Agent Skills

Use InkOS as a Pi-agent creative harness for novels, short fiction, scripts, storyboards, interactive films, open worlds, derivative works, and long-document translation.

Install

npx skills add https://github.com/narcooo/inkos --skill inkos
SKILL.md

InkOS Creative Harness

InkOS is a text-creation harness. The main Pi agent understands the request, activates professional Skills, invokes typed capability actions, and reports only persisted results.

Core contract

  • Discussion stays discussion. Creation and mutation happen only through capability actions.
  • A successful action is proven by its ActionResult and artifact revision, never by assistant prose.
  • User intent, current Work constraints, established facts, and explicit prohibitions outrank model defaults.
  • Skills provide professional methods but never grant write permission.
  • Review returns evidence-backed observations. It does not score prose, reject a chapter, or authorize an automatic rewrite.
  • Revision is an explicit action. Preserve source revisions and report the new artifact revision.
  • Context is assembled per task: LLM semantic selection chooses relevant source sections after deterministic candidate discovery. Protected intent and selected current facts are never compressed; lower-priority context is semantically compiled only on budget overflow.

Creation profiles

  • longform-novel: foundation, chapter planning, writing, review, revision, state projection, forecast, cover.
  • short-fiction: complete short story, plan, review notes, sales package, cover.
  • script: screenplay, vertical short drama, audio drama, interactive script.
  • storyboard: editable storyboard, typed image-prompt assets, image generation.
  • interactive-film: story tree, flags, script, storyboard, image assets, playable graph, export.
  • play-world: open-world or guided interaction with persistent entities, relationships, time, evidence, inventory, scenes, and images.
  • fanfic, spinoff, imitation, continuation: derivative creation with explicit lineage and source canon.
  • translation: multilingual long-document translation with terminology and resumable artifacts.

Natural-language use

Use normal conversation to discuss or request work. The agent chooses the relevant profile and Skill. A creation proposal may require confirmation; recoverable edits inside the current Work can run directly.

Force a Skill when needed:

@inkos-story-review review chapter 8 against the current reader promise
@inkos-imitation-writing keep the reference cadence but create a new story

Inspect available Skills and Works through the agent tools rather than guessing storage paths.

Explicit CLI actions

inkos studio
inkos tui
inkos book create --title "My Story" --genre other --lang en
inkos write next <book-id> --count 3
inkos revise <book-id> <chapter> --mode rewrite
inkos review <book-id>
inkos short run --title "A Short Story"
inkos fanfic init --title "Derived Story" --from source.txt --mode canon
inkos export <book-id> --format epub

Internal planning, semantic context selection, writing, state projection, and review are one Harness trajectory. review displays persisted qualitative observations.

Revision modes

  • spot-fix: exact local text patches applied by the host.
  • polish: prose-surface changes only.
  • rewrite: rewrite the affected passage or chapter while preserving authority.
  • rework: restructure scenes when the user explicitly requests structural change.
  • anti-detect: explicit prose cleanup using the activated de-slop Skill.

Context and persistence

  • Human-readable story assets remain Markdown.
  • Host-consumed state and action results use typed JSON/tool schemas.
  • Planner, writer, reviewer, reviser, settler, Play, translation, and production-document compilers submit typed results through Pi tools.
  • Conversation compaction and story-context compaction are separate. Both preserve protected current intent and use semantic compilation only after budget overflow.
  • Work manifests, artifact revisions, episode events, and model-call traces provide observable lineage.

Failure handling

  • Model/provider failures remain operational errors; they are never converted into fictional world outcomes or content verdicts.
  • A failed optional asset is reported as an observation without invalidating a successfully persisted manuscript.
  • Interrupted actions preserve the last committed artifact and can be retried from persisted state.
  • Never claim a file was created or changed without an artifact reference from the action result.

Search skills and MCP servers

Fuzzy search across 23,137 skills and servers