prompt

Analyze one or more GitHub issues, classify complexity, and produce a copy-paste-ready prompt with a model recommendation. The goal is quality-conservative right-sizing — never under-resource a task, but don't waste Opus tokens on a typo fix.

Model Lineup & Effort Levels

The fleet is Fable, Opus, Sonnet, Haiku — the same four families named in CLAUDE.md, .claude/rules/subagent-orchestration.md "Model Selection", and .claude/agents/README.md. In the Claude Code picker, Opus is the default.

Always name the family, never the version (#791). A bare family name means the current non-legacy model of that family, so it never goes stale — earlier generations still listed in the picker are Legacy and are never recommended. The bare aliases used elsewhere (agent frontmatter, spawn sites) resolve the same way: opus, sonnet, haiku, and fable each resolve to the newest non-legacy model of that family. Write an explicit model ID (e.g. claude-opus-5) only where a tool literally consumes the string. Full rule and its scope: .claude/agents/README.md "Model naming".

The picker's second control is effort, and this skill recommends one on every issue. Use the picker's own labels — Low, Medium, High, Extra, Max — because a recommendation the user cannot map onto the control in front of them is one they ignore. This is the one place the label↔token mapping is taught, for the cases where something downstream consumes the raw token: Low (low), Medium (medium), High (high), Extra (xhigh), Max (max). Everywhere else, and in every recommendation this skill emits, the label alone is the value.

Ultra code (ultracode) is not a step on that ladder — it is a session-level orchestration mode, opted into for the whole session rather than set per turn (the per-call effort enum stops at max). Name it only as a step-up note alongside a real level, never as the recommended level itself.

The internal Heavy/Standard/Light tier vocabulary and decision tree are unchanged; each tier maps to a recommended effort in the output: Heavy → Extra, Standard → High, Light → Low. Users may adjust within a tier's range — a Heavy task with correctness-critical work steps up to Max; a Heavy task needing multi-agent orchestration steps up to Ultra code; a borderline Standard task steps up to Extra or down to Medium. Model choice has its own step-up: the hardest long-horizon / orchestration work can move from Opus up to Fable (see Heavy, below).

Fast mode: the picker has a Fast mode toggle (also /fast) that gives Claude Opus faster output — it speeds output without downgrading to a smaller model. It is offered on Opus; it is not offered for Sonnet, Haiku, or Fable, so it never pairs with a Light-tier recommendation. /prompt does NOT accept a --fast flag and does not factor Fast mode into recommendations — it is a user-toggled picker option. A --fast flag is a possible follow-up, not part of this skill.

Separately from Fast mode: for Light-tier work, Haiku is a valid cheaper alternative to Sonnet; the output notes this on Light-tier recommendations.

MANDATORY OUTPUT FORMAT: Every per-issue prompt block printed to the transcript MUST open and close with ~~~ tilde fences. NEVER use backtick fences as the outer prompt-block delimiter. This governs fallback mode and print-on-demand replay — the two paths that print a block. In chip mode the prompt rides inside the chip rather than being printed, so it needs no fence; its content is otherwise identical.

Per-block model and effort labels (mandatory): The **Model:** line is the first content inside each tilde-fenced block (immediately after the opening ~~~), immediately followed by the **Effort:** line — no blank line between the two:

**Model:** {MODEL} — {REASON}
**Effort:** {LEVEL} — {REASON}

{MODEL} is the bare family name for that issue's issue_tier (from Step 5 — not the batch tier) and {LEVEL} is that tier's picker label. Each {REASON} is a concise task-type phrase of at most 10 words derived from that issue's signals (dominant drivers such as rules/CLAUDE.md, orchestration, file count, AC count, skills, dependencies, or scope keywords). Both labels must be inside the tilde fence so a pasted block is self-explanatory without surrounding prose. In chip mode the same two lines MUST open the chip's prompt text and appear in the visible short summary — chips preset neither picker control, so the user needs both before clicking and the spawned session needs them after.