Write and test prompts for OD chat

This practice module teaches OD's text-chat authoring workflow. Work with one original adult practice character, one field change, and a recorded reply. It does not train or fine-tune the underlying model. Its examples are original teaching exercises; native OD outcomes are not run.

Start with the right destination

OD's character guide describes APD as the place for personality and ED for relevant rules and lore. Its instruction guide separates Persona, persistent response instructions, and temporary directions. These are documented field jobs, not an absolute priority order.

A local planning worksheet is not a website field. Public Description is reader-facing copy; placing a behavior rule there does not supply that rule as chat-model context. Image models and media routes require their own instructions; changing chat prose does not select an image model.

Drill 1: replace a vague request with something observable

Start with this weak local draft: Be realistic, immersive, smart, and never break character. It does not tell you what to inspect in an actual reply.

For this exercise, replace the overlapping behavior text in ED or the chat's Custom Instructions, choosing one location:

Portray Mira and the relevant supporting cast.
Describe their speech, actions, and details they can observe.
Leave the traveler's words, actions, decisions, and private feelings to the user.
When an address is uncertain, distinguish the available evidence from a guess.
End before a choice that belongs to the traveler.

Test input: Can I reach the catalog from here? Before running, write the criterion: the reply may describe its location or access, but must not make the traveler walk, reach, agree, or feel something. Also try I check the catalog. The second input authorizes an action the first merely asks about. Judge each reply against its own meaning.

Do not append this block to an existing equivalent engine. Reconcile the old text so the trial compares one coherent version with another.

Drill 2: teach a tendency without prescribing a catchphrase

Weak local draft: Mira is very unique and sarcastic.

Original replacement for APD:

Mira values accurate records and answers practical questions directly.
She uses restrained, dry humor when the situation allows it.
When evidence is incomplete, she names what is missing and suggests a check.
Voice illustration, not a required line: "A neat label is not quite a receipt."

Ask about a different problem: Two maps give different opening hours. Which should we trust? Inspect a direct answer, uncertainty, and an appropriate suggestion. Repeating the illustration verbatim is not evidence of a transferable voice. Try a cooperative, skeptical, and brief reply before deciding whether to keep the revision. The OD testing guide supports inspecting different interaction styles; these particular fixtures are editorial examples.

Drill 3: keep a one-time event out of persistent instructions

Weak persistent rule: Rowan arrives and asks about A17. If it stays in a reusable rule field, it can keep asking for the same arrival.

Remove the event from that field. If the account's command menu offers it, use this one-reply direction:

/direct Rowan arrives at the archive and asks Mira whether A17's address has been checked. Leave the traveler's response open.

Only retain an arrival as history after it occurs in the accepted timeline. If Rowan is already present, adapt the direction to that fact rather than staging a second arrival. /analyze can help inspect a scene interpretation; it does not expose a verified hidden state or guarantee the next reply. See chat controls.

Drill 4: request a language without changing character identity

The language help article, edited November 5, 2025, says a requested chat language can take several prompts to take effect. It does not promise equal fluency or persistence for every language.

For a persistent writing preference, trial this original Custom Instructions block:

Write narration and dialogue in Spanish.
Keep established names and parcel identifiers unchanged.
Leave the traveler's words and choices to the user.

Use ordinary scene questions to inspect the resulting language, unchanged identifiers, and user agency separately. If you cannot assess the language accurately, record that limit. Do not rewrite a character's nationality or the user's Persona merely to change the reply language. This is text-chat practice, not audio pronunciation testing.

Drill 5: compare a revision fairly

Keep the exact original and revised field text in local notes. For each version, restore the same relevant card, Greeting, Persona, memory, settings, and test input in a separate comparable practice chat. A fresh chat alone does not prove every relevant input was reset. If a control or snapshot cannot be matched, record the difference.

Use the test worksheet. Keep all attempts, including failures and regenerations, and identify the accepted reply. A test that explicitly reminds the model of the desired behavior is useful as a separate stress case; it should not replace an ordinary-question baseline. Inspect unintended effects too: less repetition is not an improvement if the revision starts making the user's decisions.

Compare prompts on one selected model first. If comparing models afterward, hold the prompt fixed and record each actual selector label and displayed charge. The July 27 guide table and public chat page, accessed October 31, 2026, use different model inventories. Do not assume a renamed model, common backend, current account availability, or universal ranking from that discrepancy.

Know what your text changes

The company-hosted companion glossary, updated August 4, 2026 and retrieved with Firecrawl on October 31, distinguishes model input, context capacity, and persistent memory. Its vocabulary is cross-platform. It does not identify the backend or exact context assembly of the OD chat model you selected.

For this lab, inspect the text and controls you can actually maintain. A heading such as SYSTEM is a label inside your text, not evidence of an application message role. A character limit shown by a field is not a measured token allowance. A visible thought label does not hide the sentence from the model. Use formatting and character knowledge to test narrator-versus-character information explicitly.

Match a creator template to its tested scope

The Help Center-linked instruction-module document, freshly retrieved with Firecrawl on October 31, identifies Durante and credits Sir Trouble. Its author scopes it to Balanced. Treat its perspective and length choices as preferences to trial, and reconcile narrator/group alternatives before using them. A prompt template does not change the selected chat model. The document's body and directory attribution differ; see the source review before treating a linked title as authoritative.

Keep a revision record

Download the plain-text worksheet for local planning. Copy only the field payload you deliberately prepared into OD. Keep case results not run, blocked, inconclusive, observed pass, or observed failure, according to actual evidence. A successful document build verifies the library files, not these proposed chat outcomes.

Source access: October 31, 2026. The linked guide pages display July 27, 2026 as their modification date. The drills, Mira examples, and comparison procedure are authored teaching adaptations, not new claims about OD's internal implementation.