Voice typing for prompt engineering on Mac

THE CASE

Prompt engineering is voice-density work

Halopen is a native macOS dictation app for prompt engineers — hold the function key, speak the system prompt, the eval rubric, or the long-context spec, release; the verbatim text lands in the Anthropic Console, the OpenAI Playground, the ChatGPT and Claude desktop apps, Cursor’s chat panel, or any markdown file where a prompt lives. Cursor-context biasing handles model names like claude-opus-4-7 and snake_case hyperparameters in the right idiom.

Prompt engineering is one of the rare disciplines where the artifact and the process both reward more words. A good system prompt has more rules than a bad one. A good eval rubric names more dimensions and more pass/fail conditions than a bad one. A good long-context prompt names more constraints, more format requirements, and more negative instructions than a bad one. Density is the craft. The prompts that work in production are thousands of words; the rubrics that ship are dozens of paragraphs; the long-context prompts include the document, the task, the format, and the prohibition on every category of error the model has been seen to make.

That density cost something to write. Every additional rule in a system prompt is a sentence the engineer had to type. Every additional eval dimension is a paragraph. Every negative constraint in a long-context prompt is a clause that competes with the next clause for keyboard time. The friction shows up most clearly on the third revision — by the time you’re rewriting the system prompt for the third time, the marginal cost of a typed rule starts to feel like enough to skip the rule entirely. The model then makes the error the skipped rule would have prevented; you add the rule back; the loop is the rule, the typing tax, and the regression.

Voice removes the typing tax. Speaking the third revision of a prompt is no harder than speaking the first; the rules accrete; the rubric gets the dimension that catches the failure mode; the long-context prompt gets the negative constraint the model needed. The artifact gets denser without the writer paying the density tax. That’s the wedge for prompt engineering specifically — Halopen lets the prompts that ship be the prompts the engineer would have written if writing were free, and prompt-engineering rewards exactly that prompt.

SYSTEM PROMPTS

How to dictate a system prompt

System prompts have predictable structure. The role definition opens. The rules follow. The output format gets named. The examples land at the bottom. The negative constraints — the things the model should never do — sit between the rules and the examples or at the very end as a final guardrail.

Spoken, that structure becomes a paragraph rhythm. The role definition is one breath. The rules are a list of breaths, one rule per breath. The output format is one breath. The examples are one breath each. The negatives are one breath each. The whole prompt arrives at the cursor in the same shape the structure already has.

EVAL RUBRICS

How to dictate an eval rubric

Eval rubrics have one structure repeated per dimension: dimension name, what passing looks like, what failing looks like, an example of each. The repetition is the surface tax — typing the same structure five times for five dimensions is exactly the work that prompt engineers leave half-finished under typing cost. Voice closes the gap.

Pattern is dimension + pass criterion + fail criterion + example. Each dimension is one paragraph. A five-dimension rubric is five paragraphs and roughly two minutes spoken.

LONG-CONTEXT PROMPTS

How to dictate long-context prompts

Long-context prompts pair a document (the context) with a task (the prompt). The document is usually pasted; the prompt is usually typed. The prompt is where the constraint clauses live, and where typing cost compresses the prompt away from its full form. Voice keeps the prompt in its full form.

Pattern is task statement + format requirement + constraint list + negative constraints + tie-breaker. The tie-breaker is the clause that tells the model what to do when two interpretations of the constraint are both valid. It’s the half typed long-context prompts always skip; spoken, it costs nothing to include.

PROMPT ITERATION

The third-revision multiplier

The single biggest place voice pulls ahead in prompt engineering is the third revision. The first version of a prompt costs roughly the same to type or speak; the writer is composing the structure from scratch and most of the time goes into thinking, not into pressing keys. The second version is where typing starts to feel slow — every paragraph already exists in the previous version, and rewriting it just to add one new constraint feels disproportionate. The third version is where typed iteration usually compresses: the writer rewrites the changes as deltas, the prompt drifts toward incoherence, and the model’s behavior gets unpredictable in a way that’s hard to attribute to any single revision.

Voice keeps the third revision coherent. Each revision is a fresh take; speaking the prompt from scratch with the new constraint added is the same effort as speaking the first version was. The prompt that ships at revision twelve is the prompt the engineer would have written at revision twelve if writing were free. The constraint clauses that accumulated through eleven revisions all survive. The model reads the dense version.

The same multiplier shows up in eval rubrics. The first dimension is composition; the fifth dimension is repetition; the tenth dimension is repetition the writer didn’t want to do typed. Spoken, the tenth dimension costs the same as the first. The rubric ends up with ten dimensions instead of three, and the eval catches more failure modes because more dimensions are scored.

WHERE HALOPEN LANDS PROMPTS

Every prompt-input surface

Halopen lands text at the cursor through the macOS Accessibility API. Every prompt-input surface on Mac is fair game:

• The Anthropic Console system-prompt and user-prompt inputs in Safari, Chrome, Arc, or any Mac browser.
• The OpenAI Playground prompt inputs in Safari, Chrome, Arc, or any Mac browser.
• The ChatGPT chat input — both web and Mac-native desktop app.
• The Claude desktop app’s chat input.
• The chat panel in Cursor for prompt-iteration sessions against the model directly.
• The terminal where you run Claude Code or Aider for the agent-prompt patterns.
• The markdown file in your editor where the system prompt lives in source control.
• The eval framework’s rubric input — yaml, json, or markdown — wherever it’s edited.

The full directory of Mac AI coding surfaces Halopen lands prompts in sits at the hub. The pattern across all of them is the same: the input is a standard Mac text field; Halopen lands the verbatim text at the cursor; the agent or the model reads what was said.

UNDER THE HOOD

Why this works for technical vocabulary

Prompt engineering’s vocabulary is brutal: model names like claude-opus-4-7-20260201, hyperparameter names in snake_case, JSON Schema keys, OpenAPI fragments, regex patterns, the names of every tool the model can call. Halopen handles them through cursor-context biasing — the transcription engine reads the buffer of the current input and prefers tokens that match the casing and structure of the existing text. Model names already in the prompt land verbatim on the next dictation; new model names get caught by the live partial and spelled out in flight; snake_case stays snake_case; PascalCase stays PascalCase.

The full engineering walk-through — AVAudioEngine, CGEventTap, SFSpeechRecognizer, the transcription edge function, the live partial self-correction — is in the Claude Code flagship. The privacy posture — audio held only while the function key is held, never retained, every cloud call in the local audit log — is in the verbatim manifesto. The set of /for/ entities lists every surface this works in.

QUESTIONS

Frequently asked questions

Open the playground. Hold the function key. Speak the system prompt the way you’d brief a coworker.

The third revision is the one that ships. Voice is the modality that keeps the third revision coherent. The dense prompt, the multi-dimension rubric, the long-context spec with the tie-breaker clause — all of them are the artifact prompt engineering rewards, and all of them compress under typing cost. Halopen lands the dense version at the cursor.