AI agent customizer · system prompt & skills builder · free

Shape how your agent thinks. Then prove it.

Telling an agent "be concise and push back on me" is a wish. Temper compiles behavior dials, playbooks and skills into a spec written the way models are trained to read instructions. Test it with a free model that runs right in your browser, on the ChatGPT, Claude or Gemini account you already have, or through an API, a local model or your own agent's endpoint. A separate grader checks every behavior, and every version is kept, so you know a change helped before you ship it.

Open the workbench Connect your model

Free model in your browserChatGPT, Claude, Gemini accounts · no API keyAnthropicOpenAIGeminiOpenRouterGroqMistralDeepSeekOllamaLM Studioyour agent's endpoint

Spec sheet · compiled clauses

REV 01

What people typebe concise, don't be a yes-man, and don't make stuff up

Everything you need to shape an agent

Each part of Temper has its own page. Start anywhere: most people build a spec, connect the AI they use, then test it.

Build

Workbench

Behavior dials, playbooks, rules with reasons, SKILL.md skills and examples, compiled live into a system prompt, AGENTS.md, CLAUDE.md, Cursor or Copilot rules.

Open the workbench → Connect

Your AI, your way

A free model in your browser, your ChatGPT, Claude or Gemini account with no API key, any API, a local model, or your own agent's endpoint.

Connect a model → Test

Graded behavior tests

A probe for every behavior, graded by a separate judge. Repeat runs catch flaky behavior, and failed probes come with a suggested fix.

Run the probes → Learn

How agents read you

Context windows, training side effects, skills and evals: why agents follow some instructions and ignore others.

Read the guide → Save

Encrypted library and vault

Save specs by type and keep API keys in a vault that's encrypted on your device. Sign in anywhere to pick up where you left off.

Create an account → Answers

Questions

System prompts vs AGENTS.md vs CLAUDE.md, what skills are, testing without an API key, and keeping your keys safe.

Read the FAQ →

Why not just tell it?

	Telling the agent in chat	A Temper spec
Wording	Adjectives: "concise", "critical", "careful"	Calibrated clauses with triggers, thresholds and reasons
Lifetime	One conversation, then gone	A file every session loads: AGENTS.md, CLAUDE.md, Cursor rules, SKILL.md, or your app's system prompt
Conflicts	Invisible until the agent picks one at random	Linted, and resolved by an explicit priority order
Skills	Pasted in each time, always taking up context	Loaded only when a task matches the description
Security	Rarely considered	A prompt-injection guard that treats tool output as data
Verification	A feeling after a few chats	Graded probes on your own model, repeat runs, pass rates per version

Workbench: build your AI agent's spec

Work through the six steps, or import a prompt you already have. The spec sheet on the right recompiles as you type, flashes the lines that changed, and flags anything a model is likely to misread. Your draft and versions are saved in this browser.

Spec ready? Connect the AI you use →

Connect ChatGPT, Claude, Gemini or any model

Plug in the AI you actually use. No API key? Choose Free model in your browser to run an open model on your own graphics card, fully automatic, or Your chat account: Temper hands you prompts to paste into ChatGPT, Claude, Gemini or any chat app, and reads the replies you paste back. The agent under test runs your spec; the judge grades replies and powers prompt import, fix suggestions and skill drafting. They can be different apps or providers, which keeps the grader independent of the model it grades.

Roles

Agent under testRuns your compiled spec in probes and the chat. Judge & helperGrades probes, imports prompts, suggests fixes, drafts skills.

Connections

Requests go straight from your browser to the URL you enter, with your key; Temper's server never sees them. Keys stay in this tab unless you tick "Remember on this device", or sign in to keep them in your encrypted vault on every device. Use a key with a spending limit.

Add a connection

Provider

Name

Model

Chat app

Free and paid accounts both work. Temper never signs in to it: you paste prompts in and replies back.

Base URL

API key

Workspace ID (only if your key isn't tied to a workspace)

Copy it from the workspace's page in the Anthropic Console. Or create the key inside a workspace and leave this empty.

Model

Reply field in the response JSON (optional)

Leave empty to auto-detect reply, text, output, content or OpenAI/Anthropic shapes.

Remember the key on this device. Anyone using this browser profile could read it.

Connected? Test your agent →

Test your agent's behavior

One test per behavior, plus any you write yourself. Each probe sends a scenario designed to pressure that behavior to a fresh agent running your spec. The judge then checks the reply against the exact clause the spec promises. Repeat runs show whether a behavior is reliable or only sometimes right.

Runs

Not run yet.

Write your own probe YOUR TEST CASES

Name

Passes if the agent…

What the user says

Custom probes are saved with your spec and included in Run all.

Talk to your agent

A free-form conversation with the agent under test, running the current spec.

No messages yet.

Happy with the results? Export your spec from the workbench, or save it to your encrypted library →

How AI agents read your instructions

Every feature in Temper comes from how agents run and how the models inside them are trained. Here's the short version.

An agent is a loop

The model reads everything in its context window, picks the next action (a reply or a tool call), sees the result, and goes again until it decides the task is done. It has no memory outside that window. Your instructions compete for attention with files, tool output and conversation, so every always-loaded sentence has to earn its place.

Skills load in three tiers

Agent Skills (the open SKILL.md format used by Claude Code and a growing list of other agents) use progressive disclosure. Only the name and description sit in context all the time. The body is read when a task matches the description, and bundled files are opened only when needed. You can install many skills cheaply, and the description decides whether a skill fires at all. That's why Temper lints descriptions harder than anything else.

Name + description, always loaded~100 tokens

SKILL.md body, when the task matches< 5k tokens

Scripts and reference files, when openedas needed

Training decides how your words land

Models are pretrained on large amounts of text, fine-tuned on instruction and response pairs, then optimized with feedback: human preference ratings (RLHF), AI feedback against written principles, and reinforcement learning on tasks with checkable outcomes. Each stage leaves habits you can design around, whichever lab trained the model.

Literal followingRecent models do what you wrote, not what you meant. Vague words like "good" or "clean" get filled in with an average meaning learned in training.Temper: calibrated clauses, vague-word lint

Rules travel with reasonsA rule with its "why" generalizes to cases you didn't list. A bare rule gets applied too rigidly or not at all.Temper: every rule has a "because" field

AgreeablenessPreference training rewards answers people like, which pulls models toward agreeing with you. Permission to disagree has to be explicit.Temper: the Disagreement dial

Examples are magnetsModels copy the surface of examples: length, format, phrasing. A good and bad pair teaches which feature actually matters.Temper: contrastive examples

Shouting backfiresCRITICAL and NEVER in capitals were written for older, less attentive models. Current ones over-apply them.Temper: emphasis lint

Completion pressureReinforcement learning on tasks rewards finishing. Agents lean toward doing more and declaring success early unless scope and stop conditions are stated.Temper: Scope dial, self-check guard

Behavior is a distribution, not a single answer

Models sample each token from a probability distribution, so the same prompt can produce different replies. One good reply proves little. Labs measure behavior across many runs and across models, because the same spec can hold on one model and slip on another, and smaller, faster models follow long specs less reliably. The probe bench does the same at a small scale: repeat runs, any model you connect, and a judge that checks one clause at a time.

ReliablePasses every run. Ship it.

FlakyPasses sometimes. The clause is ambiguous or competing with another; use Suggest a fix.

BrokenNever passes. Something in the spec, or the model's training, is pulling the other way.

Ready to put this into practice? Build a spec in the workbench →

AI agent customization: questions and answers

Common questions about customizing AI agents, writing system prompts and skills, and testing them.

How do I customize an AI agent?

Describe its role and context, then set the behaviors that matter: initiative, disagreement, answer length, how it handles uncertainty, scope and planning. Choose a problem-solving playbook, add house rules with the reason for each, add skills for recurring tasks, and give one good and one bad example reply. Temper compiles this into a system prompt, or an AGENTS.md, CLAUDE.md, Cursor or Copilot file, and then tests that your agent actually behaves that way.

What's the difference between a system prompt, AGENTS.md and CLAUDE.md?

A system prompt is the instruction text an app sends with every request to a model. AGENTS.md is a file in a code repository that coding agents such as Codex and Cursor read automatically. CLAUDE.md is the equivalent for Claude Code. They hold the same kind of instructions, so Temper exports one spec in every format, plus versions sized for a custom GPT, ChatGPT custom instructions, a Claude Project and a Gemini Gem.

What are AI agent skills (SKILL.md)?

Skills are reusable instructions in the open SKILL.md format. Only each skill's name and description stay in the agent's context; the full instructions load when a task matches the description. That lets an agent carry many skills without wasting context, as long as each description says clearly what the skill does and when to use it. Temper's skill builder checks exactly that.

Can I test my agent's prompt without an API key?

Yes. Connect the ChatGPT, Claude or Gemini account you already have, free or paid: Temper gives you a prompt to paste into a new chat and reads the reply you paste back. Or run a free open model directly in your browser on WebGPU, fully automatic, with no account at all.

How do I stop an AI from agreeing with everything I say?

Models are trained partly on human preference ratings, which pulls them toward agreeing with the user. Permission to disagree has to be explicit. Temper's Disagreement dial compiles clear instructions, such as holding a position unless you bring new evidence, and a probe checks that the agent really pushes back on a flawed plan.

Why test a prompt more than once?

Models sample their output, so the same prompt can produce different replies. One good answer proves little. Repeat runs show whether a behavior is reliable or only right some of the time, and testing on the model you'll actually use matters because the same instructions can hold on one model and slip on another.

Is Temper free, and are my API keys safe?

Temper is free. Without an account, everything stays in your browser. With an account, specs and API keys are encrypted in your browser before upload: the server stores only ciphertext and never receives your password, so even a breach of the server can't expose them. Two-factor authentication, a recovery key and a log of sign-in activity are built in.

Still wondering how it works in practice? Read the guide or open the workbench →

Account

Sign in

Loading your account…