Skip to content

tough-tongue/toughtongue-skills / skills/ttai-agent

ttai-agent

Verified

Creates, 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/scenario/ai-instructions.md

View on GitHub
# scenario.ai_instructions

The agent's system prompt — the most important long-form field. This file is the
**shape**; a situation recipe fills in the content. Do not copy a recipe's
sections into every Scenario.

Write it with `ttai:create_scenario` / `ttai:update_scenario`
([workflow.md](workflow.md)). In the web app it is the AI Instructions field.

## Contents

- What it is
- Required shape
- Spoken lines
- Fundamental sections (not situation-specific)
- Voice rules that always apply
- Dynamic variables
- What to load next

---

## What it is

`ai_instructions` is the compiled brain of the scenario. The model sees it every
turn. Every extra sentence is a sentence it may ignore — be dense, structured,
and free of unresolved placeholders (except intentional `{{ dynamic_vars }}`).

Write enough to specify the role, reality model, flow, tools, boundaries, and
evaluation target. Do not use a word-count target: concise, unambiguous rules
outperform a long prompt with duplicated instructions.

## Required shape

- Use `##` headings. One idea per section. No essay walls.
- Write as **speech**, not as a document — this is spoken aloud.
- **One question per turn.** Never stack questions.
- Name the party's role clearly: who the AI is, who the human is.
- **Two beats then wait** (AI-led openings: caller, coach, interviewer, demo).
  Identity / reason / permission → STOP → branch on the reply. Do not collapse
  the opening into one breath. A roleplay counterpart never introduces itself:
  its opener is one bare line (see the sales-roleplay recipe).
- **Follow the thread.** If they say something interesting, pursue it — do not
  jump phases robotically.
- **Silence can be the job.** Meeting observers and notetakers stay quiet until
  addressed. Do not fill pauses.
- Every `{{ var }}` has a missing-value fallback: a `default:` filter or an
  `{% if var != blank %}` branch ([dynamic-variables.md](dynamic-variables.md)).
  An unset variable renders as empty text, so a bare line around it breaks.

## Spoken lines

- **Suggested phrasing goes in plain text.** No quotation marks, no code
  fences, marked `tone only`. The model rephrases a plain, labelled example;
  it recites a quoted or fenced one word for word, every session, and restarts
  it when interrupted.
- **Quote only what must be exact**, marked `word for word, because …` (a
  compliance disclosure, a client's tested script, a probe the rubric keys
  on). An unlabelled quote is read as verbatim; treat it as a bug.
- **Never quote the opening line.** Runtime greets with "your opening line"
  from FLOW when `skip_auto_start` is false. Write Beat 1 as a directive
  ("Start with: … Then STOP and wait."). Quoted openings restart on
  interruption. There is no `strategy.welcome_instructions` field — do not
  send one.
- **Anti-robotic rule.** Never repeat the same sentence twice; vary
  acknowledgments. Put one concrete varied-acknowledgment example in STYLE,
  labelled tone only — not a paragraph of examples.

## Fundamental sections

Use this skeleton. Rename or drop a section only when the situation file says
so. Do not add Cascade/STT blocks unless the model is Landmass `cascade`.

```
## CONTEXT
Who, company, product, why this conversation exists. Real names beat
invented ones. Template variables for per-session facts.

## ROLE
Who the AI plays, register (warm SDR, skeptical buyer, patient coach),
language + locale.

## FLOW
Ordered beats. First spoken turn (if skip_auto_start is false): a
directive — "Start with: … Then STOP and wait." Never quote the opening.
Then branches. What must never be skipped. One question per turn.

## TOOLS
When to call `end_session` (and any other enabled tools). Timing rules —
without them the agent never hangs up or hangs up mid-sentence.

## GUARDRAILS
Never reveal you are an AI unless asked. Never speak tool names. Never
invent facts that contradict CONTEXT.

## STYLE
Spoken cadence. Short sentences. Fragments fine. Never repeat the same
sentence twice; vary acknowledgments. One good/bad pair, labelled tone only.
```

Situation files add domain sections (outcomes, objections, rubric target) on top
of this skeleton — they do not replace it. Map `SESSION FLOW` / `YOUR ROLE` to
FLOW / ROLE if you see those names in an older scenario.

Before drafting, load [authoring.md](authoring.md) to define the human outcome,
reality model, difficulty, and evidence of success. The skeleton is the
structure; that contract supplies its content.

## Voice rules (every spoken scenario)

- Plain text in the model's **output**. No markdown, bullets, or URLs in what
  will be spoken.
- Spell out numbers and symbols when they will be read aloud.
- Realtime (Galaxy, Ocean, Landmass `gemini`/`openai`) hear audio. Landmass
  Cascade sees **transcripts** — only then load
  [../recipes/cascade-tts.md](../recipes/cascade-tts.md) (STT errors, SSML,
  beats).

## Dynamic variables

`{{ var_name }}` is filled per run: `?t_var_name=` on the run link (the web app
otherwise prompts the user), or `dynamic_vars` on calls and meeting bots. The
field runs full Liquid, so a missing value takes a `default:` filter or an
`{% if %}` block, and one Scenario can branch on a variable (persona level,
language, offer). Patterns, limits, and a worked example:
[dynamic-variables.md](dynamic-variables.md). Prefer `appearance.voice` for
Realtime. Cascade `tts_voice_id` lives in
[model-selection.md](model-selection.md).

## Load next

- Model / voice stamp — [model-selection.md](model-selection.md)
- Strategy, tools, appearance — [control.md](control.md)
- Variables and branching — [dynamic-variables.md](dynamic-variables.md)
- Cold call / SDR — [../recipes/cold-call.md](../recipes/cold-call.md)
- Sales roleplay — [../recipes/sales-roleplay.md](../recipes/sales-roleplay.md)
- Coaching — [../recipes/coaching.md](../recipes/coaching.md)
- Demo — [../recipes/demo.md](../recipes/demo.md)
- Interview — [../recipes/interview.md](../recipes/interview.md)
- Meeting facilitation —
  [../recipes/meeting-facilitation.md](../recipes/meeting-facilitation.md)
- Landmass Cascade speech —
  [../recipes/cascade-tts.md](../recipes/cascade-tts.md)
- Create / edit / refine procedure — [workflow.md](workflow.md)

## Key Files

- [authoring.md](authoring.md) — design contract
- [model-selection.md](model-selection.md) — model stamps
- [control.md](control.md) — strategy and tools
- [workflow.md](workflow.md) — create/update procedure
- [dynamic-variables.md](dynamic-variables.md) — variables and branching
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: 472kUpdated Claude CodeCodexVerifiedInstall
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: 323kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 323kUpdated Claude CodeCodexVerifiedInstall
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 availableStars: 323kUpdated Claude CodeCodexVerifiedInstall