Callsmith-skills

Callsmith

Give your coding agent a voice-app brief. Callsmith helps it choose a workable stack, make the important safety decisions, and write a build plan that another engineer can follow.

Install

npx skills add https://github.com/Nachi-Kulkarni/Callsmith_skills/tree/main/skills/callsmith

Start a new agent session and describe the software you want to build. Callsmith should load automatically when the task involves software that listens and speaks. Use /callsmith to force the main compile flow, or a command such as /callsmith audit to choose a specific playbook.

The skill works with Codex, Claude Code, Cursor, OpenCode, and other agents supported by the skills installer. The download is about 270 KB and contains no credentials, benchmark runners, or local project files.

What Callsmith does

Voice apps have a lot of small details that are easy to miss. Audio formats need to match. Someone must own interruptions. Recording needs a clear consent rule. Transfers and failed tools need a safe fallback.

Callsmith gives your coding agent checked details for 22 voice providers and a simple process for turning your brief into a build plan. It helps the agent:

• choose providers that can work together;
• spell out consent, data retention, and human handoff choices;
• describe the audio path and who owns each part;
• catch missing details and conflicting decisions before implementation.

Your agent still writes the application. Callsmith does not host calls, generate a framework project, or replace tools such as LiveKit and Pipecat.

flowchart LR
    A["Your idea"] --> B["Coding agent using Callsmith"]
    P["Provider details"] --> B
    F["Safety rules"] --> B
    B --> C["A clear build plan"]
    C --> D["Your voice app"]

A quick example

A clinic assistant may say that recording consent matters while leaving unsafe choices in its configuration:

{
  "surface": "inbound_pstn",
  "recording_consent": "none",
  "transcript_retention": "seven_days",
  "human_handoff": "ticket"
}

Callsmith makes those choices explicit and applies the project's safer medical defaults:

{
  "domain": "medical",
  "recording_consent": "announce",
  "transcript_retention": "thirty_days",
  "human_handoff": "transfer",
  "turn_gap_p95_target_ms": 900
}

The full clinic example includes the build plan and the matching machine-readable choices.

If you want to check those files yourself, install the optional CLI:

npm install -g github:Nachi-Kulkarni/Callsmith_skills
callsmith check --answers examples/clinic-triage/voice.answers.json
callsmith contract validate --file examples/clinic-triage/callsmith.recipe.md \
  --answers examples/clinic-triage/voice.answers.json --domain medical

What we have tested

Current models design voice agents well. Callsmith makes their designs verifiable.

Diagnostic numbers from live paired CSB runs (2026-08-15, harness prompt revision 2 + published interface): the same agent, same brief, same budget — once working normally, once with the Callsmith skill. Both arms receive the identical output interface (canonical enums, receipt schema); the skill is the only difference. Scores are deterministic gates on the artifacts, never a judge.

10–11 briefs per model · deterministic gates · paired arms — single run per scenario (diagnostic grade; 3-repetition runs with tightened intervals come before any product claim). OpenCode runs are diagnostic-grade isolation; Codex runs are publication-eligible. This measures design artifacts, not call quality, latency, uptime, or deployed cost.

Per-gate (both models, normal arm): reality traps 100%, stack physics 100% — the models are good. The lift lives entirely in floor completion and contract consistency, measured against Callsmith's own floors and contract schema (the benchmark scores conformance to this product's rules, not independently verified regulatory truth) — the two things the skill enforces. The frontier model gained more than the cheap one: strong reasoning writes confident designs, and nothing without a verifier makes it re-check them.

* one WITH arm was invalidated by the actor omitting its recipe file (correctly caught by validity gates), not by the skill.

Two further limits on this table: (1) both runs executed their two arms in parallel, sharing one model subscription — the repo's own publication standard now requires sequential arms because differential throttling between simultaneous arms is an uncontrolled confound, so these diagnostic runs would not be publication-eligible as run. (2) The lifts were measured before a follow-up interface fix (the receipt example's policy.basis enum discoverability, commit aa7bddf): the run report itself attributes part of BASE's contract-consistency failures to that gap, so a re-measured lift may be smaller. Treat the table above as an upper bound until the next diagnostic repeats it under sequential arms and the fixed interface.

Read the run report, test rules, and full limitations. Earlier (2026-07) diagnostic numbers published in this README were withdrawn — see the note atop the old report.

Things you can ask Callsmith to do

Optional CLI checks

callsmith packs
callsmith pack show twilio
callsmith pack validate
callsmith verify-packs
callsmith verify-packs --due --within 45
callsmith check --answers voice.answers.json
callsmith contract validate --file callsmith.recipe.md --answers voice.answers.json
callsmith doctor

The CLI checks provider data and build plans. It does not generate an app. Older generation commands such as init, forge, scaffold, and simulate have been removed.

Update, remove, or pin a version

# install for every supported agent found on this machine
npx skills add https://github.com/Nachi-Kulkarni/Callsmith_skills/tree/main/skills/callsmith -g

# update installed skills
npx skills update

# remove Callsmith
npx skills remove callsmith

# install this exact release
npx skills add https://github.com/Nachi-Kulkarni/Callsmith_skills/tree/v1.8.0-agent-compiler/skills/callsmith

Node 22 is recommended for the skills installer. The optional Callsmith CLI supports Node 18 and newer.

We tested a copied skills.sh installation in Codex and a clean install of the packed npm CLI. The repository also includes native manifests for Claude Code, Cursor, OpenCode, Pi, Kimi, Devin, Grok, and Antigravity. Those manifests are checked for structure, but we do not claim a full runtime test in every client.

Supported providers

• Phone providers: Exotel, Twilio, Plivo, Telnyx, and Vonage
• App frameworks: LiveKit, Pipecat, and custom FastAPI
• Realtime models: Gemini Live and OpenAI Realtime
• Speech and language providers: Deepgram, AssemblyAI, OpenAI, Anthropic, Google, ElevenLabs, Cartesia, and Sarvam
• Voice activity detection: Silero, WebRTC VAD, and Deepgram endpointing

Callsmith never makes up details for an unknown provider. New provider information must be added as a dated, sourced JSON file under providers/ and pass callsmith pack validate.

Limits

Callsmith is not a hosted voice platform, a legal certification, or a drag-and-drop bot builder. Its safety rules are conservative product defaults, not legal advice. Provider APIs also change, so production work should check current official documentation before shipping.

Project links

• How Callsmith makes product decisions
• How the skill works
• Build-plan format
• Safety rules
• Changelog
• Contributing
• Security policy
• Release process

Security reports should use a private GitHub advisory, not a public issue.

License

MIT