Chapter 7

Start your first conversation

The first six chapters prepared the tools, introduced the interface, and configured your key. In this chapter, you finally use Pi Agent. You will ask three realistic household questions, learn when to follow up, and check an answer instead of trusting fluent wording. Pi Agent is more than another chat window: it can be an adviser that is available whenever you need it.

Why devote a chapter to conversation?

After installing the add-on, touring the UI, and configuring OpenRouter, you can open pi-web, select a model, type, and send. Why spend a whole chapter on that?

Because typing a prompt and getting a useful answer are different skills. A vague request produces generic advice. Supply the facts that affect the decision, then use follow-ups to expose assumptions and limits.

  • Try three familiar household questions. These are everyday topics such as air conditioning, weak Wi-Fi, and sharing household work, not engineering exercises.
  • Learn how to follow up. The first response is usually a starting point. A good follow-up can turn generic advice into something you can use.
  • Learn how to evaluate an answer. Models can state false information confidently. Four checks and three countermeasures will help you spot hallucinations.
Tip: You can run every example in your own Pi Agent. Cost and availability depend on the current OpenRouter model and your account. Try each example as you read instead of only reading the page.

Concept: A Session carries the conversation context

Start with a simple mental model.

Each new Pi Agent conversation is a scroll: your messages and the model's responses are stored in order. On each turn, Pi Agent sends the conversation history that fits in the configured context window, and the model adds its next response.

This creates two important behaviors:

  • The model sees more than your latest sentence. Give relevant background first, then ask the question. Earlier details in the same Session can guide the answer.
  • It does not remember the previous Session. A new Session starts with new context. To continue yesterday's work, reopen the existing Session instead of creating a new one.

Some models also return reasoning information in a separate block before the final answer. Pi Agent may show it in a collapsible area. Chapters 9 and 12 explain these blocks and the formats Pi Agent supports.

Mental model: Two people take turns writing on the same sheet of paper. Each person reads the sheet before adding a reply. A new Session is a blank sheet.
What is a token? A token is a unit a model uses to process text; it is not exactly one word or character. Longer context usually means more input usage. Chapter 12 covers the practical trade-offs.

Seven steps to start your first conversation

Follow these seven steps. First, finish configuring OpenRouter in Chapter 6 and confirm that its model appears in Composer. See the Chapter 4 UI tour: the model selector is at lower left and Send is at lower right.

  1. Select New session at the upper left

    A new entry appears at the top of the Session list and the conversation area clears. Its default title may be “Untitled” or the current time. Do not put an unrelated topic in an old Session, because its history can increase usage and distract the model.

  2. Select your OpenRouter model from the lower-left model menu

    Composer is the large input area at the bottom. The model selector is at the lower left, and Send is at the lower right (see Chapter 4). Choose the OpenRouter model configured in Chapter 6, such as anthropic/claude-sonnet-4; this is the exact value entered in models[].name. Use OpenRouter here to continue the international setup from Chapters 5 and 6. Model IDs and availability can change, so check the live catalog. Chapter 13 explains how to switch providers and models.

  3. Enter the first question 
    My living-room light is a Xiaomi model. After 10:00 p.m., I want Home Assistant to dim it gradually to 20%. How should I configure that? Please respond in English.

    The prompt names the device, time, target brightness, platform, and output language. Those constraints are what make the answer specific.

  4. Press Enter or select Send

    Your message appears in the conversation and the response begins to stream. Whether a separate reasoning block appears depends on the selected model, route, and Pi Agent configuration.

    Composer normally uses Enter to send and Shift+Enter for a new line. If your installed version behaves differently, use the visible Send control and check the current pi-web release notes. Chapter 4 shows the Composer layout.
  5. Read the final answer first

    A useful response should:

    • Identify this as a Home Assistant automation task.
    • Separate the trigger (triggers:) at 22:00 from the action (actions:) that dims the light to 20%, and explain any transition. Since Home Assistant 2024.10, new YAML uses the plural keys triggers: / actions: / conditions:; older singular forms can still work but are not recommended for new automations.
    • Offer either current UI steps or YAML that you can review.
    • State assumptions and ask for missing entity details.
  6. Follow up: paste this sentence into Composer 
    How do I add this automation? I have never written YAML, so guide me step by step and respond in English.

    This does two things: it states your experience level and asks for a different output format. The model should give UI steps such as Settings → Automations → Create automation instead of unexplained code.

  7. Follow up again: ask about drawbacks 
    What are the drawbacks of this plan? If I adjust the light while watching television, will the gradual dimming override my change? Please respond in English.

    This tests a failure case. Make a habit of asking when a recommendation can fail. A strong answer should explain what happens if someone adjusts the light during the transition and whether a conditions: section is needed. If the answer uses singular condition:, note that plural keys are recommended for new YAML after 2024.10.

Concept: You have not operated any Home Assistant switch during this exercise. You are using the model in adviser mode. Review and test any proposed automation before applying it.

Three examples: Use AI as a household adviser

These examples do not require Home Assistant knowledge. Choose one and inspect how well the response handles the stated constraints.

Example A: Choose an air conditioner

You type:

I need a new air conditioner for a 20 m² living room with a 3 m ceiling, strong afternoon sun, and no ceiling insulation. My budget is [local currency and amount], and children nap in this room, so low noise matters. Should I choose an inverter model, and what cooling capacity should I discuss with an installer? Give me selection criteria rather than invented product listings, and respond in English.

A useful answer should:

  • Estimate the required cooling load, but label the proposed capacity range as an estimate that depends on climate and construction.
  • Compare inverter and fixed-speed operation without promising savings.
  • Give selection criteria such as rated capacity, efficiency, and published outdoor-unit noise levels instead of inventing models or prices.
  • Tell you to confirm sizing and electrical requirements with a qualified local installer.

Follow up: “Which assumptions have the largest effect on your estimate?” or “What should I measure before asking installers for quotes?”

Example B: Diagnose weak Wi-Fi

You type:

Wi-Fi is strong in the living room but shows only one bar and about 5 Mbps in the main bedroom. The router is an ASUS AX3000 on a cabinet near the entrance; the bedroom is diagonally opposite, behind a concrete wall. How can I test whether to move the router, replace it, or add a mesh node? Please respond in English.

A useful answer should:

  • Treat signal attenuation, interference, placement, and hardware limits as hypotheses to test instead of assuming the router is old.
  • Start with low-cost tests: move the router into the open, test at several locations, and compare bands.
  • Recommend a mesh node only after measurements show that relocation is insufficient.

Follow up: Ask how to measure signal strength, compare 2.4 GHz and 5 GHz, or choose a wired versus wireless backhaul.

Example C: Share household work

You type:

Three people share our home: I work outside the home and return around 7:00 p.m.; my spouse is a freelancer with flexible hours; and our son is in eighth grade. My spouse currently handles almost all household work and says the workload is unsustainable. How can we redistribute it fairly, and which age-appropriate tasks could our son take on? Please respond in English.

A useful answer should:

  • List visible and less visible work, including planning and reminders.
  • Allocate work by time and effort, not by who happens to be at home; flexible hours do not make one person's time less valuable.
  • Suggest age-appropriate responsibilities without assuming one arrangement fits every family.
  • Recommend a trial period and a family review.

Follow up: Ask for a two-week trial schedule, a way to handle missed tasks, or a fairer redistribution if one person still carries most of the work.

Why these examples? You already have your own judgment to compare with the answer, none has one perfect solution, and the immediate risk is low. Still verify purchases, health, safety, and other consequential claims.

Three prompt techniques that improve answers

You do not need special syntax. Apply these three habits.

Tip 1: Supply context

Weak: “Recommend an air conditioner.”
Better: “Recommend selection criteria for a 20 m² west-facing room with a 3 m ceiling, poor insulation, a stated budget, and a low-noise requirement.”

The stronger version gives the model constraints it can address. It also makes omissions easier to notice.

What context should you include? Imagine what a competent person would ask before advising you. Supply those answers proactively, and say when a fact is unknown.

Tip 2: Ask one thing at a time

Weak: Combine weak Wi-Fi, a lighting automation, and screen-time rules in one prompt.
Better: Diagnose Wi-Fi first, then open a new Session for the next unrelated topic.

Separating topics makes the answer easier to inspect and follow up. Keep related steps in one Session; move unrelated work to another.

Tip 3: Follow up

The first answer is a starting point. These prompts work in many situations:

Follow-upWhat it reveals
“What are the drawbacks?”Failure cases and trade-offs
“Is there a cheaper, simpler, or faster option?”A practical alternative
“Explain that in plain English.”A clearer version without unnecessary jargon
“Give one concrete example.”How an abstract idea applies
“Does this still apply if [specific condition]?”Whether the recommendation survives your real constraint
Concept: Treat the exchange as a discussion with a fallible adviser. One response rarely settles a complicated question.

Four checks for every important answer

Fluent writing can make a weak claim sound convincing. Use these checks.

  1. Does it give specific numbers, steps, or names?

    Useful: “Estimate the required capacity and show the assumptions” — you can verify the claim.
    Weak: “Choose a suitable inverter model” — it gives you nothing to test.
    Check: Can you identify an exact term or claim to verify in a primary source?

  2. Does it address the conditions you provided?

    You stated a budget, but the model recommends something over budget → correct it.
    You mentioned children, but the response ignores noise → follow up.
    Check: List your conditions and match each one to the response.

  3. Is it obviously biased?

    Models reflect gaps and biases in their training data. A brand comparison, investment answer, or nutrition answer may inherit a narrow regional viewpoint.
    Check: Ask for a credible opposing view and the evidence that would favor it.

  4. Have you compared the answer with another model?

    A second model can reveal disagreement, but agreement among models is not proof. For purchases, health, finance, or safety, use current primary sources and qualified professionals.
    Check: Chapter 13 shows how to compare models; you must still verify the result independently.

Tip: Checking for specific numbers is fast, but precision alone is not proof. Ask for one concrete example and its source. An unsupported exact number can be more misleading than a cautious range.

When AI is wrong: Understand hallucinations

Start with the key point: every current model can confidently state something false, including models reached through OpenRouter. This is commonly called a hallucination. The model is not deliberately deceiving you; it does not know that its generated claim is wrong. Confidence is not evidence.

Common forms include:

  • Invented products, books, or people. A plausible Xiaomi device model such as “MJZC01-XXX” may not exist.
  • Invented URLs, actions, or commands. For example, light.slow_dim is not a Home Assistant action. Do not copy an invented call such as light.slow_dim: entity_id: light.xxx. A gradual change uses a real action such as light.turn_on with the documented transition option. This is an intentional hallucination example; do not paste it into an automation.
  • Mixed-up product features. The answer may assign one brand's feature to another.
  • Outdated information. A model trained on 2024 material may present a two-year-old option as current in 2026.

Three countermeasures

  1. Verify the facts that affect your decision, not every sentence.

    Spend 30 seconds checking model numbers, commands, prices, regulations, and anything that changes what you will do. Prefer current primary sources.

  2. Ask another model.

    A second answer can reveal uncertainty, but it is not independent evidence. Resolve disagreement with primary documentation or a qualified professional. See Chapter 13.

  3. Be more skeptical when the model sounds certain.

    This is counterintuitive, but hallucinated claims can sound more certain than true ones. Ask for sources, assumptions, and uncertainty, then check the cited material yourself.

Hard stop: For health, legal, financial, electrical, gas, lock, or alarm decisions, use AI only for an initial discussion, never as the final decision. Consult a qualified professional and use deterministic safety controls.
Useful reminder: Tell yourself, “This is a draft, not a conclusion.” Review an AI answer as you would review a draft.

Your conversation remains in its Session

Close the pi-web tab and return tomorrow: the Session should remain in the list unless you delete it or its data is lost.

Chapter 8 covers these actions in detail:

  • Reopen: Select the previous Session and continue with its stored context.
  • Rename: Use the available Session menu to give it a searchable title.
  • Delete: Remove a Session you no longer need, after exporting anything important.
  • Fork: If your version supports it, branch from a point in the conversation and explore a different direction while preserving the original.

See Chapter 8: Session management for version-qualified instructions on organizing, exporting, and deleting Sessions.

Concept: Do not be afraid to create a new Session. The important rule is not to put every question in one ever-growing conversation. Long context increases provider usage and can crowd out earlier details. Use one Session per topic.

Six common first-conversation problems

Nothing happens after I send the message

Check in this order:

  1. The model selector below Composer is unset or empty. Open it and select the configured OpenRouter model.
  2. Open Models and verify that the OpenRouter provider has a key and an exact current model ID.
  3. Inspect the browser Console and Network panels. 401 Unauthorized means the key is invalid or expired; Network Error means the provider cannot be reached. Do not paste secrets into a support request.
  4. Reopen Pi Agent from the Home Assistant sidebar, then check the add-on logs before restarting it.
The response stays on “thinking”

A fixed time does not distinguish a slow model from a failed request.

  1. The account may have exhausted its quota. Inspect the request or add-on log for 402 Payment Required, 429 Rate Limit, or quota exceeded. Add credit if required or select another available OpenRouter route.
  2. Your network may block the provider route. Confirm that it can reach OpenRouter. A legacy direct GLM route may use open.bigmodel.cn; opening https://open.bigmodel.cn can help confirm basic access to that host.
  3. The model service may be unavailable. Check the provider's current status page and try again later.
  4. If the route is healthy, a reasoning model may simply take longer; wait or select a lower-latency model for the next attempt.
Text appears as empty boxes or corrupted characters
  1. Reload the page with Ctrl+R.
  2. Try a current browser and confirm that the page is being served as UTF-8.
  3. Test a private window with extensions disabled. If that works, re-enable extensions one at a time.
The answer stops mid-sentence

The response may have reached the model's single-response output limit, the connection may have ended, or the provider may have returned an error. For example, max_tokens=4096 is a documented default for a legacy direct GLM route; see the Z.ai parameter documentation.

  1. The simplest response is to ask “Continue from the last complete sentence” if the request ended normally.
  2. Inspect the request and model configuration before changing output limits.
  3. Split a large request into smaller parts.
A reasoning block grows for a long time

Some reasoning models take longer before producing a final answer.

  1. Wait while the request is active. Do not select Stop unless you intend to end it.
  2. For future low-risk questions, choose a non-reasoning or lower-latency model.
  3. If the request appears stalled, inspect Network and the add-on log for an error.
The model seems to forget earlier details

Two common causes are:

  1. The conversation exceeds the model's effective context. Ask for a checked summary and start a new Session with it.
  2. Your follow-up is ambiguous. Replace “that” with the exact device, plan, or claim you mean.

Conversation FAQ

How long should one conversation be?

For personal questions, 10-30 turns can be a normal in-depth discussion. Technical work may reach 50-100 turns, but there is no reliable round limit; usage depends on accumulated tokens and the selected model.

Watch the context indicator when available:

  • Keep one topic in each Session.
  • Summarize resolved work before starting a new Session.
  • If the conversation exceeds 150 turns or the model becomes noticeably slower, start a new Session or fork with a verified summary.

Chapter 12 explains context and reasoning costs.

Does prompt language affect answer quality?

In 2026, mainstream models can answer well in many languages. Performance still varies by model, language, and subject, so use the language in which you can state constraints precisely.

English may help when:

  • You are asking about a niche technical subject whose primary documentation is mainly in English.
  • You need the AI to produce English code comments or documentation.
  • You are practicing English writing.

Otherwise, use the language in which you can state constraints most precisely, and request the output language you need.

Can I apply an AI-generated Home Assistant automation directly?

Review it before applying it. Use three checks:

  1. If you understand it, validate and test it manually. Use the current Home Assistant tools appropriate to that configuration.
  2. If you do not understand it, ask for a line-by-line explanation. Do not apply code you cannot review.
  3. For locks, gas, alarms, or other safety controls, first test on an unimportant entity. In YAML, review the actions: block and its target entity_id especially carefully.

Do not rely on promised future Pi Agent features. Use only controls present in your installed version.

Does deleting a Session make the model forget it?

Separate two ideas:

  • Deleting the Session removes your stored copy from Pi Agent, subject to any backups.
  • Starting a new Session: the model still does not remember you because the previous Session's context is not automatically included.

This separation is intentional, but it does not prove that a cloud provider retains nothing; review that provider's current data policy.

  1. For short-term continuity, paste a concise, verified background note at the start of a new Session.
  2. For long-term records, store the information deliberately in a secure location rather than relying on a hypothetical future memory feature.
Can I attach images or files?

Support depends on both your pi-web version and the exact model route.

  • Images: Confirm that the selected model and route accept image input.
  • Text-only models: Image input may fail or be ignored.
  • Files: Confirm supported formats and inspect what text is sent before sharing sensitive data.

Do not use a failed upload as a capability test if the file contains private information.

Can I interrupt a response?

If your version shows Stop while generating, select it to end the request.

After stopping:

  • The partial response may remain in the Session.
  • Ask it to continue from the last complete point if appropriate.
  • Give a new direction if the answer went off topic.
  • Use the message menu if your version supports deleting that message.

Stop a response when it is unsafe, irrelevant, or based on a mistaken premise.