BootLoops 1.0 is a harness for large language models doing precision quantitative science. It has two halves. The computing packages, ports and engine upgrades are the toolkit, which lives in its own repository, bootloops, published beside this one by the same organization. This repository, skills, is the other half: the working protocols of the program, written as standalone instruction files an LLM agent can load and follow, plus their packaging as Claude Code and Codex plugins. The protocols are as much a part of the harness as the packages: the code supplies capability, these supply the discipline that makes its output trustworthy.

The skills are plain markdown in the Agent Skills format (a SKILL.md with frontmatter: name, when to use, and a self-contained body), the format that Claude Code and a number of other agent tools read. None depends on any private infrastructure, and they are written so that any capable agent can follow them. They encode what "done" means: a result reproduces an independent route at points no fit ever saw, with a positive control proving the check can fail; integer-relation discipline with the constant ring declared before the search and refusal over invention when no relation is found; planted-truth controls that recover a known answer before any real data is touched; provenance bookkeeping so no oracle that fed a fit ever certifies the result; timing discipline (measure a small run before a big one); and the heavier protocols for proving with agents, simulating referees, auditing the literature behind a novelty claim, verifying bibliographies, and linting scientific prose for hype and honesty failures.

Twelve skills in two groups; skills/README.md has one line per skill on what each governs.

• The protocol layer (general purpose; load these for any quantitative work): acceptance-gate, constant-recognition, planted-truth, independence-bookkeeping, timing-discipline, reading-contract, tool-stewardship.
• The research skills (heavier machinery for specific kinds of work): prove-protocol, referee-sim, lit-review, ref-check, prose-lint.

The two groups are also the two plugins, bootloops-protocols and bootloops-research.

These protocols make agent-produced results checkable; they do not replace review by a qualified person. The verdict labels used in the skills (CLOSED, VERIFIED-CLOSED) are internal grades; every BootLoops result was reviewed by the author before release.

• skills/: the library, one directory per skill. This is the single source of truth; every other copy in the repository is generated from it.
• .claude/skills/bootloops-setup/: the installer skill, the only skill that is active when you open Claude Code inside a clone. It shows the roster, asks which skills you want and at what scope, and copies only your selection into place.
• plugins/, .claude-plugin/marketplace.json, .agents/plugins/marketplace.json: the plugin packaging (two plugins and the marketplace manifests Claude Code and Codex read). Generated by tools/make_plugins.py and never edited by hand.
• tools/make_plugins.py: the generator.

Nothing activates without your say-so on any route. Pick the route that matches your setup. The commands below name this repository, github.com/BootLoops-ai/skills; a copy of it published under another organization regenerates its plugin manifests with python3 tools/make_plugins.py --force --slug ORG/skills.

git clone https://github.com/BootLoops-ai/skills
cd skills

No GitHub account is needed: public repositories clone anonymously, and the same tree downloads as a plain tarball from the repository page. The skills arrive inert. Open Claude Code in the clone and run

/bootloops-setup

to pick which skills to activate and at what scope (project-level .claude/skills/, or user-level ~/.claude/skills/ for every project). Deactivate any skill later by deleting its directory from the install location. In any other agent, point it at a skill file directly ("read skills/acceptance-gate/SKILL.md and hold my result to it") or copy the chosen skill folders into your agent's skills directory.

The skills are packaged as two plugins, bootloops-protocols (the seven general disciplines) and bootloops-research (proving, referee audit, literature review, reference checking, prose linting), so you install exactly the slice you want:

/plugin marketplace add BootLoops-ai/skills
/plugin install bootloops-protocols@bootloops
/plugin install bootloops-research@bootloops

Updates arrive through the plugin system; disable or uninstall per plugin any time with /plugin. Plugin systems enable whole plugins, not single skills; for per-skill choice use the clone route.

The same two plugins ship in Codex's plugin format from the same repository:

codex plugin marketplace add BootLoops-ai/skills

then open /plugins in the Codex CLI (or the Plugins tab of the Codex app), choose the bootloops marketplace, and install bootloops-protocols, bootloops-research, or both. Installed skills are matched to tasks automatically from the next session on.

skills.sh installs Agent Skills from a GitHub repository into the skill directories of whichever agents you use (Claude Code, Codex, Cursor, Copilot, and others), with per-skill selection:

npx skills add BootLoops-ai/skills

Add --list to see the roster first, or --skill acceptance-gate to take one.

Editors and agents that read the Agent Skills format (agentskills.io) can load the skills/ directory directly; see your tool's documentation.

The skills also run in web assistants that accept custom skills: claude.ai accepts a zipped skill directory under Settings > Skills (see the claude.ai help article on using custom skills).

plugins/ and the two marketplace files are a pure function of skills/. After editing a skill:

python3 tools/make_plugins.py --force     # rewrite plugins/ and the marketplace files
python3 tools/make_plugins.py --check     # exit 0 when the committed tree is current
claude plugin validate .                  # optional: Claude Code's own schema check

Without --force the script refuses to touch an existing generated tree. The plugin names, descriptions, version and repository slug are constants at the top of the script; change them there, never in the generated files. The slug (ORG/REPO, read by /plugin marketplace add and written into the manifests) defaults to the organization currently hosting this repository; published copies under another organization regenerate with --force --slug ORG/skills and verify with --check --slug ....

Several skills (tool-stewardship above all) speak of "the toolkit": the computing packages an agent should consult before writing new code. That is the bootloops repository published beside this one, with its tools/README.md index and per-package guides. The skills work without it; they are discipline, not software. The papers, the result pages and the per-problem code live at bootloops.ai.

Maintenance. This repository is maintained by Matthew D. Schwartz, not by Anthropic. It is not an officially supported Anthropic product, and Anthropic does not provide support, updates or fixes for it.

Reporting issues. Please report bugs and security problems through this repository's GitHub issues.

Security considerations. Treat input files from others as code. These are research tools meant to be run locally on inputs you trust. Many of them evaluate the contents of their input files (JSON, YAML, .m, .ms, .jl, pickle and similar), so a file received from someone else can run arbitrary commands on your machine. Only run files you wrote yourself or got from a source you trust, or run them in a sandbox or container. The integrity checks and certificates in this repository guard against accidents. They are not a security boundary.

BootLoops skills 1.0. Copyright (c) 2026 Anthropic, PBC. Created by Matthew D. Schwartz; written by Claude (Anthropic) under his supervision. This is not an officially supported Anthropic product; it is maintained by Matthew D. Schwartz (https://www.bootloops.ai).

The skill texts and the other prose in this repository are released under CC BY 4.0 (LICENSE-CONTENT); the scripts and plugin manifests are released under the MIT License (LICENSE). NOTICE has the attribution. The repository bundles no third-party components.

Formats and tooling. The skills follow the Agent Skills format, originally developed by Anthropic and released as an open standard (agentskills.io); the plugin packaging follows Anthropic's Claude Code plugin and marketplace specification, OpenAI's Codex plugin specification, and the Agent Plugins 1.0 portable manifest schema (agent-plugins.org). The npx skills installer route is the skills.sh directory maintained by Vercel. We thank the maintainers of all four; none of them is affiliated with or endorses this repository.

Acknowledgments. Each skill ends with a short "Sources and acknowledgments" note naming the prior work it codifies; the protocols are assemblies of practices that physics, statistics, software engineering and the systematic-review community worked out long before us, and we have tried to say so where it matters.

To cite BootLoops: M. D. Schwartz, BootLoops 1.0 (2026), bootloops.ai.