpts + sources and never conclude — judgment never delegates down.

One file, one run. The pipeline produces exactly one research file — the synthesizer writes it from inside the workflow, and the orchestrator verifies it landed (backfilling from the structured return if not). No intermediates by construction.

Adversarial verify. Each load-bearing claim gets a skeptic that tries to refute it, drops what it can't corroborate, flags single-source claims, and re-rates confidence. The first plausible answer is often a confident-sounding wrong one.

Plan, not just findings. A wall of facts is half the deliverable. Every RR run must produce a concrete, opinionated recommendation.

Updating

Compare the version field in your installed SKILL.md frontmatter against the repo's latest:

cd /path/to/rr-repo && git pull
cp SKILL.md workflow.js /your/project/.claude/skills/rr/

License

MIT

Ver o original no GitHub

Bom para

Relacionados