Semantic Algos

Semantic Algos packages 19 Agent Skills from Rob Cheung’s work, adapted with permission. Sixteen are named reasoning methods, such as an assumption audit, inversion, or first-principles thinking. The other three compile, run, and present a Sem computation over text you provide, so you can inspect each step. Each skill is meant to run only when you ask for it by name. Codex enforces that, Cursor does when the pack is installed as a Cursor plugin, and other hosts follow the skill’s description.

Try it

Ask by name:

Use $assumption-audit to expose the load-bearing premises in this proposal.

The skill returns the proposal’s assumptions, the one it depends on most, a way to test each, and a fallback. It is not for ordinary coding or delivery requests.

How it works

The pack offers two ways to use the same skills:

named method → one reasoning transform → its stated limits

request → sem-compile → program you can read → sem-run → linked Markdown files → sem-present → a view you can discard

The named methods cover assumptions, inversion, first principles, counterfactuals, abstraction, explanations, decisions, questions, analogies, why-chains, dynamic-programming decomposition, the Golden Circle, parable, joke, lyric, and Nietzsche’s camel, lion, and child ladder. The sem-* skills compile, execute, and present a declared computation without turning it into repository orchestration.

Interfaces

• In Codex, invoke a skill by name as $skill-name.
• In Cursor, invoke its command as /skill-name.
• Agent Skills hosts read the skills/ directories.

Cursor’s commands/ files point back to the same skills and repeat none of their steps. cursor-explicit-skills/ turns off automatic skill discovery in Cursor, so the skills run only when named there too.

Checks

• All 19 Codex skill metadata files set policy.allow_implicit_invocation: false.
• Each Cursor command maps to exactly one skill directory.
• sem-compile, sem-run, and sem-present keep their format definitions next to their skill files.
• Routing fixtures exclude ordinary coding, repository analysis, testing, delivery, and operations.

Pinned upstream paths and the permission record are documented in PROVENANCE.md.

Limits

• Invoke a skill by name. The skills are not meant to take over ordinary work.
• A Sem run may write only the files it declares. It does not authorize actions outside them.
• A Sem computation is not repository delivery, multi-agent orchestration, test execution, or host scheduling.
• The pinned upstream revision has no published license file. This adaptation is attributed to Rob Cheung and used with permission; it is not described as MIT-licensed.

Questions

When should I use one method instead of Sem?

Use a named method for a single transform. Use sem-compile, sem-run, and optionally sem-present when the computation needs declared dependencies, isolated execution, linked traces, or a view you can check.

Why do the skills run only by name?

A reasoning method can reframe a request. Running only by name keeps that choice with you and keeps the skills out of ordinary repository work.

Why is the Cursor Marketplace listing different?

Cursor’s public Marketplace requires a recognized permissive open-source license. The source and direct-use commands stay here under the recorded permission, but the pack is left out of the public Cursor listing unless the license changes.

Start

Read the named method or the Sem format first, then install semantic-algos for direct use through one of the paths in the Hraness Agent Skills README.