A Hermes (and Claude Code) skill that delegates hyper-complex tasks — specs, architecture decisions, deep research, big ideas — to a Claude Code (Fable 5) worker spawned inside herdr. The user talks to Hermes; Hermes asks Fable.
Hermes is fast and conversational. Fable 5 is Anthropic's most capable generally available model. For work that benefits from deeper reasoning — writing a production-grade spec, reasoning about a hard architectural trade-off, doing a comprehensive research synthesis — routing through Fable gives better output without the user leaving their Hermes session.
- Accepts a question or task from Hermes (or directly via CLI).
- Spawns a fresh interactive Claude Code TUI worker inside herdr — interactive
is mandatory: it's what draws on the Max plan (headless
claude -pdoes not). - Delivers the prompt via file protocol: the worker reads a prompt file, writes its complete answer to an answer file, and echoes a sentinel. No TUI scraping, no truncation — the file is the deliverable.
- Surfaces the answer plus the Claude session UUID (resume follow-ups with
claude --resume <uuid>), then closes the worker (or keeps it with--keep).
Confirmed working on this machine via dry runs on 2026-06-11 (Fable 5, Claude Max).
"ask fable to…", "send this to fable", "let fable think about this", "get a fable spec for…", "fable, what do you think about…"
When Hermes hears these, it loads the skill and runs the workflow.
./scripts/install.sh # hermes + claude + ~/.local/bin/ask-fable
./scripts/install.sh --hermes # hermes only
./scripts/install.sh --claude # claude only
./scripts/install.sh --bin # CLI only
./scripts/install.sh --uninstall# Ask a question
ask-fable "Design a rate-limiting architecture for a multi-tenant SaaS API"
# From a file (good for long specs)
ask-fable --file my-problem.md
# Use Sonnet instead for quick/cheap tasks
ask-fable --model claude-sonnet-4-6 "What's the right data structure for X?"
# Give the worker a repo as context
ask-fable --cwd /path/to/repo "Write an SDD spec for the auth module"
# Keep the worker open for iterative follow-ups
ask-fable --keep "Draft the spec; we'll iterate"
# Override timeout for long research (ms; default 600000 = 10 min)
FABLE_TIMEOUT=1200000 ask-fable --file big-research-prompt.md- Re-state the question fully — the Fable worker has no prior conversation context.
- Prefix the prompt with a Hermes-caller framing:
You are a Fable 5 worker called from a Hermes session via herdr. The user wants: <restated question>. Output: <format — structured spec / prose / list / etc.> - Relay verbatim or lightly edit to remove the
━━━framing. - "ask Fable" is explicit authorization to spawn — no confirmation needed.
ask-fable-skill/
├── README.md
├── CHANGELOG.md
├── LICENSE
├── skill/
│ └── SKILL.md ← loaded by Hermes / Claude Code
└── scripts/
├── ask-fable.sh ← one-shot CLI wrapper (symlinked to ~/.local/bin/)
├── install.sh ← skill + bin installer
└── lint.sh ← frontmatter + file checks
Semver. MAJOR = breaking workflow change, MINOR = new pattern/option added, PATCH = fix / clarification.
MIT