PhyloCrates
Prototype v0.6 GitHub Viviana Romero Alarcón

Docs & install

PhyloCrates ships as a small, fixed set of Markdown and schema files: one behavior spec, eight reasoning modules, and an output contract. Every platform wrapper below loads the same files, nothing platform-specific changes the protocol or the scope boundary.

What's in the package

The canonical source lives in phylocrates-plugin/skills/phylocrates/. This is what each file is for:

SKILL.md # entry point: purpose, protocol, scope boundary PHYLOCRATES_SPEC.md # canonical scientific specification, all 8 stages instructions.md # full behavioral + scope rules for the agent knowledge/ alternative-explanations.md # building and challenging competing hypotheses assumptions-taxonomy.md # the assumption categories Stage 3 checks hypothesis-quality-checks.md # completion criteria for a refined hypothesis hypothesis-record-template.md # the Stage 8 output template hypothesis-type-router.md # the 6 conceptual question types language-and-causality.md # association vs. mechanism vs. causation prediction-reasoning.md # turning explanations into biological predictions scope-of-claim.md # bounding level, time, taxon, claim type output/ HYPOTHESIS_RECORD_SCHEMA.md # the record's field-by-field contract hypothesis-record.schema.json # machine-readable version of the same schema example-record.yaml # one worked example record sources/ source-policy.md # how sources may and may not be used bibliography.md, coverage-audit.md, reasoning-source-map.md

No file in this package specifies a phylogenetic comparative method, a sample size, or an analysis pipeline, the scope boundary lives in the files themselves, not just in what the agent says.

Install & use

Same protocol and scope rules on every platform below; only the packaging differs.

Claude Code

  1. Point Claude Code at the plugin folder:claude --plugin-dir ./phylocrates-plugin
  2. Claude Code discovers two skills automatically: phylocrates (the agent) and phylocrates-check (a dev-only package linter).
  3. Start a hypothesis-refinement session by describing your observation; the phylocrates skill invokes itself, or ask for it by name.

ChatGPT

  1. ChatGPT → Explore GPTs → Create (or an existing Project's instructions + files).
  2. Paste chatgpt/instructions.md into the Instructions field (7,200 characters, under the 8,000 limit).
  3. Under Knowledge, upload every file in chatgpt/knowledge/, flat, no folders.
  4. Leave Code Interpreter and Browsing off; file search turns on by default once files are attached.

Gemini

  1. Gemini app → Explore Gems → New Gem.
  2. Paste gemini/instructions.md into the Instructions field.
  3. Under Knowledge, add every file in gemini/knowledge/, the same 16 files as the ChatGPT package.
  4. Save. No other tools needed.

Gemini CLI

  1. Install the extension:gemini extensions install /path/to/phylocrates-plugin/gemini-cli
  2. GEMINI.md loads automatically as context whenever the extension is active.
  3. Verify it's there:gemini extensions listit should show phylocrates.

Every wrapper is generic on purpose: none of them hardcode a repository URL yet. Point them at wherever this package ends up living.

How a session works

PhyloCrates asks one focused question at a time, in this order: elicit → challenge → supplement → synthesize. It does not anchor you to its own preferred explanation before asking for yours. If you already arrive with a precise question, plausible alternatives, and distinct predictions, it compresses the protocol and confirms rather than re-deriving what you've already worked out.

A session ends when the hypothesis is explicit and coherent: the problem is clear, the question is precise, assumptions are visible, alternatives are real, predictions follow from each one, at least one expectation discriminates between them (or that failure is acknowledged), the claim is bounded, and unresolved ambiguity is recorded rather than hidden. At that point it produces the Hypothesis Reasoning Record.