Skip to content
Halopen
Get Pro · $19

Halopen for AI engineers

The best Mac dictation tool for AI engineers

AI engineering is half writing prose and half writing code. Halopen is the Mac voice layer that makes the prose half — system prompts, eval rubrics, agent specs, model docs — happen at the speed of speech, in every surface where the work lives.

Download Halopen See pricing

Free forever for the first 8,000 words a month · macOS 14.0+ · Apple Silicon & Intel

Why this fits

Halopen, paired with AI engineers.

Halopen is a native macOS dictation app built for AI engineers — verbatim system prompts, eval rubrics, and agent specs captured in the Anthropic Console, OpenAI Playground, Google AI Studio, Braintrust, LangSmith, Helicone, Promptfoo, Together Playground, Cursor, and Claude Code, with `claude-opus-4-7`-style model IDs and snake_case hyperparameters preserved. Hold the function key, speak the variation, release; the verbatim text lands at the cursor on Apple Silicon and Intel.

AI engineering is the discipline of building production systems on top of LLMs as primitives. The work is half code — pipeline glue, retrieval logic, queue management, observability — and half prose: the system prompts that ship to production, the evaluation rubrics that gate releases, the agent specs that define what gets built next, the model docs that explain to teammates why this prompt and not the previous one.

The prose half is where typing-bound throughput hurts most. A 600-word system prompt that takes 12 minutes to type takes 4 minutes to dictate. A 30-question eval rubric that takes an hour of typing takes 18 minutes of dictation plus a quick editing pass. Halopen lifts the typing tax off the prose half so the prose is as iterable as the code.

The compounding effect: AI engineers who add voice typing tend to ship more eval coverage, more variant tests, longer system prompts with more specificity, and faster doc cycles for teammates. The discipline's prose surfaces stop being the slow half of the work.

The workflow

How to use Halopen with AI engineers.

  1. 1

    Open the prose surface

    OpenAI Playground, Anthropic Console, Google AI Studio, your eval harness, your prompt-management tool, a Jupyter notebook, Notion for the agent spec, the README of an internal repo. Halopen works in every Mac text input.

  2. 2

    Hold the function key

    The recording pill appears. Halopen is listening. The cursor stays in the system-prompt field, the eval rubric cell, the spec doc, the agent description.

  3. 3

    Dictate the prompt or the spec

    "You are an assistant that triages incoming customer support tickets. For each ticket: identify the product surface (auth, billing, API, dashboard, mobile), assign a severity tier from one to four based on user-impact, extract any reproduction steps stated in the ticket, and propose a draft reply that acknowledges the issue and sets a realistic expectation. Output JSON conforming to the schema below; do not include prose outside the JSON. If the ticket appears to be a bug report, set escalate to true and include the suggested engineer assignee based on the surface mapping in the table below."

  4. 4

    Run the eval

    The full prompt lands at the cursor verbatim. Run it against your eval set; review the output; iterate. Voice version A → eval → voice version B → eval → variant rubric expansion → eval. The iteration count climbs.

  5. 5

    Document for teammates

    Same hotkey in Notion or your repo README. Dictate the why behind each prompt change, the eval results, the rationale for the production version. The doc gets written this week, not next sprint.

What matters for AI engineers

The Halopen features that earn their place.

  • Verbatim — your prompt design survives

    Halopen does not paraphrase. The "respond in JSON only" rule, the "always cite the source" constraint, the "if the diff touches billing, flag it as high-risk" conditional — all land in the prompt as you said them. Prompt design is a craft of specificity; the dictation layer treats it as one.

  • Works in every prompt surface on Mac

    OpenAI Playground (Chrome), Anthropic Console (Safari), Google AI Studio, your eval harness, your prompt-management tool, a Python REPL, a Jupyter notebook, the system-prompt field of any AI coding agent, Notion for the spec, your repo README — every one of them is a Mac text input. Halopen lands text in all of them.

  • Long-form holds — for the 600-word system prompt

    Continuous holds up to 10 minutes per take. A full production system prompt ships in one hold; a 30-question eval rubric ships in two; the agent-spec doc ships in three. The live preview shows the partial transcript as you speak.

  • Live preview catches misreads — model names, parameter names, JSON keys

    Custom model identifiers, internal tool names, parameter keys, JSON-schema field names — the live partial transcript shows what Halopen heard. Re-state or spell out anything that came through wrong; the correction replaces the misread.

  • Hold-to-talk — bounded audio for sensitive prompts

    The microphone is hot only while the configured key is held. No wake word. No always-on transcription. No surprise audio uploads. The audit log records every cloud call. Useful when system prompts contain proprietary logic or production-tied content.

  • Native Swift, idle in tens of megabytes

    Halopen runs natively on Apple Silicon and Intel. Idles in tens of megabytes with near-zero CPU — leaves resources for your eval harness, your Jupyter kernel, the agent processes you have open alongside.

A real Halopen session

A production system-prompt rewrite dictated into the Anthropic Console:

Halopen output

“"You are an assistant that drafts replies to incoming customer-support tickets for a Mac dictation product. For every ticket: classify the product surface from this set — onboarding, hotkey, accuracy, privacy, billing, sync, account, refund, other. Identify the user emotional tier from this set — calm, frustrated, urgent, churn-risk. Extract any reproduction steps the user mentioned. Identify whether the ticket contains a feature request, a bug report, both, or neither. Then draft a reply with these properties: open with one sentence acknowledging what the user said in their own words; provide a specific next-step they can take in two or three sentences; close with a one-sentence offer of follow-up. Use the operator voice — first person, contractions allowed, no marketing-speak, no AI-flavored language. If the ticket contains a refund request, route it to the human queue and don't draft a reply. Output JSON conforming to the schema below; do not include prose outside the JSON. If any field is unknown, set it to null."”

  • · 230-word production system prompt dictated in a single 80-second hold
  • · Multi-tier classification logic preserved verbatim
  • · Domain enums ("onboarding, hotkey, accuracy...") captured exactly
  • · Voice-style rule ("no AI-flavored language") preserved exactly as stated
  • · Voice version: ~80 seconds; typed version would have been ~5 minutes

Why Halopen

The dictation tool that earns its place.

AI engineering is bottlenecked at the prose surfaces. The system prompt, the eval rubric, the agent spec, the model card — all of them get under-specified when typing is the only path. Under-specified prompts produce under-specified behaviour; under-specified evals miss regressions; under-specified specs leave teammates guessing. The keyboard is silently shaping the production system.

Halopen is the calmest way to remove that constraint. Verbatim by default so prompt design survives. Long-form holds so a full production prompt ships in one take. System-wide so the same hotkey covers Playground, Console, eval harness, Jupyter, Notion, the README. The dictation layer earns its place by becoming invisible — what changes is that the prose surfaces stop being the slow half of AI engineering.

For AI engineers specifically, Halopen is the lowest-friction path to better-specified production systems. Voice typing isn't a productivity tool; it's a quality tool that happens to also be faster.

The editorial deep-dive

Voice typing for Claude Code on Mac.

The cost-of-articulation thesis, the verbatim wedge applied to AI coding, and how Halopen + Claude Code works under the hood — with two real prompt walkthroughs (M-E-R-I-A schema migration, Postgres EXPLAIN ANALYZE refactor).

Read the long-form essay →

Halopen for AI engineers — FAQ

Questions worth answering.

What's the best voice typing app for AI engineering on Mac?

Halopen. Hold-to-talk, verbatim by default, system-wide on macOS, long-form holds up to 10 minutes per take. Works in OpenAI Playground, Anthropic Console, Google AI Studio, every chat client, every eval harness, every AI coding agent, Notion, your repo README — every Mac text input. Free for the first 8,000 words a month; Pro is $19/mo or $179/yr.

How is AI engineering different from prompt engineering for voice typing?

Prompt engineering is the iteration discipline — variant testing, eval rubric design, prompt-versioning. AI engineering is the broader role: prompt engineering plus the production code, observability, retrieval logic, agent infrastructure, and team docs around it. Halopen serves both — for prompt engineering specifically see /for/prompt-engineers/.

Does Halopen preserve technical terminology across prompts and code?

Yes. Halopen biases the transcription engine with cursor-adjacent text and your active app context, so domain terms common to AI engineering — JSON schema, function-calling, tool use, RAG, embedding, completion, chain-of-thought, system prompt, few-shot, eval, observability — tend to land correctly. For unfamiliar names (custom model IDs, internal tool names), the live preview surfaces misreads before they ship.

Can I dictate the long production system prompts I write?

Yes. Continuous holds up to 10 minutes per take, which is enough for a 1,000-word system prompt or a 30-question eval rubric. The live preview shows the partial transcript as you speak so you can confirm the wording is landing as intended. Most AI engineers find a 600-word system prompt takes 4-5 minutes of dictation plus a quick editing pass.

Is Halopen safe for production-tied or proprietary prompts?

Audio leaves your Mac only while you hold the configured key, only to the transcription service, and only for the seconds you're holding it. Halopen does not retain audio. Halopen does not capture your screen. Halopen does not log transcripts. The audit log records every cloud call so you can verify the privacy posture against your team's production-data handling requirements.

Will it work with my existing AI engineering stack?

Yes — every component of a typical AI engineering stack is, fundamentally, a Mac text input. OpenAI Playground, Anthropic Console, Google AI Studio, LangSmith, Braintrust, Helicone, your custom eval harness, your prompt-management tool, your Jupyter notebook, your Notion docs, your repo README. Halopen lands voice-typed text in all of them with the same hotkey.

How much does Halopen cost?

Halopen Free is 8,000 words a month, forever — enough to dictate dozens of production prompts before the cap. Pro is $19/mo or $179/yr for unlimited words. Pro Lifetime is $499 one-time. 14-day no-questions refund.

More Halopen pairings

Halopen for whatever you use most on Mac.

See every Halopen pairing →

Power-user cheat sheet

Take Halopen with you when you work with AI engineers.

One short email, then the Halopen power-user cheat sheet — hotkeys, best-fit apps, custom vocabulary tips, voice patterns for prompt engineering. No spam. Unsubscribe in one click.

 

Try Halopen with AI engineers

Hold the function key. Speak.

Halopen Free is 8,000 words a month, forever. Open Halopen, hold the function key, and listen for what you sound like.

Download Halopen See pricing