Ai Research Explore
ai-research-explore is an agent skill that runs conservative, auditable exploration on current_research into candidate-only explore_outpu
npx skills add https://github.com/lllllllama/rigorpilot-skills --skill ai-research-explore| Installs | 54k |
|---|---|
| GitHub stars | ★ 412 |
| Last updated | May 27, 2026 |
| Repository | lllllllama/rigorpilot-skills ↗ |
What it does
You need to explore potentially novel ideas on top of current_research without corrupting trusted code, benchmarks, or verification lanes.
Who is it for?
Solo ML or agent builders using RigorPilot-style lanes who already have current_research and want hypothesis-driven exploration with written audit trails.
Skip if: Quick one-off refactors, production hotfixes, or work when exploratory research was not explicitly authorized on current_research.
When should I use this skill?
Exploratory work has been explicitly authorized on top of current_research and you need candidate-only auditable outputs.
What you get
You get isolated exploratory runs with CHANGESET.md, TOP_RUNS.md, and status.json in explore_outputs, ready for review before any transplant into the trusted lane.
- explore_outputs/CHANGESET.md
- explore_outputs/TOP_RUNS.md
- explore_outputs/status.json
Related skills
Microsoft FoundryDeploy, invoke, evaluate, and troubleshoot Microsoft Foundry agents with Docker, ACR, and observability workflows.400k1.2kAzure AiWire Azure AI services—especially Content Safety in Java—into a solo builder’s backend when you need moderation, blocklists, and SDK patterns without digging through full Microsoft docs.397k1.2kAzure Hosted Copilot SdkDetect @github/copilot-sdk in a repo and scaffold, extend, or deploy Copilot-powered apps on Azure with BYOM support.369k1.2k
Lark EventRun long-lived Lark event consumers that stream IM, VC, and Minutes events as NDJSON for bots and agent subprocesses.244k
Setup Matt Pocock SkillsBootstrap AGENTS.md/CLAUDE.md and docs/agents/ with issue tracker, triage label vocabulary, and domain doc layout before using triage, PRD, TDD, and architecture skills.223k121k
Running Claude Code Via Litellm CopilotPoint Claude Code at a local LiteLLM proxy forwarding to GitHub Copilot Chat API for cheaper runs.220k61How it compares
Use instead of ad-hoc agent rewrites when you need candidate-only exploration with frozen scope—not a general code-review or debugging skill.
FAQ
Who is ai-research-explore for?
Indie and solo builders running agent-coordinated research experiments who must keep exploratory changes auditable and separate from trusted implementation.
When should I use ai-research-explore?
During Idea research to test novelty hypotheses, during Validate prototyping when benchmarks need bounded sweeps, and during Build agent-tooling when coordinating isolated runs on current_research.
Is ai-research-explore safe to install?
Review the Security Audits panel on this Prism page and inspect the skill repo before granting git, shell, or network access for branch worktrees and run exploration.
SKILL.md
display_name: Rigor Explore short_description: Rigor Explore compatible slug for candidate-only current_research exploration into auditable explore_outputs. default_prompt: Use current_research as the explicit exploratory context, coordinate isolated code and run exploration conservatively, treat novelty as a hypothesis until evidence supports it, and write CHANGESET.md TOP_RUNS.md and status.json into explore_outputs. # Research Explore Policy ## Purpose Use this skill only when exploratory work has been explicitly authorized on top of `current_research`. In RigorPilot terms, the goal is meaningful and potentially novel candidate work, not verified novelty. ## Requirements - keep work on an isolated branch or worktree - record `current_research` in a durable form - treat all outputs as candidate-only exploratory records - coordinate code and run exploration conservatively instead of freeform rewriting - keep the trusted lane and exploratory lane clearly separated - keep improvement mining bounded to the frozen task family, dataset, benchmark, evaluation source, and provided SOTA references - require source-backed idea cards before transplant-style implementation planning - keep patch plans minimal, reversible, and auditable - keep research lookup free-first and provider-optional; missing external keys must not block the flow - prefer local curated literature, including Zotero when available, before broader lookup, without requiring a provider - treat `seed_only` lookup records as weak evidence only - distinguish `external_provider`, `parsed_locator`, `repo_local_extracted`, and `seed_only` evidence in downstream ranking and support summaries ## Avoid - implicit experimentation - claiming exploratory gains as trusted reproduction success - claiming novelty, contribution, or SOTA superiority before literature contrast, ablation evidence, and fair comparison - requiring non-bundled skills to complete the workflow - using this skill for narrow code-only or run-only asks - open-ended scientific brainstorming without a frozen campaign anchor - broad multi-module rewrites or metric-surface edits by default - presenting cache-first locator parsing as complete current-literature retrieval # Idea Evaluation Framework `ai-research-explore` uses a bounded, candidate-only evaluation scheme for Rigor Explore idea ranking. ## Hard Gates - baseline gate must not be `abandon` - `single_variable_fit >= 0.6` - `interface_fit >= 0.5` - `patch_surface <= 0.7` - `dependency_drag <= 0.7` - `eval_risk <= 0.6` - `short_run_feasibility != blocked` ## Soft Ranking Positive contributions: - `expected_upside` - `single_variable_fit` - `interface_fit` - `rollback_ease` - `innovation_story_strength` - `source_support_strength` - `execution_feasibility` Negative contributions: - `implementation_risk` - `eval_risk` - `patch_surface` - `dependency_drag` - `execution_cost` - `baseline_distance` ## Provenance Each ranked card should record where each field came from: - campaign input - read-only repo analysis - source lookup cache - source mapping and patch planning - execution feasibility or smoke evidence ## Guardrails - ranking is for candidate prioritization only - ranking does not prove novelty - ranking does not prove benchmark completeness - ranking does not prove verified SOTA superiority # Research Campaign Spec ## Purpose Use `research_campaign.json` or `research_campaign.yaml` when `ai-research-explore` is operating as Rigor Explore: - the task family is already chosen - the dataset is already chosen - the evaluation method is already chosen - the provided SOTA table is already frozen by the researcher - the remaining work is campaign governance, implementation, and candidate filtering This file is an advanced reference. The public entrypoint only requires the campaign core to be frozen; the detailed blocks below are guidance for richer campaigns, not fields the agent must invent on every run. `variant_spec` still exist