calle-script-advisor
Draft and check CALL-E phone-call task text and result schemas for clarity, safety, and extraction quality before any call is placed.
- License 许可
- MIT
- 最近更新
- 2026-10-08
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- calle-script-advisor
- license
- MIT
- description
- Draft and check CALL-E phone-call task text and result schemas for clarity, safety, and extraction quality before any call is placed.
02 资源文件 5
- 参考文档references/examples.md4.8 KB
- 参考文档references/safety.md3.9 KB
- 参考文档references/script-patterns.md3.4 KB
- 脚本scripts/check-call-script.mjs18.6 KB
- 脚本scripts/check-call-script.test.mjs11.4 KB
03 SKILL.md 正文
CALL-E Script Advisor
CALL-E's task field is free-form natural language, and its quality determines whether a call succeeds. A vague task produces a vague conversation, low extraction confidence, and an unusable result - and by the time that is visible, the call has already been placed and paid for. This skill turns CALL-E's own API guidance into an executable check that catches that before the call happens.
When To Use
Use this skill whenever you are drafting or reviewing a CALL-E task and result_schema, before any call is placed. That includes writing a new call from scratch, adapting an existing task for a new scenario, and reviewing a task or schema someone else wrote.
The Drafting Procedure
Before writing any text, gather:
- The goal - why this call is happening, in one sentence.
- The single decision the call must establish - the one thing that determines what happens next. If there seem to be several, pick the one that matters most and treat the rest as optional context.
- Who is calling and why - the identification line the agent should open with: who they are, who they are calling on behalf of (if anyone), and the reason for the call.
- How to close the call - the instruction that ends the conversation cleanly once the ask is resolved.
- What to do on voicemail - whether to leave a message, what it should say, and whether to try again.
Then, in order:
- Write the
tasktext: identification, purpose, the single ask, edge-case handling, and the close. Seereferences/script-patterns.mdfor the anatomy and a before/after rewrite. - Write the
result_schemato capture exactly the decision the call must establish - not everything that was discussed, just the answer this call exists to produce.
The Checking Step
Run scripts/check-call-script.mjs against the task and schema, and iterate until there are no errors:
node scripts/check-call-script.mjs --task-file <path> --schema-file <path>
node scripts/check-call-script.mjs --task "..." --schema '{...}'
Both arguments are optional independently - lint a task alone, a schema alone, or both. The script prints findings grouped by severity, a score out of 100, and exits 1 if any error-severity finding exists, so it can gate a workflow.
Errors are not negotiable - fix every one. Warnings are judgement calls: read the suggestion, decide whether it applies to this call, and either fix it or consciously accept it. A warning left in place should be a deliberate choice, not an oversight.
See references/examples.md for three complete worked examples that pass the linter with zero errors.
Rules That Are Not Negotiable
- Never invent a phone number. Use only the number the user provided or that the calling system resolved.
- Never infer region, locale, or timezone from a phone number.
- Never solicit sensitive personal data: Social Security numbers, credit card or bank account numbers, CVVs, PINs, passwords, dates of birth, or a mother's maiden name. See
references/safety.md. - Always include an
unknown(orunclear/not_stated/undetermined) enum member for anything the call may fail to establish. CALL-E's own docs recommend this explicitly, and it is what lets a downstream integration tell "the answer is no" apart from "there was no answer" - the siblingplugins/zapier-callerefuses to mark a call actionable when a required field comes back with one of these values, so the enum member is not decoration. - Ask the schema for its own evidence. Beside any field an automation will act on, add a
<field>_quotestring described as the recipient's exact words that establish the answer above. It costs nothing - the extraction model is already reading the transcript - and it is the only thing that separates an answer drawn from the call from one drawn from thin air, since both arrive as equally well-formed JSON with equally high confidence. The siblingplugins/zapier-calleverifies that quote against the transcript and sends the call to human review when it appears nowhere the recipient spoke. Ask for a full spoken sentence rather than a word: a one-word quote can match by coincidence. List the_quotefield inrequiredalongside the field it supports. A quote the model may omit is a check that switches itself off on exactly the call that had no evidence to offer, since a missing key is nothing for a verifier to examine; requiring it turns a silent pass into a loud failure. - Mark a field
requiredonly when the workflow genuinely cannot proceed without it.requiredis a contract: an integration that checks it will send the whole call to human review when the field is absent or unknown. Requiring a nice-to-have field turns every partially-successful call into manual work. - Never let the script imply a commitment the caller cannot honour - no promised callback times, approvals, or guarantees the agent is not authorized to make.
Read references/safety.md before drafting any task that touches consent, disclosure, calling windows, or a do-not-call request.
Progressive Disclosure
This file stays focused on the workflow. Detail lives in references/:
references/script-patterns.md- the anatomy of a good task, with a before/after rewrite.references/safety.md- consent, disclosure, sensitive data, calling windows, and do-not-call handling.references/examples.md- three worked examples (appointment confirmation, on-call acknowledgement, lead qualification), each verified against the linter.
04 同仓库更多 Skills
xitter
Interact with X/Twitter via the x-cli terminal client using official X API credentials. Use for posting, reading timelines, searching tweets, liking, retweeting, bookmarks, mentions, and user lookups.
writing-plans
Use when you have a spec or requirements for a multi-step task. Creates comprehensive implementation plans with bite-sized tasks, exact file paths, and complete code examples.
whisper
OpenAI's general-purpose speech recognition model. Supports 99 languages, transcription, translation to English, and language identification. Six model sizes from tiny (39M params) to large (1550M params). Use for speech-to-text, podcast transcription, or multilingual audio processing. Best for robust, multilingual ASR.
weights-and-biases
Track ML experiments with automatic logging, visualize training in real-time, optimize hyperparameters with sweeps, and manage model registry with W&B - collaborative MLOps platform
webapp-testing
Toolkit for interacting with and testing local web applications using Playwright. Supports verifying frontend functionality, debugging UI behavior, capturing browser screenshots, and viewing browser logs.
web-pentest
Authorized web pentest: recon, proof-based exploits, report.