본문으로 건너뛰기
cccskills
무료GitHub에 공개

geo

Design, implement, or evaluate GEODE's GEO capability with source-grounded stage diagnosis, frozen measurement contracts, and no-model preflight before live engine runs.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

설치 방법 보기

포함된 파일(1)

  • SKILL.md4.8 KB

SKILL.md(원문)

설치하기 전에 에이전트에게 전달되는 지시 내용을 확인할 수 있습니다.

GEODE GEO scaffold

Use this scaffold for repository work on /geo, generative discovery, AI-search visibility, citations, or the GEO benchmark.

  1. Read .geode/skills/geo/SKILL.md for the runtime behavior contract.
  2. Read docs/eval/geo-visibility.md before defining a metric or run. Freeze a prospective run spec before any score-bearing commercial-engine call.
  3. Read docs/plans/2026-08-22-slash-control-enforcement.md for typed state, then docs/plans/2026-08-20-slash-goal-geo.md before changing slash routing, planning authority, or Goal coupling.
  4. Diagnose the exact stage and preserve its denominator: F fetch/index eligibility, R retrieval inclusion/rank, C citation selection, P visible placement/rank, A answer absorption, Q support/credibility/answer quality, and O first-party outcome. Preserve not_measured; never collapse the vector into one GEO score.
  5. Run deterministic site preflight and relevant unit/E2E checks before asking for live-test approval. Never infer organic lift from a fixed-context test. Model tool calls cannot mint approval or preregistration receipts; exercise the operator-owned /geo approve-live <receipt> and /geo preregister <receipt> boundaries in state-machine tests. Require phase-bound evidence before each transition. Permit diagnostic completion from live_observe; require experiment only for a preregistered treatment claim.
  6. Confirm .geode/skills/geo/SKILL.md is the runtime source selected by SkillLoader; use get_geo/update_geo as the state authority. Record primary sources, facts, inferences, unobserved stages, repetitions, locale, account state, and timestamps in the result.

Long-running execution

For broad or explicitly long-running work, represent progress as a verifiable evidence frontier: questions, dependencies, source locators, contradictions, failed attempts, uncertainty, and unvisited leads. Do not store or expose hidden chain-of-thought.

  1. Keep deterministic preflight, prerequisites, synthesis, verification, and every write in the parent AgenticLoop.
  2. Send at most three independent read-only branches through the existing delegate_task batch. Use durable spawn_agent only when a branch must be steered across waits. Keep prerequisite branches serial and children depth one. Built-in role workers inherit the parent's active model and credential source unless an explicit binding is part of the run contract.
  3. Join all branches before synthesis. Bound each result to claims, evidence locators, contradictions, confidence, and unresolved gaps; do not copy raw logs into the parent context.
  4. At each wave boundary, checkpoint evidence, uncertainty, failures, and unvisited leads. Treat the checkpoint as revisable, not as authority.
  5. Permit one focused follow-up wave for a material gap or contradiction. Stop when required coverage, new primary evidence, and contradiction resolution cease improving.
  6. Gate URL reachability, file facts, counts, hashes, citation locators, and benchmark invariants with deterministic checks. Same-model critique is not a truth oracle.

Reuse SkillRegistry → AgenticLoop plus the existing collaboration tools. Do not add a GEO-only executor, natural-language subcommand parser, coordinator, generic Tree-of-Thought/MCTS engine, or TaskGraph coupling without benchmarked need.

For execution-strategy changes, compare the frozen A/B/C profiles in docs/eval/geo-visibility.md with repeated runs; a single smoke cannot establish that longer or parallel execution is better.

Trajectory and publication

Keep the native result, verifier receipt, attempt lineage, normalized behavior, and publication decision as separate authorities. Do not manufacture an Inspect .eval archive from a GEODE slash transcript: bind .eval only when Inspect produced it. For ordinary /geo runs, export the parent and every child from canonical sessions.db as geode.trajectory@1 with digested content.

Treat terminal transcripts, messages, evidence JSONL, worker results, SQLite/WAL, profiles, usage, diagnostics, provider bodies, and hidden reasoning as withheld-private by default. After an exact-byte privacy review, stage only scope-complete normalized trajectories with geode session stage-trajectory-release; use the explicit --allow-replay-incomplete gate for reviewed content-digested releases. Copy only the allowlisted release and reviewed aggregate sidecars into a fresh geode-eval-artifacts PR, then bind the remote merge commit and manifest digest back into the GEODE eval ledger after independent read-back. Never recursively mirror a runtime home or artifact root.

리뷰

아직 리뷰가 없습니다. 사용해 본 소감을 남겨 주세요.

댓글

로그인 없이 쓸 수 있습니다. 사용 후기, 질문, 작동하지 않은 환경 등을 편하게 남겨 주세요.

아직 댓글이 없습니다. 첫 댓글을 남겨 보세요.

보내면 바로 공개됩니다. 이메일 주소나 전화번호 같은 개인 정보는 쓰지 마세요. URL은 2개까지입니다.

0 / 1,000자

이 양식은 reCAPTCHA로 보호되며 Google 개인정보처리방침 및 서비스 약관이 적용됩니다.

같은 저장소의 스킬

Audit agent runtimes and their tests for lifecycle inference, capability overload, context pollution, correlated verification, prompt-only control, false success, dead compatibility paths, test-only seams, and complexity without contribution. Use for whole-repository agent audits, slop or dead-code triage, test contribution reviews, large refactors, tool/context simplification, and cleanup requests that need proof before deleting code or tests.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

mangowhoiscloud/geode152026년 10월 19일 업데이트

Diagnose agent operational failures through default-policy checks, multi-gap analysis, explicit service ownership, request-context lifetime, and degradation-versus-correctness checks.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

mangowhoiscloud/geode152026년 10월 19일 업데이트

Align, run, audit, and publish GEODE benchmark comparisons against Agent-World arXiv 2604.18292v1 across MCP-Mark, BFCL V4, and tau2-bench. Use for Agent-World or AgentWorld comparison requests, cross-suite agent benchmark design, Tau2 comparator/profile diagnosis, paired thin-runtime-versus-GEODE experiments, mean-accuracy@8 replication, or benchmark comparability and artifact reviews.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

mangowhoiscloud/geode152026년 10월 19일 업데이트

Verification checklist to prevent fake success. Detects test deletion/disabling, coverage regression, lint bypass, secret exposure. Triggered by "deception", "fake" ("가짜"), "fake success", "verification" ("검증"), "checklist" ("체크리스트"), "deletion detection" ("삭제 탐지"), "regression" keywords.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

mangowhoiscloud/geode152026년 10월 19일 업데이트

Implement proven backend architecture patterns including Clean Architecture, Hexagonal Architecture, and Domain-Driven Design. Use when architecting complex backend systems or refactoring existing applications for better maintainability.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

mangowhoiscloud/geode152026년 10월 19일 업데이트

Auto-search and summarize latest AI/agent papers.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

mangowhoiscloud/geode152026년 10월 19일 업데이트

mangowhoiscloud의 스킬 모두 보기

이 스킬의 문제 신고