Optimize Your Prompt

Transform a rough idea into an engineered, provider-shaped prompt.

Domain
General
Coding & Tech
Creative Writing
Draft
Target
0%
Orchestration Idle

Nothing compiled yet

Write a draft on the left and hit Transform. The engineered prompt, its rationale and the answer run will appear here.

Claude Prompt Optimizer

Current Claude models already plan and check their work. A rough instruction becomes a clearer prompt — usually shorter — with a goal, the context that matters, and a concrete way to tell it is done.

Instruction

One line is enough. Missing goals, success checks and output shapes get added. Pressure language, self-checks and scripts Claude no longer needs get cut. Ctrl+Enter rewrites.

Compiler

Same Venice catalogue as the other tabs. The rewrite is for Claude; this model only writes it.

Claude rewrite Idle

Nothing rewritten yet

Type an instruction on the left. The improved prompt, what changed, and which practices applied will land here.

Good and bad practice

Newer Claude models already plan and check their own work. Overprompting makes them slower, wordier and more expensive. State the goal, the context that changes the answer, what done looks like, and the real constraints — once, in plain language — then let the model work. Audit and trim the prompt when it grows. Drawn from Addy Osmani’s writing on overprompting Claude, including the dated patterns a prompt audit removes.

Good practice
Good State the goal and why it matters

The model can guess a task. It cannot guess the outcome you need or who it is for.

Before 
Summarize this report.
After 
Summarize this report so a director can decide in two minutes whether to fund the project. Open with the decision the report supports.
Good Short context. Extra detail lives in a doc

Keep the facts that change this answer. Point at the rest instead of pasting it.

Before 
Write the outage note. Here is our entire 40-page voice guide, last year's brand deck, and every previous incident email…
After 
Write the outage note for new customers. Follow the voice rules in docs/voice.md. This note only needs what broke, who is affected, and when it will be fixed.
Good Say what done looks like

A concrete check beats "make it good": tests passing, acceptance criteria, or an output format.

Before 
Fix the flaky login test and make the suite healthier.
After 
Fix the flaky login test. Done when npm test passes twice in a row and the change does not weaken an assertion to get there.
Good One constraint, plain language, then stop

Say the real limit once, with the reason. Then leave the method to the model.

Before 
You are a world-class expert. First plan, then draft, then critique, then rewrite. Be thorough. Do not be lazy.
After 
Recommend one vendor. The note has to fit in a Slack message, because that is where the decision will be made. Name the price and the risk.
Anti-patterns
Bad Self-verification boilerplate

"Double-check your work" is taken literally and duplicates planning the model already does.

Before 
Summarize the report. Before you answer, double-check your work, verify every claim, and if anything is wrong, fix it and check again.
After 
Summarize the report in one page: the decision, the numbers that support it, and what is still unknown. Ask if the report is not attached. Do not invent figures.
Bad ALL-CAPS emphasis

CRITICAL / MUST / NEVER used as volume stops meaning anything, and the anxious register leaks into the answer.

Before 
CRITICAL: You MUST ALWAYS use a formal tone. NEVER use contractions. YOU MUST NOT miss any detail.
After 
Write a formal summary without contractions. Include every figure from the source, because the reader will quote them.
Bad Rigid step-by-step templates

A hand-written script stacks on top of the model's own plan. Keep steps only when order is the constraint.

Before 
STEP 1: Restate the question. STEP 2: List assumptions. STEP 3: Think step by step. STEP 4: Draft. STEP 5: Critique. STEP 6: Rewrite.
After 
Answer the question. State any assumption that would change the answer. End with the recommendation and the evidence it rests on.
Bad Excessive or outdated few-shots

Examples are the strongest signal in a prompt. A pile of them freezes an older model's length and tone.

Before 
Here are eight sample summaries from 2023. Match their length, their headings, and their "In conclusion" close exactly.
After 
Three bullets: decision, evidence, open question. No sample to imitate.
Bad Contradictory or repeated rules

Current models follow conflicting lines more literally, and spend the turn reconciling them.

Before 
Be exhaustive. Keep it under 50 words. Always use bullets. Never use bullets. As stated above, be exhaustive.
After 
Three bullets. Each one is a decision the reader can act on.
Bad Stale thinking budgets

Manual thinking budgets and "think harder" were settings for older models. They do not belong in the prompt.

Before 
Use a thinking budget of 8000 tokens. Think harder. Show your full reasoning in the answer.
After 
Recommend one option. Put the reason in two sentences after it.
Bad Bloated context

A long dump makes every instruction less likely to be followed. Trim to what this task needs, and re-audit when the prompt grows.

Before 
Read the whole wiki, the old CLAUDE.md, three postmortems, and the style guide pasted below, then answer the one-line question at the end.
After 
Answer the question using only the incident dated 12 March. The wiki and the style guide stay where they are; do not restate them.

Engineer the Goal

A goal is not a wish. Turn a rough intention into an observable outcome, the evidence that would prove it, the work it deliberately excludes, and a rubric that can fail the result — before anything is optimized or harnessed.

Intention
Model & identity
Goal dossier Idle

No goal engineered yet

Write the intention on the left. The outcome, evidence, exclusions, assumptions, constraints and rubric will appear here as one charter.

Model Tuner

Pick the model the prompt will run on. Learn what its maker recommends, then turn a one-line idea into a prompt written the way that model reads best.

Provider
Model
Your prompt
General
Writing
Coding
Agentic
Extraction
Analysis
Compiler

Any Venice model can write the prompt. The target model above decides its style.

Model guide Idle

Build Your Agent

Generate precise system instructions, description and starters for a custom agent.

Brief
Model
0%
Orchestration Idle

No agent configured

Describe the agent on the left. Its name, description, system instructions and conversation starters land here.

Harness Builder

A harness is not a prompt — it is the environment that turns an agent into a repeatable operator. Describe the work you want harnessed and get the whole structure: rules, playbooks, subagents, skills, MCP wiring, hooks, permissions, execution layer, state schema, evals, context budget and the loop that makes it better every run.

Use case
Autonomy
Target harness
Claude Code
Codex CLI
Cursor
Gemini CLI
Include
Model
Orchestration Idle

No harness built

Describe the workflow. Each layer streams in as it is written — rules first, then the specialists, controls and the improvement loop.

Agent Skills

Build a portable SKILL.md package. The Agent Skills standard is no longer Claude-only — Claude Code, Codex CLI, Cursor and Gemini CLI all load the same folder.

Skill type
🔄
Workflow
Multi-step processes
🔧
Tool
Format handlers & APIs
📚
Reference
Guidelines & standards
⚡
Capabilities
Extend agent abilities
Definition
Include
Model
0%
Orchestration Idle

No package built

Name the skill and describe it. The analysis, SKILL.md, scripts and folder tree render here, ready to download.

Agent Rules

Write the file every coding agent reads before it touches your repo. AGENTS.md is the open standard — Codex CLI, Cursor, Copilot, Gemini CLI, Aider, Windsurf and Zed all read it natively. Pick any harness that needs its own format and get that too, derived from the same source.

Project
Commands

The highest-value thing in the file. Exact commands, so an agent never guesses at your scripts.

House rules
Target harnesses
AGENTS.md
Claude Code
Cursor
Copilot
Gemini CLI
Windsurf
Aider

Codex CLI, Cursor, Copilot, Gemini CLI, Aider, Windsurf and Zed all read AGENTS.md — only add a harness here when you want its native format as well.

Model
Orchestration Idle

No rules written

Describe the project and its exact commands. The AGENTS.md, any harness-native variants, and an install guide land here.

Plugin Builder

Design a complete Claude Code plugin — commands, agents, skills, hooks and MCP config — packaged into build instructions any AI can scaffold from.

Definition
Model
0%
Orchestration Idle

No plugin designed

Name the plugin and describe it. Its component graph — commands, agents, skills, hooks, MCP servers — renders here.

Loop Design

Turn a plain-English request into a Loop Engineering spec — trigger, actions, proof, memory and stop conditions — pressure-tested by a second-pass Loop Critic.

Request
Model
0%
Orchestration Idle

No loop compiled

Describe the repeating job. The compiled spec, the critic's findings and four export formats appear here.

Content Loop

Compile a repeatable content engine: pillars, a dated calendar, channel-native drafts, Venice-generated visuals and the feedback loop that makes cycle two better than cycle one.

Subject
Channels
LinkedIn
X
Newsletter
Blog
YouTube
Instagram
TikTok
Cadence
6
Visuals
Model
Orchestration Idle

No content engine yet

Set the subject, channels and cadence. The strategy, calendar, drafts, visuals and automation spec build up here stage by stage.

Gauntlet Loop

Forge a tool-creation gauntlet: parallel sub-agents that each own one dimension, an adversarial critic that scores against a named gold standard, and a loop that will not exit until the blind comparison comes back in your favour.

Objective
Domain
Software
Game / 3D
Design
Writing
Research
Loop parameters
6
5
Model
Orchestration Idle

No gauntlet forged

Name the objective and the standard it must beat. The agent roster, critic rubric, blind comparison protocol and runnable mega-prompt assemble here.

Be Ambitious Prompting

Felix Rieseberg (Anthropic) says we're not ambitious enough with models. His fix has three levers: Altitude — aim at the goal beyond the goal; Depth — work with the instruments of the field's best, checked by an outside judge; Range — replace the nouns with verbs, then assemble a crew of unexpected fields. Better, not bigger. Based on the article · summarized by RoboNuggets · skill from robonuggets/be-ambitious.

The three types

⛰️ Altitude

Stop fixing single crashes; keep the apps crash-free. Ask: what do you want to happen next? Build a 3-rung ladder: understand → decide → replace the ritual.

🔬 Depth

Use the raw data and real instruments the top experts use, set an "unreasonable" goal and treat it as reasonable, adversarially test what everyone believes — and always install an outside judge.

🎭 Range

A QBR isn't "a deck", it's "getting busy people to believe a story in 30 minutes". Verbs reveal the crew: analyst, journalist, short seller. Do what they do — don't talk like them.

Ambition Multiplier

Agent skill library — downloadable

Every generator in this app has a matching, tested SKILL.md you can drop into any agent (Claude Code, Codex CLI, Cursor, Hermes). Generated with the skill-creator format: frontmatter, triggers, steps, pitfalls — each one executed and verified against the live Venice API before publishing.

Skill Vault

Every generator in this app ships as a portable, tested viv-*.md skill. Drop any folder into Claude Code, Codex CLI, Cursor, Gemini CLI or Hermes — they all load the same Agent Skills standard. Each skill was executed and verified against the live Venice API before publishing. All MIT licensed.

How to install
  1. Download a viv-<name>.md file below.
  2. Rename it to SKILL.md inside a folder named after the skill, e.g. viv-be-ambitious/SKILL.md.
  3. Drop the folder into your agent's skills directory (e.g. ~/.claude/skills/, ~/.hermes/skills/, or your project's .agents/skills/).
All skills — click to download
Origin: robonuggets/be-ambitious ↗