Pantheon

Pantheon

Pantheon is a bounded evolution system for Codex skills. Treat each skill as an uploaded working memory: not a personality to imitate, but a tested procedure that can survive outside the original conversation.

Core Rule

Never silently mutate installed skills. Propose, patch, audit, test, and ask for confirmation before installing or replacing a skill.

Workflow

1. Observe the mortal task

   • Identify the repeated workflow, brittle decision, tool ritual, or domain knowledge that should outlive the current conversation.
   • Prefer concrete examples over broad ambition.
   • If the request is vague, create the smallest useful skill and list the assumptions.

2. Name the god

   • Use a lowercase hyphenated machine name under 64 characters.
   • Use a human-facing display name only in agents/openai.yaml.
   • Avoid names that describe vibes only; the trigger description must say what work the skill performs.

3. Choose the body

   • Put essential procedure in SKILL.md.
   • Put detailed policies, rubrics, schemas, and examples in references/.
   • Put deterministic repeatable operations in scripts/.
   • Put templates, icons, or reusable output assets in assets/.

4. Forge the skill

   • Use scripts/pantheon.py scaffold when a draft is enough.
   • Manually refine the generated SKILL.md; Codex is expected to improve the draft.
   • Keep the body concise and operational. Remove lore that does not change behavior.

5. Judge the skill

   • Run scripts/pantheon.py audit <skill-dir>.
   • If the skill is nontrivial, run scripts/pantheon.py experiment --case <case-file>.
   • Fix failures before suggesting installation.

6. Evolve with consent

   • For an existing skill, produce a short change plan before editing.
   • Preserve working behavior unless the user explicitly wants a redesign.
   • Keep a clear diff and verification result.

Deity Roles

Use these roles as a thinking frame, not as extra agents unless the user explicitly asks for subagents.

• Archivist: Extract durable knowledge from a messy task.
• Smith: Convert knowledge into SKILL.md, scripts, references, and assets.
• Oracle: Write trigger descriptions and default prompts that invoke the skill at the right time.
• Judge: Audit for overreach, missing validation, context bloat, unsafe autonomy, and untestable claims.
• Ember: Preserve the romantic purpose: the skill should feel like a memory made useful, not a generic template.

Commands

From the Pantheon skill directory:

python3 scripts/pantheon.py distill --input path/to/notes.md
python3 scripts/pantheon.py scaffold --brief path/to/brief.md --out /tmp/skills
python3 scripts/pantheon.py evolve --brief path/to/brief.md --report reports/evolution.json --svg assets/evolution.svg
python3 scripts/pantheon.py audit path/to/skill
python3 scripts/pantheon.py experiment --case experiments/skill-forge-basic.md --workdir /tmp/pantheon-exp
python3 scripts/pantheon.py benchmark --dataset experiments/pantheon-benchmark.jsonl --workdir /tmp/pantheon-bench
python3 scripts/pantheon.py download-dataset --url https://example.com/skills.jsonl --out /tmp/skills.jsonl
python3 scripts/pantheon.py benchmark-public --name alpaca --limit 50 --report reports/alpaca-50.json
python3 scripts/pantheon.py benchmark-public --name awesome-chatgpt-prompts --limit 50 --report reports/prompts-50.json

Read references/evolution-protocol.md before evolving an installed skill. Read references/experiment-rubric.md before claiming the skill has been proven useful. Read references/language-policy.md when the user asks for Chinese, bilingual, localized, or non-English skill output.

Experiments

Use the built-in benchmark for a fast proof loop. It compares a naive baseline generator against Pantheon generation on JSONL briefs and scores trigger clarity, workflow leverage, resource fit, validation integrity, and bounded autonomy.

Use evolve when the user wants the Pantheon core loop: fork multiple skill variants, score them, select winners, merge traits, and preserve a lineage report.

Use benchmark-public for public dataset samples. Current adapters support Stanford Alpaca and awesome-chatgpt-prompts; both are transformed into skill-creation briefs before scoring.

Treat benchmark scores as engineering evidence, not truth. Include qualitative notes from the generated artifacts before claiming impact.

Quality Bar

A Pantheon-created skill is ready only when:

• Its frontmatter description contains the actual trigger conditions.
• Its workflow is shorter than the task it replaces.
• Its reusable resources are justified by repeat use or deterministic reliability.
• Its validation checks can fail for real defects.
• Its autonomy is bounded by user review before installation or destructive edits.
• Its language carries the project's myth without obscuring the engineering contract.