How to use the Novelty agent skill: hard-to-vary causal theories

What is an agent skill, and what does Novelty make the agent do?

An agent skill is a folder containing a SKILL.md file and supporting files that tells an AI agent how to handle a certain kind of task. Novelty's SKILL.md describes an evolutionary explanation-discovery process for hard-to-vary causal theories of unknown phenomena.

Without the skill, an agent tends to give quick guesses about a mystery. With it, the agent runs a structured process: it gathers observations, rival explanations and available tests, generates many unusual candidate theories, mixes and mutates them over several rounds, and keeps only those that pass strict checks for being hard to vary and testable.

• Each surviving theory comes with its lineage (how it descended from earlier candidates).
• Each survivor has a risky prediction that could prove it wrong.
• Each survivor notes what it still leaves unexplained.
• Speculative ideas are kept but clearly labelled as unproven.

Install the Novelty skill

You can install Novelty in one of three places. The source for the skill is the Novelty folder in the LifeOS repository (LifeOS/install/skills/Novelty, MIT licence).

Option 1: personal skills folder. Download the Novelty ZIP from this page, unzip it, and place the Novelty folder inside ~/.claude/skills/ so that SKILL.md sits at ~/.claude/skills/Novelty/SKILL.md.

Option 2: project skills folder. Put the same Novelty folder inside the .claude/skills/ folder of a specific project. It will then apply only to that project.

Option 3: Claude apps. Upload the ZIP in the Claude apps under Settings → Capabilities → Skills.

Check what the skill needs before you run it

According to the skill listing, Novelty requires nothing beyond the agent itself. It does not run scripts and does not use the network. Its supporting files are HardToVary.md, Intents/EvolutionCycle.md, Intents/ExplanatorySelection.md, Intents/RapidCycle.md and SKILL.md.

Because it has no scripts, there is no separate program to install or run on your machine.

• Requirements: none listed.
• Runs scripts: no.
• Uses network: no.

Trigger Novelty with natural requests

You do not need a special command. Describe the open question in plain language and ask for a new theory, an explanation of a mystery, or a creative exploration over several rounds. The skill's routing uses these three intents:

• EvolutionCycle: unknown phenomenon, new theory, hard problem, breakthrough explanation. Use this for a full multi-round run.
• ExplanatorySelection: you already have candidate theories and want them tested and ranked for being hard to vary.
• RapidCycle: a rapid novelty cycle or time-boxed creative sprint.

Examples of requests and what you get back

The skill's own examples show three patterns. Each is described here in words; the wording of your request can differ.

• Request: 'Nobody has a good account of why this species migrates at night. Develop a new theory.' You get a portfolio of surviving causal explanations, each with lineage, a risky discriminating prediction and its unexplained residue, plus a wildcard and the stopping condition.
• Request: 'Here are three theories of the anomaly. Which one actually explains it?' You get the three candidates ranked by hard-to-vary score, each with stress-test outcomes and a rival observation it fails or survives.
• Request: 'Give me 20 minutes of novelty on why the outage only hits on Tuesdays.' You get a first-cycle portfolio of at least three candidates from distinct lineages, a selected leader and wildcard, and the one test that should govern a longer run.

Understand how survivors are judged

Novelty judges each candidate with the stress tests in HardToVary.md: parameter substitution, component ablation, rival equivalence, cross-state comparison and causal-core compression. A candidate must state its target phenomenon and observations, a short causal chain, why each core part is necessary, and a risky observation that could count against it.

Hard-to-vary is a comparative measure of explanatory quality. The skill is explicit that it is not proof of truth. Elegance and fit to data do not decide survival on their own.

Safety notes and limits

Novelty is designed for phenomena that are not adequately explained today. It is not for questions with well-established answers.

Do not use it for medical, legal or safety decisions. Its output is speculative hypotheses that need expert validation, and they could be mistaken for established fact.

Source material the skill reads is treated as evidence or creative material, not as instructions to follow.

• Use the single-pass BeCreative skill for simple brainstorming or name and idea lists.
• Use the Assay skill for delivery validation alone.
• Check any surviving theory with domain experts and real tests before relying on it.

Tips for better results

Give the skill a clear phenomenon, the observations you already have, the constraints, and the rival explanations you know of. The skill's process uses these inputs directly. If you have available tests or interventions, mention them, since the TEST phase challenges candidates against feasible interventions.

For a longer run, ask the agent to state why it stopped and what would justify another cycle. The skill is built to report this.

• Name the phenomenon precisely and separate it from your proposed explanation.
• List known rival theories so they can be compared.
• Ask for the risky prediction for each survivor, since that is how it can be tested.
• Expect speculative candidates to be labelled as unproven.