DOCUMENTATION

Frequently Asked Questions

Do skills require a pre-configured Cookbook to execute?

No. Each skill adheres to a self-contained execution contract. If brand context (COOKBOOK.md or saved state) is missing, the skill defaults to Setup Mode, prompting the user interactively for required business parameters (ICP, core differentiator, offer, target channel). When context is present, the skill automatically skips discovery and executes in Apply Mode.

What are the execution modes (Setup, Apply, Tune, Validate)?

Skills dynamically resolve into four discrete operating modes based on the provided input state:

  • Setup Mode: Triggered when required context is absent. Interactively queries the user to establish baseline variables.
  • Apply Mode: Default production mode. Ingests approved brand context and applies it directly to the target deliverable without interactive setup.
  • Tune Mode: Triggered when supplementary data or revised evidence is injected. Adjusts specific sections of an existing asset while preserving approved positioning.
  • Validate Mode: Evaluates draft copy or positioning against approved brand rules, flagging hallucinations, unverified claims, or tone deviations.
How and where is brand context stored?

All brand context generated via the web builder is stored client-side in browser localStorage. No proprietary database stores your business context. You can export this context at any time as a raw, standardized COOKBOOK.md file to commit into your local codebase, .cursorrules, or CLAUDE.md.

Which AI agents, IDEs, and LLMs are supported?

Skills are distributed as model-agnostic Markdown specifications with standardized prompt contracts. They are tested and optimized for:

  • CLI & Code Agents: Claude Code, Cursor, Windsurf, GitHub Copilot.
  • Hosted Model Interfaces: Claude 3.5 Sonnet, GPT-4o, Gemini 1.5 Pro.
  • Custom Agent Pipelines: Ingestible as system prompt instructions via any standard LLM API.
Does Cookbook trigger automated actions or publish campaigns?

No. Cookbook workflows are strictly sandboxed and read-only. Skills produce structured text deliverables (briefs, copy, frameworks, analyses) for human review. They do not have external API side-effects, they will not publish to social channels, trigger email sends, manipulate ad spend, or mutate live systems.

What input formats can skills ingest for research and analysis?

Skills accept standard structured and semi-structured text inputs:

  • Plain text, raw interview notes, and call transcripts (.txt, .vtt, .srt).
  • Tabular data (CSV, TSV, Markdown tables).
  • Structured payloads (JSON, API responses, analytics readouts).

Each skill validates source metadata and input completeness before processing, flagging missing variables or ambiguous samples.

How does the system prevent hallucinations and unverified claims?

Skills enforce strict evidence-traceability guardrails. In data-synthesis workflows, statements must tie directly to user-provided source transcripts or verified metric banks. The prompt contracts explicitly instruct models to report contradictions, treat low-sample feedback as qualitative outliers rather than consensus, and never invent testimonials or performance metrics.

How should teams resolve conflicts between automated website extraction and human-approved context?

Context extracted from a public URL is always treated as an uncommitted draft. Automated inferences must never overwrite approved brand definitions. Once a human reviews, edits, and saves the draft, the state locks as the canonical source of truth until explicitly updated via Tune Mode.

Can our team create and distribute proprietary internal skills?

Yes. Any Markdown file following the standard frontmatter schema (name, category, version, required_inputs, deliverable_schema) can be hosted in an internal Git repository or loaded directly into local workspace configs.