tough-tongue/toughtongue-skills / skills/ttai-agent
ttai-agent
VerifiedCreates, edits, and refines Tough Tongue AI Scenarios, the definitions behind voice agents (sales calls, lead qualification, demos, screening, support, booking) and practice roleplays (negotiations, interviews, coaching). Writes ai_instructions, ai_model_config, strategy, tools_config, and rubrik through ttai:create_scenario and ttai:update_scenario. Also resolves the personal or organization workspace, checks plan and balance, places or cancels SIP calls, schedules or cancels meeting bots, mints private Scenario links, grounds a Scenario in the user's documents with knowledge FileBases, and troubleshoots ttai MCP sign-in. Use when the user wants a voice agent or practice Scenario created or improved ("create a voice agent", "the agent ends calls too early", "make it sound more natural"), a call placed, a meeting bot sent, or a Tough Tongue AI account question answered. Not for reports across many sessions (ttai-session-analyst) or browser demo steps (ttai-browser-demo-builder).
references/recipes/sales-roleplay.md
View on GitHub# Sales Roleplay Scenario Patterns
Rules for scenarios where the **AI plays a prospect** and the user practices
selling or negotiating.
## Contents
- Call types (cold / warm / discovery)
- Resistance pattern and guarded disclosure
- Breakthrough conditions and core objections
- Early termination rules and end outcomes
- ai_instructions and user_instructions structure
- Rubric (evaluates the REP)
- Technical config and quality checklist
Key facts:
- Uses Galaxy `medium-stable` by default — Realtime pipeline, native Vertex
voice, no TTS/STT. See
[../scenario/model-selection.md](../scenario/model-selection.md).
- Three call types: Cold Call, Warm Intro, Discovery Call — each has different
opening behavior and termination risk. The prospect's first line is one bare
line (Hello, who is this? on a cold call), never an introduction
- Pitch reception first: two or three how-does-it-work questions before any
objection; guarded, but still talks like a person
- Enforce strict resistance: ONE objection at a time, 2-3 follow-ups each
- Example lines in this recipe are tone only: write them unquoted in
`ai_instructions` and label them so
([ai-instructions.md](../scenario/ai-instructions.md#spoken-lines))
- Define breakthrough conditions, core objections, and early termination rules
- Always enable the `end_session` tool
- Three outcomes: CONVINCED, NOT READY, FIRM NO
---
## Call Types
| Type | Opening (one bare line, then wait) | Risk | Trust |
| -------------- | ---------------------------------- | ------ | ------- |
| Cold Call | Hello, who is this? (guarded) | HIGH | 2-3 min |
| Warm Intro | Hi, how can I help? (neutral) | MEDIUM | 1-2 min |
| Discovery Call | Hi! Ready to discuss... (engaged) | LOW | Minimal |
Risk = termination risk; Trust = time to build trust.
The prospect's opener is one short line and never an introduction, a reason for
the call, or a permission ask — those are the rep's moves, and the rubric scores
the rep on them. The two-beat opener in
[ai-instructions.md](../scenario/ai-instructions.md) is for AI-led calls and
does not apply here. The rep opens the call (`skip_auto_start: true`); the
prospect's line is its reply. Write it unquoted in FLOW, marked tone only, or
`word for word` when the exact line matters.
Specify the call type in `ai_instructions` — it determines opening behavior and
termination rules.
---
## Resistance Pattern — CRITICAL
AI naturally wants to be helpful. Sales prospects are NOT helpful. Enforce
strict resistance:
- **ONE objection at a time** — never combine multiple concerns
- **2-3 follow-up questions per objection** — test the salesperson's
understanding
- **Accept solid answers** — acknowledge when concerns are genuinely addressed,
then move on
- **Reject vague claims** — "Trust me" → "I need specifics"; "Industry leading"
→ "What numbers?"
```
### Pitch Reception (before any objection)
- Listen to the pitch. Ask two or three questions about how it works and
whether it is proven before raising any objection.
- Only after that, move to the first objection.
### Resistance Pattern
- NEVER combine multiple objections - focus on ONE issue at a time
- For each objection, ask 2-3 follow-up questions to test understanding
- Accept reasonable arguments backed by specifics, examples, concrete plans
- Don't accept vague claims, but DO accept solid responses - move to next concern
- Guarded, not mute: answer in real sentences with in-character detail. Never
stonewall with one-word answers.
- Every refusal has an in-character reason (budget, a spouse to consult, a past
failure), never a bare "I can't say"
```
---
## Discovery Process — Guarded Disclosure
Make salespeople WORK for context. Never dump information unprompted.
### Phase 1: Vague initial responses
- User: How many customers do you serve daily?
- AI: Quite a few, it's a busy shop
- User: Can you give me a number?
- AI: Maybe 100-150? Not sure exactly
### Phase 2: Progressive detail
Answer specific questions with specific info. Naturally embed uncertainty: "I
think", "maybe", "not sure exactly". Sometimes self-correct: "Actually wait,
maybe closer to 20 per day."
### Phase 3: Ambient context
Mix answers with real-world interruptions and context:
- Handle a customer: "Sorry, one minute — [handles customer]"
- Weather / environment: "Hot today, makes people thirsty"
- Family: "My wife usually handles that, I'd have to check"
### Anti-pattern: info dumping
- BAD — dumps everything unprompted: I have 150 customers daily, 40-50
remittances, BRILink for 3 years...
- GOOD — requires targeted extraction: User: Tell me about your business / AI:
I run an FMCG shop, been here 5 years / User: How many transactions daily? /
AI: Maybe 100-150 customers? It varies
```
### Discovery Process - Realistic Information Disclosure
- NEVER dump all details at once - require targeted questions
- Answer vaguely first, specific only when asked precise follow-ups
- Mix answers with small talk about [shop/office/family context]
- Make user work to uncover problems through progressive questioning
- Sound human: "I think", "maybe", "not sure exactly", "let me remember"
```
---
## Breakthrough Conditions
The prospect becomes receptive ONLY when ALL conditions are met PLUS urgency:
```
### Breakthrough Conditions
You become receptive ONLY when ALL [3-5] are proven:
1. [Primary value proven with customer-specific math]
2. [Key fear addressed with concrete safety plan]
3. [Reliability guaranteed with examples]
PLUS urgency (any one):
- Limited-time pricing / Seasonal deadline / Competitive pressure
```
### Examples by domain
**B2B Sales**: ROI proven with customer-specific math (their numbers, not
generic), integration plan for current systems, support guarantees (SLA,
replacement). Urgency: Q4 budget deadline, competitor launching soon.
**B2C Sales (diet/fitness)**: Past struggles validated without blame, realistic
timeline (2-3 kg/month), health-focused beyond weight, maintenance plan.
Urgency: discount ending, event in X weeks.
**Compliance-sensitive**: Legal compliance confirmed with specific regulations,
safety/privacy plan with concrete isolation, proof from similar risk-averse
customers. Urgency: regulatory deadline, seasonal peak.
---
## Core Objections
Pick 5-7 total, use 2-4 per conversation. Categories:
1. **Price/Value** — "Too expensive", "Hidden costs?"
2. **Risk/Safety** — "What if it fails?", "Is this compliant?"
3. **Integration** — "How does it work with my current system?"
4. **Proof/Credibility** — "Show me examples", "Who else uses this?"
5. **Timing** — "I'm happy with current solution", "Maybe later"
Lead with the persona's PRIMARY fear. Follow with practical concerns.
When the user provides real call transcripts or lost-deal notes, extract the
actual objections verbatim and use those instead of generic ones — that is what
makes the practice transfer to the real conversation.
---
## Early Termination Rules
```
### Early Session Termination Rules
**STRICTLY ENFORCED:** You MUST END SESSION if conversation becomes unproductive.
After [timing by call type], assess quality.
**Must END immediately if:**
- Repeats same points without new info
- Ignores your concerns or is pushy/dismissive
- Generic pitch not tailored to your situation
**How to end:**
- "[Context-appropriate dismissal]"
- Then follow the live `end_session` turn rule; never call it in the same
turn as the dismissal
```
---
## End Outcomes (Exactly 3)
| Outcome | Condition |
| ------------- | ----------------------------------- |
| **CONVINCED** | All breakthrough + urgency met |
| **NOT READY** | Missing 1-2 key conditions |
| **FIRM NO** | Poor pitch or termination triggered |
Behavior per outcome:
- **CONVINCED** — express interest, ask about next steps.
- **NOT READY** — politely decline, leave the door open.
- **FIRM NO** — direct rejection, then follow the closing turn rule.
---
## ai_instructions Structure
```
## ROLE
- [Persona details, background, current situation]
## Current situation
- [Call Type]: A sales rep is [calling/visiting] you to pitch [Product]
- FLOW first turn: Start with a guarded greeting (call-type table), one bare
line, no introduction. Then STOP and WAIT. Do not quote a canned line.
## BEHAVIORAL GUIDELINES: MOST IMPORTANT TO FOLLOW
### Core Principle / ### Discovery Process / ### Pitch Reception
### Resistance Pattern
### Breakthrough Conditions / ### Core Objections (One at a Time)
### Early Session Termination Rules
## Conversation Flow
### Opening / ### Engagement Progression (Phase 1-3) / ### Decision Outcomes
### Your Background Context
- ONLY reveal when explicitly asked
```
## user_instructions Structure
```
**Your Goal:** [Clear objective]
## Winning Strategy (X-Y minute pitch)
### 1. Opening (timing) / ### 2. Discovery / ### 3. Value Pitch
### 4. Address [Primary Fear] (CRITICAL) / ### 5. Handle Objections
### 6. Create Urgency / ### 7. Close
## What Scores High Points
## What Kills the Deal
```
---
## Rubric Standard
Evaluates the **SALES REP** (the human user) — not the AI prospect.
Every rubric uses this structure: each stage scored **1-10** independently, each
stage has a **weight** signaling evaluation depth, overall score = weighted
average displayed as X/10.
```
# [Scenario Title] — Evaluation Rubric
**Overall Score: Weighted average of all stage scores (each 1-10)**
## Stage 1: [Opening] (Weight: 30%)
Score this stage 1-10. Focus: [what to evaluate]
**1-3 (Poor):**
- [observable failure criteria]
**4-6 (Acceptable):**
- [observable decent-attempt criteria]
**7-8 (Good):**
- [observable solid-execution criteria]
**9-10 (Excellent):**
- [observable textbook / above-and-beyond criteria]
## Stage 2: [Objection Handling] (Weight: 40%)
Score this stage 1-10. Focus: [what to evaluate]
...
## Stage 3: [Closing] (Weight: 30%)
Score this stage 1-10. Focus: [what to evaluate]
...
```
### Standard stage breakdowns
Stages in order, with weights:
- **2-5 min pitch** — Opening Hook (30%) → Engagement & Objections (40%) → Close
(30%).
- **Gatekeeper + DM** — Gatekeeper Nav (20%) → DM Qualification (25%) +
Articulation (30%) → Close (25%).
### Leniency rules (build into every rubric)
- "Send me info + call back next week" = NOT a failure — score **5/10 Close**.
- "Never argued with an objection" earns minimum **5/10 on Objections**.
- Calls cut short by a genuinely busy prospect should not penalize dimensions
the rep never got a chance to demonstrate.
- Textbook call that books the meeting = **8-9/10**. Perfect with creative
extras = **10/10**.
---
## Technical Config
```json
{
"ai_model_config": {
"provider": "Galaxy",
"model": "medium-stable"
},
"strategy": {
"skip_auto_start": true,
"system_instructions_template": "minimal",
"silence": {
"silence_threshold": 6000,
"end_session": false,
"force_agent_to_speak": false
},
"conductor": {
"enabled": true,
"messages": [
{
"time_seconds": 700,
"message": "Move toward a decision — CONVINCED, NOT READY, or FIRM NO. State the close; wait for the rep's reply before end_session.",
"end_turn": true
}
]
}
},
"tools_config": {
"tools": {
"end_session": {
"should_register": true,
"add_to_system_prompt": true,
"tool_settings": { "disconnectDelaySeconds": 5 }
}
}
},
"session_analysis": { "is_auto_analysis": true, "is_auto_submit": true },
"appearance": { "voice": "Puck" }
}
```
Note `skip_auto_start: true` — the sales rep (the human) opens the call and the
prospect's bare line is its reply. For a scenario where the prospect must speak
first, set it to `false`; the same bare line is Beat 1 in FLOW.
`force_agent_to_speak: false` — after a silence, a small model writes the nudge
in character. The fixed check-in line (`true`) pulls the prospect out of
character; keep `true` for AI-led callers, not for roleplay counterparts.
Galaxy `medium-stable` uses native Vertex voice — no TTS/STT/LLM sub-fields. Set
`appearance.voice` to choose the voice persona.
---
## Quality Checklist
- [ ] Call type specified (cold / warm / discovery); the prospect's first line
is one bare line, no introduction
- [ ] Pitch reception: two or three how-does-it-work questions before the first
objection
- [ ] Resistance pattern: one objection at a time, follow-ups per objection;
guarded but still talks like a person
- [ ] Example lines unquoted, labelled tone only or word for word
- [ ] Guarded disclosure rules present
- [ ] Breakthrough conditions (3-5) + urgency triggers defined
- [ ] Core objections (5-7 defined, 2-4 used per session) from real context
where available
- [ ] Early termination rules with end_session instruction
- [ ] Exactly 3 outcomes: CONVINCED / NOT READY / FIRM NO
- [ ] Background context marked "reveal only when asked"
- [ ] Rubric evaluates the REP with weighted categories
- [ ] `end_session` enabled; `force_agent_to_speak: false`; conductor timer
matches expected call length
## Key Files
- [../scenario/authoring.md](../scenario/authoring.md) — realism and difficulty
principles
- [../scenario/model-selection.md](../scenario/model-selection.md) — model stamp
- [../scenario/control.md](../scenario/control.md) — strategy / tools
- [../scenario/rubric.md](../scenario/rubric.md) — evaluation design
- [../scenario/ai-instructions.md](../scenario/ai-instructions.md) — prompt
shape
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |