Shared communication
Apply TW plain-language and guidance rules to user-facing work. Explain terms in context; preserve the user's chosen Newcomer, Beginner, Proficient, or Veteran level without inferring changes. Newcomer and Beginner levels receive detailed explanations; every level receives necessary explanations and bug explanations unless the user explicitly waives that scope. Soul and prose modes never disable this baseline. Load the shared reference for task entry, level management, or explanation decisions.
Plugin runtime
When this Skill is invoked through the TavernWeave plugin, or the task runs in dots cloud or on a connected computer, read the shared plugin execution contract at task entry. Apply its host, tool and persistence checks before host-specific commands; keep this Skill's scope, authorization and acceptance gates.
Write the information a scene needs, then make sure it can reach the model handling that scene. Do not begin by dividing an encyclopedia into equal-sized entries or adding a controller.
Enter at the requested scope
- Write or refine a body: read writing-contracts.md. Preserve confirmed meaning and source syntax. A prose-only request does not require a world-size interview, runtime matrix, template installation, or export.
- Plan or change entry boundaries/loading: read planning-and-loading.md. Inspect actual entries, configuration, intended scenes and available mechanisms together. Ask Small / Medium / Large only for a new design or material redesign; reuse an explicitly confirmed choice.
- Review an existing book: stay read-only, select the relevant checks from those references, and report observed issues separately from missing material and untested runtime behavior. Do not rebuild the book merely to match a new-card convention.
- Investigate a wording or context-sensitivity failure: optionally use behavioral-validation.md for equivalent-wording, local-definition and position comparisons. Read research-sources.md when the rationale or its limits matter. These are not compulsory entry-level tests and do not authorize model charges or runtime execution.
Before file writes, use $consult-tavernweave-library with this route, read A0 and
the smallest matching guide sections, and retain the route receipt with the existing
project authority. Dated guides are evidence; current authoring policy and verified
target behavior decide which examples apply. Do not copy a snapshot's recursion,
numeric order recipe, model-brand depth, or runtime assumptions into the design.
When the selection/processing/request distinction is not already clear, read the task-relevant part of SillyTavern runtime foundations. Start with its pure-text turn. Optional template, MVU and UI chains are needed only when the project adopts them.
Work from the existing decisions
Apply the shared communication reference's active clarification method: separate the user's goal, hard constraints and suggested means. “Break it down, preferably using a template” may be a proposal; “All entries must be split” is a constraint. Reuse the context, ask only what changes the plan, and continue reversible work where safe. Unknown runtime capabilities are facts to verify, not preferences for the user to guess. Show conditional plans when necessary.
- Start from the existing Content map, creative decisions, source evidence and state contract. Keep the established canon, decision ledger and state store rather than creating parallel copies.
- Identify what the receiving model must decide in representative scenes, the facts needed together, and how they can actually reach that request.
- Choose boundaries and loading jointly. Start with native mechanisms where suitable; use simple conditional output or an existing controller only when it solves a demonstrated need. Partial mixing within one book is allowed.
- Write the body with its semantic invariants, then pair configuration when the task creates or changes an actual entry. Preserve stable IDs, unknown fields, unrelated settings and protected text in local edits.
- Check the changed scope. Separate source checks, import, final-request evidence, model behavior and author acceptance; static tests do not establish runtime success.
Use examples.md for the same bridge requirement through native, simple conditional and controller designs; the port/secret, unreachable-entry, partial-mix and neutral routing cases. The examples are optional and adjustable, not exclusive templates or required chapter/keyword/word-count quotas. The same-name example also separates a concept's local meaning from its human label.
Ownership and handoff
| Need | Owner |
|---|---|
| Worldbuilding, whole-card Content map, greetings, character design and final card responsibility | tavern-card-builder |
| Source extraction and provenance before choosing runtime entries | build-work-library |
| Wording refinement within the protected semantic contract | rewrite-natural-prose |
| E03: target versions, APIs, template stages and exact read behavior | version verification |
| E05: typed targets, aliases, schema/initialization, state scopes and missing values | target/state contract |
| E06: shared content identity, deduplication and final context budget | output budget |
| E07: isolated real-runtime probes and actual request evidence | runtime checks |
| Source-preserving component edits, import/export and release | component update, card pipeline |
This skill owns E01/E02 selection and E04 author-facing precedence. It states required capabilities and expected outputs; it does not verify an API from its name, install extensions, implement a new scheduler, write state, execute an inherited card, or authorize publication. Template selection is a reader, never an incidental state writer. Native recursive activation remains disabled in both directions; effective keyword scan depth remains 2 under the existing TW policy, distinct from insertion depth.
Return the requested prose or the applicable entry plan/body/configuration, reasons, dependencies, unresolved decisions and evidence. Keep short tasks short.