Learn Cue · context and coding workflows
Use Cue Voice with Claude Code, Codex and Your Task Context
Speak the intent, supply the evidence, and give the right Agent a bounded task. Cue's contribution is the voice entry point and task interface—not a promise that every Agent shares the same tools or history.
By the Cue product team · Updated . Based on Cue source and a read-only client review. These are practice recipes, not recorded workflow results or speed benchmarks. Use the controls available in your installed version.
Voice input, an Agent and a model do different jobs
1. Dictate into the tool you already use
Use Cue Dictation in an editable prompt field in Claude Code or Codex, where your installed version and input permissions support insertion. Speak a draft, check the text, then submit it yourself. Cue is helping you enter the request; the receiving app manages its own Agent, context, tools and account. Start with the voice typing exercise in a harmless note before using a coding prompt.
2. Select an external Agent inside Cue
In Cue versions with the Agent selector, Claude and Codex appear as local Agent choices. They need a working local connection and any sign-in required by that Agent. Seeing an entry does not mean it is connected. Select the Agent first, then one of the models and controls that it actually offers.
3. Use Cue Agent with a supported model
Choose Cue in the Agent column, then choose an available model. Selecting a Claude model here is not the same as running Claude Code. It changes the model for Cue's Agent, not the entire execution environment. A model name does not automatically grant a terminal, email access, a repository or another app's subscription.
Prepare the Agent before handing it a real task
- Start a new, non-sensitive practice task. Do not experiment in a conversation that is changing production files.
- Open the Agent/model control near the composer. In the reviewed Cue UI, “Choose agent, model, and reasoning” separates Agent, Model, and Reasoning / Speed.
- Check connection readiness. If Claude or Codex needs setup, use the explanation shown by Cue. Review any proposed install or repair command and complete the provider's sign-in yourself when required. Do not paste credentials into a chat prompt.
- Check the selected Agent and model again. The menu stores the selection for the current task; availability can vary by account and connector. If a model is locked or absent, use an available choice or resolve access first.
- Give only the necessary context. For code, identify the intended repository and relevant files. For a summary, supply the short source text. Do not assume a coding Agent received everything visible on your screen.
- Request a draft or inspection first. Ask the Agent to restate the task and what it can access. Approve implementation only after that matches your intent.
For the standalone Claude Code or Codex route, keep using that tool's normal project and approval controls. For the Cue-integrated route, follow the local connection and permission state shown in Cue. The two routes are alternatives, not steps that must both run.
Turn a meeting recap into a brief that an Agent can use
After a meeting, a short spoken recap can separate decisions from loose discussion. This exercise uses fictional notes; it does not require recording anyone.
Mina will draft 12 survey questions by 14 October. Rowan will review them by 16 October. We discussed three possible launch dates but chose none. We still need an owner for the accessibility review.
Step 1 — capture: dictate that recap into an empty note. Check the people, dates and “chose none” before asking an Agent to transform it.
Step 2 — structure: supply the checked text to Cue Agent and speak this request:
Use only these notes to prepare a handoff brief with decisions, tasks, open questions and missing information. For each task list owner, deliverable, deadline and its supporting source sentence. Keep unspecified owners and dates unknown. Draft only; do not create tickets or calendar events.
Step 3 — verify: the brief must preserve 12 survey questions, both named owners and their dates; leave the launch date undecided; and flag the missing accessibility owner. “Three possible dates” must not become a committed launch date. This is an illustrative success checklist, not a recorded Cue answer.
Step 4 — hand off one bounded action: after review, give the relevant Agent one task from the brief, plus its source. Do not send the entire meeting transcript by default. Ask it to restate the target and missing information before taking action.
If you need to capture a meeting as it happens, use the separate meeting-notes workflow, with participant consent and the recording controls in your installed app. A generated summary is not a substitute for checking the original record. After-meeting recap is a different, shorter workflow.
Use voice to describe a bug, then give the coding Agent evidence
Speech is useful for describing intent and reproduction steps. Paste exact file paths, error messages and identifiers as text: retyping them from a voice transcript can introduce ambiguity. Use the same task brief whether the recipient is Claude Code, Codex or Cue Agent with authorized code tools.
A concrete example: a stretched video thumbnail
Work in a disposable demo repository or a branch you control. Provide the actual component path and the image's dimensions. The example below deliberately asks for inspection before edits:
In this repository, the video thumbnail stretches vertically at a narrow width. Inspect the component and its styles, then propose a fix that preserves the image ratio at 320, 390 and 1280 pixels. Keep the copy, price and download behavior unchanged. Identify the nearest browser regression test. Do not edit files or run commands yet; first tell me what evidence you need.
After the scope is correct, authorize the bounded change in a follow-up: “Add the regression test, demonstrate the failure, make the smallest layout fix, and run the relevant checks. Show the diff and test results. Do not deploy.”
Useful follow-ups you can speak
- “Keep that solution. Explain why the existing test missed the bug.”
- “Read the failing test output. Separate the observed failure from your hypothesis.”
- “Summarize the changed files, checks run and remaining risks for another reviewer.”
The workflow is consistent with Claude Code's guidance to provide specific context and a way to verify work, and OpenAI's task guidance on clear instructions and useful context. These sources describe their own tools; they are not evidence of a Cue integration test.
Manage context with a short, inspectable handoff
Keep a task brief you can read and correct. This is a user-managed working document, not a claim that Cue automatically shares all conversation history between Agents.
Goal: [one result].
Source: [specific notes or file paths].
Confirmed facts: [facts with their source].
Constraints: [what must stay unchanged].
Current state: [what was actually completed].
Validation: [checks and observed results].
Open questions: [unknowns].
Next action: [one bounded step requiring approval if consequential].
Speak the goal and constraints, then verify exact identifiers. At a task boundary, ask the active Agent to draft this handoff; compare it with the real files or destination app before using it.
When changing Agents in a task that already has content, use an explicit new task or the supported fork/handoff flow in your version. Preserve the original conversation. Give the receiving Agent the checked brief and only the source files it needs. Ask it to state what context it received; hidden state, permissions and past tool results should not be assumed to transfer.
For an email task, keep the recipient and approved facts in the brief. For meeting notes, keep undecided items separate. For code, keep the repository, changed files and real test results. Remove obsolete assumptions instead of accumulating unrelated requests in one conversation.
Where GPT-6 and Claude Fable 5.1 fit
A newer model is a candidate for a task, not a replacement for the source brief or verification. Provider specifications are documented separately for GPT-6 Astra and Claude Fable 5.1.
Availability check, September 7, 2026: the reviewed Cue client listed Fable 5.1 under Cue models, with a Cue Plus upgrade control. GPT-6 availability in that account was not confirmed. A model listing is not an end-to-end execution test, and provider availability does not establish Cue account access.
- Select the Agent first: Cue, or an available connected external Agent.
- Search that Agent's current model list for the desired name or ID. If it is absent or locked, do not assume it will be selected by an “Automatic” option.
- Start with the model's default reasoning and speed controls offered by that Agent. Not every provider supports the same settings.
- Run the same harmless task with the same supplied context. Compare factual corrections, scope compliance and usable output before changing more settings.
For the recap above, count wrong or invented facts. For the coding task, check the diff and test results. If comparing speed, include transcription correction and review time—not just answer-generation time. This guide is not a benchmark and does not claim that one model is faster or better in Cue.
Cue is free to start; Cue Plus is $19.99/month. Check the current plan and the selected provider's access requirements. Do not assume a Cue plan includes an external Agent subscription, or that its model choices match Cue's Hosted list.
Recover without losing context or repeating an action
My spoken request has the wrong file name or date
Correct the text before submission. For exact identifiers, paste the value from its source and keep voice for the explanation. If already submitted, stop the run if appropriate and clarify the specific correction before authorizing more work.
Claude or Codex appears but cannot run
Read the connection or authentication message. Check local setup and the selected Agent, then use the recovery option actually shown by Cue. Do not switch to a different model silently or claim the requested Agent completed the task.
The next Agent does not know what happened
Supply the verified handoff brief and relevant source. Ask what it can access before continuing. Keep the original task; do not delete it as a way to fix missing context.
A task failed after making a partial change
Cancel is not an undo operation. Inspect the actual destination before retrying, and ask only for remaining work. If an email, ticket or file might already exist, resolve that uncertainty before repeating the original request.
Is voice helping? Try a small repeatable task using typing and voice. Record total time, corrections, clarification turns and whether the result passed the same checks. Keep the workflow if it reduces effort without losing accuracy; this is your own comparison, not a claimed Cue benchmark.
Start small, then connect the workflow
Begin with the dictation tutorial and the first Voice Agent task. Once those work, try the fictional recap and review every fact before handing it off.
Then choose a concrete task: draft a project follow-up email, extract source-linked actions from meeting notes, or give a coding Agent a reproducible bug report. Each exercise includes source material, a reusable request and checks for the result.
Ready to work with code? Follow a small bug fix with Cue Voice and Claude Code, then use a bounded Codex code review to inspect a candidate patch. These are disposable practice exercises with supplied code and verification steps, not recorded Cue runs. Keep implementation and release approval separate.
For setup help or corrections, contact Cue or email eli@sophoninc.com. Include the Cue version, selected Agent and error message with personal data and credentials removed. Review privacy and data handling before supplying sensitive context.