Skills

Here’s the specific kind of friction that makes agent-assisted writing feel worse than it should. You have a rule you want the agent to follow — “don’t use the word delve,” “prefer ‘said’ in dialogue tags,” “check the Legendry before inventing new city names,” whatever — and you keep having to tell the agent the same thing on every request. You type the rule into the prompt. The agent follows it. Next request, same rule, type it again. Fifty requests later, you’re still typing the same five rules at the top of every prompt, and about one time in three you forget one and the agent produces output that breaks the rule you forgot to include. The Skills system exists to eliminate this friction. Write an instruction once, save it as a skill, and it gets automatically injected into every agent prompt that targets it. No more retyping. No more forgetting. The agent always has your rules loaded.

That’s the whole pitch. Skills are reusable, persistent instructions that give your agents rules to follow across every request. Ishvana ships with a bundled set of defaults, and you can write your own.

Bundled skills

Ishvana ships with nine bundled skills covering core writing craft and Machine Tells awareness. They’re active by default, but you can disable any of them per-project.

Writing craft

Show Don’t Tell — prefer concrete sensory detail over abstract emotional statements.
Dialogue Craft — write natural dialogue with invisible “said” attribution. No “hissed,” “breathed,” “murmured” overuse.
Precise Language — choose specific, concrete words over vague or grandiose ones.
Scene Grounding — anchor scenes in specific physical and temporal detail.
Tension and Pacing — vary sentence rhythm and control information release.

These are the fundamentals of fiction craft. Leaving them active gives Hawken a baseline of good prose conventions that would otherwise need to be in every prompt.

Machine Tells awareness

Avoid Machine Prose Patterns — avoid vocabulary and phrases that mark model-generated prose. Delve, tapestry, testament, weave, navigate, harness, myriad, realm. The full Machine Tells vocabulary.
Character Naming — avoid the fantasy names models default to. Elara, Kael, Lyra, Theron, Zephyr. The name set that appears in roughly 40% of model-generated fantasy fiction.

These two skills exist specifically because model-generated writing drifts into a recognizable vocabulary and name set unless explicitly told not to. Leaving them active is the cheapest way to make output stop sounding like it came out of a model.

Genre

Dark Fantasy Conventions — genre conventions for dark and grimdark fantasy fiction.
Worldbuilding Consistency — maintain internal consistency when generating world details.

These are more project-specific. Disable if your project isn’t dark fantasy; enable if it is.

Creating your own skills

Skills are Markdown files with a simple format:

Name — a unique identifier.
Description — one-line summary shown in the UI.
Agent target — which agent should use this skill (or all agents).
Tags — labels for filtering and discovery.
Body — the actual instruction text in natural language.

Create skills from the Etherforce → Skills tab in the Analysis workspace, or write Markdown files directly in your project’s skills/ folder. Both paths produce the same result — a skill file that gets loaded by the skill system and injected into prompts targeting the matching agent.

Example skills people actually write:

“Always use American spelling, never British.” Simple, specific, saves you from “colour” vs. “color” drift.
“When describing fight scenes, include a specific tactile detail from the environment.” Forces Hawken to ground action rather than handwave.
“Characters with the trait ‘Aware’ speak in fragments and never finish sentences.” Project-specific craft rule.
“Dialogue attribution should use action beats, not dialogue tags.” Industry-specific convention.
“The magic system’s cost is physical exhaustion. Describe it viscerally, not abstractly.” Canon enforcement.

None of these are bundled defaults because they’re specific to individual projects. That’s the whole point of custom skills.

Skill management

Edit — change the Markdown file. Updates take effect on the next agent request.
Delete — remove via the UI or delete the file from disk.
Seed defaults — copy bundled skills into your project directory for customization. Once customized, your version takes precedence over the bundled one.
Filter — search skills by name, description, or agent target.

Skills are per-project, which matters because different projects have different needs. Your fantasy epic’s skill set is different from your noir detective story’s skill set. Switching projects automatically loads that project’s skill set.

Agent targeting

Each skill can target a specific agent or all agents:

Target Hawken for writing-specific instructions. All nine bundled skills target Hawken by default because they’re writing-focused.
Target all agents by leaving the agent field empty. Useful for project-wide rules like “always check lore before inventing new cities” — a rule that applies to any agent doing any kind of work on the project.

Targeting matters because it controls which agents consume the skill’s tokens. A Hawken-only skill doesn’t affect Lagan’s research responses. An all-agents skill is loaded into every request regardless of who’s responding.

Token budget

Tokens are the currency of LLM requests, and skills spend them on every request they target. The bundled skills are compact and well-worth the cost, but third-party or custom skills can be long and expensive. Check the token count of your skills in the Skills tab’s detail view, and disable any you’re not actually using.

Per-project customization

Skills are per-project, which matters more than it sounds. A single Ishvana install with five projects can have five completely different skill sets:

Project A: all nine bundled skills plus a custom “Regency romance conventions” skill.
Project B: only the Machine Tells awareness skills because the rest don’t apply.
Project C: custom skills specific to the Swordsfall universe, overriding some of the bundled defaults.
Project D: a minimal skill set for a short-story project where verbose skills aren’t justified.
Project E: fresh skills derived from a different style guide.

Project-level skills can override bundled ones by name. Customize a bundled skill for a specific project, and your version replaces the default for that project only.

Where skills fit in the Etherforce pipeline

Skills get injected into the agent’s prompt at the system-prompt level, before the actual user message. They’re not part of the conversation history — they’re part of the agent’s instructions. This matters for context compression: as conversations grow and older turns get compressed to save tokens, skills stay intact because they’re system-level, not conversational.

For the full explanation of how skills fit into the LLM pipeline, see Etherforce and the conceptual pages in the Engine section.

What’s next

Etherforce The LLM engine where skills get injected into agent prompts.

Hawken The primary consumer of writing skills — style, craft, and genre instructions.

ProseGuard The deterministic counterpart to skills — rule-based enforcement instead of agent guidance.