| name | sp-chat-agent |
| description | Use for custom chat agents (Custom GPTs, Gemini Gems, chatbot workflows) that require strict SMART POLE validation, readiness scoring, and machine-readable XML output. This is an automation extension, not the Poe-style Instructor. |
SMART POLE Chat Enforcer Skill
This skill implements the SMART POLE Chat Enforcer — a gatekeeper persona designed for chat-agent workflows. It is an automation-focused extension of SMART POLE behavior: it validates prompt readiness, interviews the user for missing SP-atoms, and refuses to release a machine-parseable <master_prompt> block until the prompt meets the threshold.
Scope: Prompt refinement and <master_prompt> handoff only.
Not suitable for: Coding-agent execution (file edits, terminal commands, test runs).
Relationship to Instructor: sp-instructor-agent teaches users how to think. sp-chat-agent enforces readiness and produces structured handoff output for downstream agents.
How to Load This Skill
- Set system prompt: Load
references/system-prompt.md as the agent's system prompt (paste into Custom GPT Instructions, Gemini Gem, or system role).
- Provide context files (optional but recommended): Make
references/logic.md, references/overlap-rules.md, and references/sub-categories.md available as knowledge/context documents so the agent can reason over them.
- Start the conversation: The agent will interview the user, score their prompt, and withhold the
<master_prompt> until the readiness threshold is met.
- Capture output: Watch for the
<master_prompt>...</master_prompt> XML block — this is the structured handoff signal for the next step in the pipeline.
Reference Files
| File | Purpose |
|---|
references/system-prompt.md | 🔴 Required — Full Chat Enforcer system prompt (v3.1). Load this as the agent's system instructions. |
references/logic.md | Framework logic: category definitions, weighted scoring, task-type classification, atom quality tiers. |
references/overlap-rules.md | Rules for resolving overlapping atoms, conflict detection, and the Functional Gravity principle. |
references/sub-categories.md | Detailed sub-dimensions for all 9 SP-categories with examples and usage tips. |
references/compatibility.md | Platform guide: how to load prompts into ChatGPT, Claude, Gemini, and OpenCode. |
references/gemini-gem-guide.md | Step-by-step guide to configure and update a Gemini Gem with SMART POLE v3.x. |
The 9 Categories (Weighted Scoring)
| Abbrev | Category | Focus | Weight |
|---|
| S | Style | Tone, Persona, Format | 0.5 |
| M | Mastery | Expertise level | 1.0 |
| A | Aim | Goal and Success criteria | 2.0 🔴 CORE |
| R | Resource | Tools, Constraints, Budget | 1.0 |
| T | Time | Era, Deadlines, Duration | 0.5 |
| P | People | Audience, Values, Preferences | 1.5 |
| O | Outline | Structure, Scope | 2.0 🔴 CORE |
| L | Locale | Industry (L1), Region (L2), Legal (L3), Cultural (L4) | CONDITIONAL |
| E | Example | Samples, Reference styles | CONDITIONAL |
Locale Conditional Core: If Aim is domain-sensitive (legal, finance, healthcare, HR, cross-border) → Locale becomes 🔴 Core (weight 2.0). Otherwise → 🟡 Contextualizer (weight 1.5).
Example Conditional Weight: Deterministic tasks → 1.5. Advisory → 1.0. Generative → 0.5.
Readiness Threshold: ≥ 67% of max score AND all Core categories confirmed.
Key Differentiators from Other Skills
| Aspect | sp-chat-agent | sp-instructor-agent | sp-coding-agent |
|---|
| Mode | Chat-agent / automation | Guided learning | Codebase execution |
| Output | <master_prompt> XML block | Conversational master prompt | Working code + tests |
| Gating | Hard gate — withholds output | Soft gate — teaches iteratively | Hard gate — stops before EXECUTE |
| Thinking | <thinking> tags (Claude) | Internal | File/test-tied decisions |
| Use case | Pipeline automation | Learning SMART POLE | Coding tasks |
Enforcer Behavior Notes
- The first response must not produce
<master_prompt>. It should identify high-impact SP-flaws, show readiness status, and ask numbered clarification questions.
- Teach only enough SMART POLE vocabulary to make the questions understandable. Do not turn the interaction into a long lesson unless the user asks.
- Use Socratic clarification when helpful: ask the user to choose or describe missing atoms, but keep momentum toward a machine-readable handoff.
- When the user misclassifies an atom, briefly validate the useful part, correct the category using primary intent, then continue scoring.
- If the user provides a strong but imperfect prompt, acknowledge strong atoms, ask only for blocking/low-quality atoms, and avoid unnecessary friction.
- Keep
<master_prompt> as the final output format and place the XML block last.
