{
  "title": "Articles/gradientize-task-skill-description",
  "caption": "Gradientize Task",
  "slug": "gradientize-task-skill-description",
  "tags": [
    "article",
    "gradientized-prompting",
    "hermes-published",
    "pack-2",
    "published"
  ],
  "canonical_url": "https://mosiah.org/articles/gradientize-task-skill-description/",
  "interactive_url": "https://mosiah.org/#Articles%2Fgradientize-task-skill-description",
  "markdown_url": "https://mosiah.org/articles/gradientize-task-skill-description.md",
  "json_url": "https://mosiah.org/json/gradientize-task-skill-description.json",
  "fields": {
    "sort-date": "2026-05-12T10:45:00Z",
    "caption": "Gradientize Task",
    "created": "20260512103201589",
    "modified": "20260512103201589",
    "tags": "article hermes-published published gradientized-prompting pack-2",
    "title": "Articles/gradientize-task-skill-description",
    "type": "text/vnd.tiddlywiki"
  },
  "text": "//Related:// [[sources|Article Sources/gradientize-task-skill-description]] · [[notes|Article Notes/gradientize-task-skill-description]] · [[metadata|Article Metadata/gradientize-task-skill-description]] · [[Published Pieces]]\n\n! Gradientize Task\n\n//A skill description for converting long-running agent requests into locally orderable run geometries.//\n\nLong-running agents fail when given checklists, vague goals, or ordinary functional specs. They complete visible steps, satisfy proxies, reward-hack tests, and drift away from the real objective.\n\nGradientize Task converts a user request into a run geometry: a locally orderable optimization surface for agentic work.\n\nIt outputs:\n\n* the real artifact being improved;\n* the ideal state;\n* the value criterion;\n* the invariant set;\n* the homotopy from low-resolution reality to production complexity;\n* verifier functionals;\n* anti-Goodhart constraints;\n* checkpoint, rollback, and escalation policy;\n* stopping conditions.\n\nCore principle: homotopy, not ladder.\n\nDo not create a sequence of disconnected toy tasks. Define one real problem and continuously increase resolution while preserving topology and invariants.\n\nThe skill should reject or flag prompts that are not yet gradientized:\n\n* checklist-shaped prompts;\n* sparse functional specs;\n* fake mocks;\n* staged difficulty ladders;\n* weak verifiers;\n* objectives that reward proxy wins;\n* tasks where local progress is not orderable.\n\nThe final output is not a plan. A plan may be generated later, but only after the run geometry is defined.\n\nA plan says: walk this path.\n\nA run geometry says: here is how to know uphill.\n"
}