fable-domain

The fable-method ships domain adapters that translate its loop into a sector's nouns. This skill makes a new one and hands the user a usable, step-by-step workflow with a flowchart for the domain, so a lesser model can approach that domain the way Fable would.

Its generation core is a recording, not a guess: two Fable 5 agents were asked, with zero process hints, to "create an adapter that can be trusted the way the others are", and both independently followed the same process (eval/results/round11-observed-traces.json). Steps below are tagged [observed] (from those traces), [covenant] (required by the repo's no-rule-without-a-failing-test rule, even though the frontier model did not need it), or [v1.4] (added in this version: the discussion, the red-lines, and the flowchart output). The reason the covenant and v1.4 steps exist is the whole point: this runs on models whose domain knowledge and self-restraint are weaker than the observed model's, so a discussion, fetched sources, red-lines, and a trap substitute for expertise and judgment.

What it produces (the bundle; all four, or not done)

A domain workflow with a flowchart [v1.4]. The step-by-step approach for this domain, distilled from the discussion and research, plus a mermaid flowchart, the same shape as this method's own references/flowcharts.md. This is the user-facing "here are the steps, in order" artifact. It lives in the adapter's Workflow section (see TEMPLATE.md).
The adapter, conforming to references/domains/TEMPLATE.md, every named regulation/policy/figure carrying a fetched source in its Sources section.
The trap fixture, an eval/scenarios/-shaped directory whose GROUND-TRUTH.md defines the task, the trap (the sector's central fraud), scoring caps, and ideal behavior.
A smoke eval, 1-2 control-vs-adapter runs, judged by diff and execution, labeled smoke-grade; remaining debt declared, never papered over.

Stage 1: Discuss [v1.4]

Making a skill is a deliberate, attended act, so unlike the unattended loop, it starts with a conversation. Ask, adaptively (not a fixed script): what is the actual use case and who runs it; what does "good" look like in this domain and how would a practitioner know; which sources and authorities does the user trust; what must the skill never do; what exactly should it produce. Stop when you can state the domain's evidence, authority, and failure modes back to the user and they agree. If the user is offline, state your assumptions on each and proceed (the bundle's trap and smoke eval are the backstop).

Red-lines (a hard refusal, checked during the discussion). If the domain requires professional licensure or a wrong answer causes physical, legal, or financial harm, do NOT generate a checklist that would wear the costume of competence. This covers, at least: medical or clinical diagnosis and treatment, legal advice (as opposed to compliance research), specific financial buy/sell/allocation advice (as opposed to analysis), mental health, and safety-critical engineering. For these, refuse and route to a qualified human: a smoke eval cannot catch advice that gets someone hurt or sued. Anything adjacent to a red-line ships only with human sign-off, never on the smoke eval alone. Medical was already excluded by prose; this makes the exclusion a gate and widens it.