FIELDWORK / 20 ADDITIONAL FULL GUIDES
The last mile of a useful conversation.
Go beyond the successful demo. Inspect the small decisions that make an interface understandable, an avatar interruptible, and a runtime accountable.
Choose the problem in front of you
A route, not a reading backlog.
Can a person understand—and correct—the next turn?
Start with discoverable capabilities, then inspect correction, streaming, time and language. Each guide leaves you with a concrete review artifact.
- 01
Foundations / 4 MIN
Conversational onboarding: preview the work before asking for trust ↗
Help a first-time user discover a relevant, bounded task and leave or change course without losing control.
- 02
Knowledge & actions / 4 MIN
Editable confirmation summaries: corrections must reopen the right decisions ↗
Let people inspect and correct proposed changes without a stale confirmation authorizing a different result.
- 03
Conversation design / 4 MIN
Accessible streaming responses: make partial output interruptible and readable ↗
Present streamed assistant output so keyboard and assistive-technology users can follow, stop, and revisit a stable final record.
- 04
Knowledge & actions / 4 MIN
Dates and time zones in booking conversations: confirm the local appointment, store the unambiguous fact ↗
Avoid bookings that shift at daylight-saving transitions or misunderstand an ambiguous date and local time.
- 05
Evaluation & operations / 4 MIN
Multilingual evaluation fixtures: test the interface, not only translated sentences ↗
Create a repeatable review set for localized conversation behavior, including right-to-left layouts and locale-dependent data.
What survives when the character cannot carry the interaction?
Follow speech repair into public kiosks, training scenarios, authored character boundaries and graceful fallback. Presence should serve the task, not obscure its state.
- 01
Conversation design / 4 MIN
Names and numbers in speech: repair identifiers before they become actions ↗
Collect spoken names, codes, and numbers without turning a plausible transcript into a mistaken identity or transaction target.
- 02
Embodied systems / 4 MIN
Kiosk session reset and privacy: make the next visitor truly new ↗
Design a public or shared conversational device so one visitor’s information cannot persist into the next visitor’s experience.
- 03
Embodied systems / 4 MIN
Embodied training scenarios: design roleplay that can be reviewed and transferred carefully ↗
Create an inspectable practice scenario for a conversational avatar without fabricating educational efficacy.
- 04
Embodied systems / 4 MIN
Character content boundaries: keep canon separate from session memory ↗
Let writers and designers govern a character’s claims without letting invented session details become permanent world facts.
- 05
Embodied systems / 4 MIN
Avatar fallback quality ladder: degrade presence without losing the task ↗
Keep an embodied conversation usable when visual fidelity or audio availability changes, without restarting the task or pretending the avatar is still functioning normally.
Which event owns the work—and how long can it wait?
Trace the first audible response, assemble streamed calls safely, reconcile delivery, bound retries and assign the floor in a shared room.
- 01
Evaluation & operations / 4 MIN
Latency trace budget: measure the first usable response ↗
Measure when a person can first use a response and identify the owned stages that precede it.
- 02
Knowledge & actions / 4 MIN
Streaming tool-call assembly: execute only a finished call ↗
Assemble streamed call fragments without executing partial, cancelled, or cross-turn arguments.
- 03
Evaluation & operations / 4 MIN
Webhook delivery reconciliation: separate receipt from business outcome ↗
Handle duplicate, delayed, and out-of-order events without losing the difference between event delivery and domain truth.
- 04
Evaluation & operations / 4 MIN
Retry and timeout budgets: preserve one deadline ↗
Choose retry behavior from operation semantics and remaining time rather than a copied universal setting.
- 05
Embodied systems / 4 MIN
Multi-user avatar turn ownership: the room needs an authority ↗
Keep shared-room speech, captions, actions, and permissions attributable to a room participant rather than an animated surface.
Would the answer still deserve trust when the easy case disappears?
Inspect cached state, document boundaries and deliberate non-answers. Build synthetic fixtures honestly, then rehearse an exit from the provider.
- 01
Knowledge & actions / 4 MIN
Conversation cache invalidation: freshness is scoped ↗
Cache only values whose audience, authorization, version, and freshness rule are explicit.
- 02
Knowledge & actions / 4 MIN
Corpus chunking evaluation: test boundaries before tuning retrieval ↗
Choose chunk boundaries with controlled evidence fixtures instead of assuming one size or overlap works for every corpus.
- 03
Evaluation & operations / 4 MIN
Retrieval abstention test set: make no-evidence measurable ↗
Release grounded answers only after testing questions that should remain unanswered, conflicted, or unsupported.
- 04
Evaluation & operations / 4 MIN
Synthetic conversation test data: useful fixtures without invented users ↗
Create clearly fictional, nonidentifying fixtures while measuring what their construction fails to represent.
- 05
Foundations / 4 MIN
Provider exit portability: rehearse the reversible path ↗
Preserve owned records, documented tool contracts, and asset rights so leaving a provider is a tested transition rather than an emergency assumption.
Put the runtime reading to work
Trace the pause.
Explore a serial model from the end of speech to the first audible response. Change the inputs, see the contribution of each stage, and keep the model’s limits in view.
What this edition can—and cannot—claim.
These 20 pieces extend the existing 56-guide edition. Hubs, tools, cards and outlines are not included in that count. Each guide has a distinct reader need, a worked example, an original conceptual diagram, a practical review table and claim-scoped primary-source notes.
Prepared and source-checked 19 September 2026. No historical publication date, independent product test, customer result or interview is implied. Worked examples and decision procedures are editorial synthesis, not proof of vendor performance. Source imagery elsewhere on the site remains local-review-only pending owner clearance.
Read the earlier 20-guide resource edition ledger · Inspect the source library · Download this edition’s coverage record