RUNE skill for Hermes

A Hermes Agent skill that rewrites a rough request into an explicit 8-layer prompt, plus an optional shell wrapper around the RUNE wand CLI.

Measured result: in our blind pilot (50 pairs, 2 Gemini models) it did not make answers better. Amplified prompts were preferred in 19.6% of decided pairs (95% CI 10.2–29.3%, 46 decided of 50), so they lost. The pilot covers only those prompts and models. Method, per-domain results and limits are in RUNE docs/BENCHMARKS.md.

Why

Agent instructions often leave role, scope, permissions, stop conditions and output format implicit. This skill gives Hermes a fixed checklist for making those parts explicit, which is useful when you want to read, review or reuse the structure of a request (subagent briefs, cron prompts, plans).

It is a structuring aid, not a quality booster. Given the pilot result above, do not expect better answers just because a prompt was amplified.

Quick start

Install the skill into Hermes (only SKILL.md is needed):

git clone https://github.com/neurabytelabs/rune-skill
cd rune-skill
mkdir -p ~/.hermes/skills/prompt-engineering/rune-prompt-amplification
cp SKILL.md ~/.hermes/skills/prompt-engineering/rune-prompt-amplification/SKILL.md

Start a fresh Hermes session with the skill loaded:

hermes chat -s prompt-engineering/rune-prompt-amplification \
  -q "RUNE this into a launch plan: ship a private beta for my agent mesh"

New skills may only appear in a new Hermes session. Use the full categorized path to avoid name collisions with other RUNE-related skills.

Optional: the main.sh wrapper

main.sh calls the RUNE wand CLI and strips ANSI colors so its output can be piped. It uses wand if it is on PATH, otherwise $RUNE_DIR/wand.py (default RUNE_DIR is $HOME/Documents/GitHub/rune).

git clone https://github.com/neurabytelabs/rune
cd rune && python3 -m pip install -e .    # requires Python >= 3.11
export RUNE_DIR="$PWD"

rune-wand is not published on PyPI at the time of writing, so install from source.

Commands that do not call a model:

bash main.sh version
bash main.sh grimoire

Commands that call a model need a provider (see Configuration):

echo "Explain quantum computing" | bash main.sh       # default: inscribe (show enhanced prompt only)
bash main.sh "Write a marketing email for my SaaS"
bash main.sh cast "Design a REST API for a todo app"  # enhance and run
bash main.sh validate "Check this prompt quality"
bash main.sh duel "Compare sorting algorithms"        # raw vs enhanced
bash main.sh swarm "Evolve the best coding prompt"

The wrapper accepts these wand subcommands: cast inscribe duel grimoire test validate forge stats cost config fuse bind lineage swarm version. Any other first argument is treated as prompt text for inscribe.

Configuration

RUNE reads provider settings from environment variables and/or ~/.rune/config.toml. Never commit real keys.

mkdir -p ~/.rune
cat > ~/.rune/config.toml <<'EOF'
[llm]
api_url = "https://your-openai-compatible-endpoint/v1/chat/completions"
api_key = "your-api-key"
default_model = "your-model"
timeout = 300
EOF

Or:

export RUNE_API_URL="https://your-openai-compatible-endpoint/v1/chat/completions"
export RUNE_API_KEY="your-api-key"

For backwards compatibility main.sh also sources ~/.secrets if that file exists. New setups should use the options above.

How it works

flowchart LR
    A[Request] --> B{Complex?}
    B -- no --> C[Answer directly]
    B -- yes --> D[Fill layers L0-L7]
    D --> E[Four-point check]
    E --> F[Answer or visible RUNE pass]

Loading

SKILL.md tells Hermes to answer simple requests directly and, for complex ones, to work through eight layers internally:

Layer	Name	Covers
L0	System Core	role, stance, behavioral rules
L1	Context Identity	domain, history, audience, constraints
L2	Intent Scope	goal, success criteria, output shape
L3	Governance	safety, permissions, non-goals
L4	Cognitive Engine	reasoning strategy, decomposition, critique
L5	Capabilities	tools, files, integrations, retrieval
L6	QA	final check (below)
L7	Output Meta	language, tone, structure, format

The L6 check asks four questions, named after Spinoza's terms: does the answer help the user act (Conatus), is it coherent (Ratio), is it clear (Laetitia), is it not overengineered (Natura). If one fails, the answer is revised.

Layers stay hidden unless the user asks for a visible "RUNE mode", in which case Hermes prints a compact L0–L7 breakdown. SKILL.md also contains short reusable patterns for planning, coding, debugging and cron-prompt design.

Repository contents:

• SKILL.md — the Hermes skill (frontmatter + instructions).
• main.sh — optional wand CLI wrapper.
• package.json — package and Hermes metadata.

Status / limits

• The blind pilot found amplified prompts lost to raw prompts (see above). It covered one prompt set, Gemini generation models only, and a single sample per cell; it does not show how the skill behaves in Hermes on other models.
• SKILL.md is instruction text; there are no automated tests in this repo.
• main.sh depends on the separate RUNE repo, which is not on PyPI.
• Declared platforms in SKILL.md: macOS and Linux.
• OpenClaw support is legacy: main.sh still works as an executable-style skill, but Hermes is the primary target.

To check changes to this repo:

bash -n main.sh
bash main.sh version
bash main.sh grimoire

Related

• RUNE framework — the wand CLI and the benchmark harness
• RUNE Playground
• Hermes Agent docs

License

MIT. See LICENSE.

Author: Mustafa Saraç · NeuraByte Labs