Reusable AI Workflows — Any Model

Your best prompts.
One word away.

ICENI turns the prompts you type over and over into single-word commands that produce structured, consistent, machine-parseable output — on any model.

Type review. Get an expert-grade code review. Every time.

What You Type vs. What ICENI Sends

Before — You Type This (130 tokens)

Please review the following code very carefully for any and all bugs, edge cases, security vulnerabilities, and potential improvements. For each issue you find, provide a thorough explanation, the severity level, and a concrete suggested fix. Be comprehensive. def transfer(amount, from_account, to_account): accounts[src] -= amount accounts[dst] += amount return accounts[src]

Unstructured output — varies every time

After — You Type This (1 word)

/review [code pasted]

ICENI sends structured XML calibrated per model

100%

Parse Success Rate

ICENI XML: 5/5 parseable as structured data. Baseline prose: 0/5.

46%

Inter-Agent Routing Compression

130 tokens per hop → 70 tokens per hop in multi-agent swarms.

8/10

Task Types Where ICENI Wins

Wins on structure-heavy tasks: review, audit, refactor, API review, docstring.

~80

Words You Don't Type Per Call

The review alias replaces an 80-word instruction block with one word.

Where ICENI Wins (and Where It Doesn't)

Measured on real tasks. ICENI's instruction overhead is ~70 tokens — so it wins on real inputs and loses on toy snippets.

Input Size	Example	ICENI Result	Why
20 tokens (toy)	`divide()` function	Baseline wins by 6 pts	70-token overhead dominates
72 tokens (small)	`buggy_example.py`	Baseline wins by 2 pts	Overhead still significant
320 tokens (real)	`user_api.py`	ICENI wins by 7 pts	Structure pays off
577 tokens (large)	Order-processing module	ICENI wins by 7 pts	XML structure organizes complex analysis

The rule: On real code (300+ tokens), ICENI's calibrated structure produces better-organized output. On tiny snippets, a plain prompt is fine.

What ICENI Is (and What It Isn't)

ICENI is honest about its limits. It doesn't claim to make every prompt better.

ICENI wins on: Code review, security audit, API review, refactoring, docstring generation — any task where structured output (severity levels, line numbers, categorized findings) matters.
ICENI is neutral on: Test generation, error explanation, deployment checks — prose works fine; ICENI adds consistency, not advantage.
ICENI loses on: Dependency audits (baseline produces tables naturally), commit messages (too simple to need structure) — skip these.
The real saving isn't tokens per call — it's re-runs avoided. A vague prompt gives a bad answer ~25% of the time. A calibrated alias gets it right first. That's the cost saving.

How It Works

Install a pack (or create your own)

iceni pack install code-quality gives you review, security-audit, api-review, refactor and docstring in one command. Or build your own: iceni create review --goal "Review code for bugs" --constraint "cite line numbers". Every alias gets a cryptographically signed identity.

Use it anywhere

Run iceni connect-desktop once, restart Claude Desktop, then type /review in any conversation. ICENI renders your intent as a structured prompt calibrated for that specific model — XML for Claude, markdown for GPT, concise for Kimi.

It improves as you use it

ICENI's discovery engine watches your conversation logs and suggests new aliases for patterns it detects. The calibration engine refines each model's render based on which outputs you accept. No manual tuning required.

Deterministic, parseable, trustable

Every alias version is Ed25519-signed and content-addressed. Output is structured XML that downstream agents can parse deterministically — 100% parse rate vs. 0% for prose. No re-parsing with an LLM required.

Stop typing the same instructions.

ICENI is open-source, runs locally, and connects to Claude Desktop via MCP. Your aliases are stored in SQLite on your machine — nothing leaves your system.

Get Started