Prompt triwizard

Prompt Triwizard

Writes, refines, and evaluates LLM prompts (system prompts, user prompts, few-shot examples, and tool-call schemas) for agent pipeline nodes — producing high-quality, style-correct output every time.

---

Reuse Within a Session

Once this skill has been loaded, its guidance (structural hierarchy, principles, anti-patterns) stays in the conversation context. When the same prompt needs another modification within the same session (e.g. a re-fix request after testing), continue applying this guidance without re-invoking the skill.

---

Step 1 — Read the Context, Deliver in Kind (Common Step)

Prompts usually live as string literals inside code files (most often .py, but it varies — .ts, Jinja templates, etc.). Locate the prompt first; grep for the prompt text if you're unsure which file holds it.

Read only:

• The target prompt file
• Sibling files holding related prompts (other nodes of the same pipeline)
• Schemas, Pydantic models, or tool definitions the prompt feeds into

Don't read a file just because it shares the directory. Skip .md/.json/.yaml/config unless it's the schema or tool definition the prompt feeds.

Goal: understand the project's existing depth, tag conventions, and schema bindings — this shapes how specialized the new prompt should be.

Deliver in kind: match the file/format the original prompt lived in (most often .py, but it varies — .ts, Jinja templates, etc.); if there's no existing file to match (Mode 1), default to .py — use .ts or Jinja instead only if the project already shows that convention elsewhere. When the prompt takes variables, wrap it as a function call (Python/TS) or a render() call (Jinja) rather than a bare string. Name the system and user pieces — whether a constant or a function — so the literal tokens system_prompt and user_prompt appear in the name (e.g. SYSTEM_PROMPT, build_system_prompt, USER_PROMPT, build_user_prompt).

---

Step 2 — Receive the Style and Language

Style is the primary axis of this skill — it determines the structural format of the prompt you produce, and everything downstream (Step 3's use-case content, Mode 1/2/3's output) is written in the style chosen here. The output language determines what language the prompt is written in.

Style

Priority order:

1. User explicitly names a style → use it
2. Code read in Step 1 shows a model string (claude-*, gemini-*, gpt-*) → offer it as a suggestion, then confirm
3. Ask: "Which format should I use? Anthropic / Google / OpenAI. I'll default to Anthropic if you skip."
4. No response → use Anthropic as the default

If sibling prompts already exist in the project, ask: "Should I also convert the existing prompts to the chosen style?"

Language

Priority order:

1. User explicitly names a language → use it
2. No explicit language → detect the language the user is writing in, then ask: "Should I write this in [detected language] or English?"
3. No response → use the detected language as the default

---

Step 3 — Identify the Use Case

Use case is a secondary axis, layered on top of the style chosen in Step 2: style controls how the prompt is formatted (tags, headers); use case controls what additional instructional content the prompt needs (persistence rules, null-case handling, grounding policy, etc.) beyond the style's own baseline. Determine the use case from what the node actually does:

Use case	Signal
Conversational	Multi-turn dialogue, persona/tone, no fixed completion point
Single-shot transform	Classification, extraction, summarization — one input, one output, done
Agentic tool-use	Calls tools in a loop, reasons between calls, has a stop condition
RAG grounding	Context/documents already injected; answers in one shot from what's provided
Deep research	Calls search tools iteratively across multiple angles before answering

Priority order:

1. User explicitly names a use case → use it
2. Code read in Step 1 shows the shape (multi-turn loop, single function call, tool-call loop, injected context, iterative search) → infer it, then confirm
3. Ambiguous → ask before proceeding

A node can combine traits from more than one use case (e.g., a conversational agent that also calls tools). When it does, read every applicable usecases/ reference and merge their instructional content — don't force a single label where the task genuinely spans two.

---

Step 4 — Identify the Mode

Mode	Triggered by	Output
Mode 1 — Create	Request for a new prompt	Full system + user prompt from scratch
Mode 2 — Refine	Request to improve, fix, or rewrite an existing prompt	Before/after diff with diagnosis
Mode 3 — Eval	Request to test or measure a prompt	Eval report with scores and regression check

---

Step 5 — Apply References

Read the references below in the order indicated. They are located in the references/ directory beside this file, split into common/ (shared quality backbone), styles/ (fixed set of 3 format styles — the primary axis, from Step 2), and usecases/ (what additional instructional content the prompt needs, by agent/task purpose — the secondary axis, from Step 3).

Reference	Read when
common/core-principles.md	All modes, always — quality backbone
common/anti-patterns.md	After drafting — self-check before delivery
common/eval-rubric.md	Mode 2 (diagnosis) and Mode 3 (eval loop)
styles/anthropic-claude.md	Style = Anthropic
styles/google-gemini.md	Style = Google
styles/openai-gpt.md	Style = OpenAI
usecases/conversational.md	Use case = Conversational
usecases/single-shot-transform.md	Use case = Single-shot transform
usecases/agentic-tool-use.md	Use case = Agentic tool-use
usecases/rag-grounding.md	Use case = RAG grounding
usecases/deep-research.md	Use case = Deep research

Each usecases/ file includes a per-item placement table showing which style container that content belongs in — read the style reference first, then cross-check the usecase file's placement table against it rather than treating the two as independent lookups.

---

Mode 1 — Create a New Prompt

1. Clarify requirements: task, inputs, outputs, tool calls (if any), output schema
2. Read core-principles.md → choose the right technique (zero-shot, few-shot, CoT, ReAct)
3. Read the chosen style reference (from Step 2) → this is the format the prompt will be written in
4. Read the use-case reference(s) from Step 3 → determine what additional instructional content the prompt needs, and place each piece per its placement table for the chosen style
5. Self-check with anti-patterns.md
6. If the prompt feeds a Pydantic model or tool schema: verify field descriptions are in sync

Deliverable: Full system prompt + user prompt template, labeled clearly — delivered per the Step 1 rule.

---

Mode 2 — Refine an Existing Prompt

Refine covers two distinct jobs. They share the same diagnosis, then split:

• Path A — In-place — the prompt is already in the target style; improve its content without restyling.
• Path B — Re-style — the prompt is in a different style (or styleless), and you carry it into the target style.

A re-style is never a marker swap. Each style encodes a different philosophy — Anthropic's motivation-over-negation and positive framing, Google's RULE N routing with [REQ]/[OPT] severity, OpenAI's outcome-oriented output contracts. Carrying a prompt across styles means re-expressing its intent in the target philosophy, which changes wording and structure by design. And the original may be rough: when it is, improve the content while you restyle — the two jobs merge.

Common diagnosis (both paths)

1. Read the existing prompt in full → detect its current style. Check exclusive, structural markers, not just "has XML tags" (multiple styles use backtick-wrapped XML tags):
   • **FOLLOW-UP RULES** / *RULE N: NAME* bold+italic pattern anywhere → Google
   • A single XML tag wraps the entire system prompt, closing right before the user turn (which is itself then wrapped in a task-specific tag) → Google
   • Bare # (H1) headers structure the top level, with no enclosing wrapper; any XML tags present are supplementary and plain, never backtick-wrapped → OpenAI
   • Many independently-named XML tags scattered as top-level siblings, with any Markdown headers nested inside a specific tag's content (never floating between tags) → Anthropic
2. Read eval-rubric.md → score on 4 axes: Clarity / Specificity / Actionability / Scope
3. Read core-principles.md → identify which principles are violated
4. Read the chosen style reference (from Step 2) in full
5. Read the use-case reference(s) from Step 3 → check whether required instructional content (persistence, null-case handling, grounding policy, etc.) is present and correctly placed for the chosen style

Choose the path

• Current style == target style → Path A
• Current style != target style (or styleless) → Path B
• If it's unclear whether the user wants a restyle or only a content fix in the existing style, ask before proceeding.

Path A — In-place refine

Present diagnosis and changes using the before/after format:

❌ Original:
[exact original text]

🤔 Diagnosis:
[which axis fails, and why]

✅ Improved:
[exact replacement text]

Rule: One change per block. Multiple changes in separate blocks.

Path B — Re-style refine

The prompt is rewritten whole, so block-level diffs don't apply. Instead:

1. Re-express the prompt in the target style's philosophy, not just its markers (reasoning induction, priority markers, data placement, rule structure all convert).
2. Preserve the original's core intent and hard constraints — verify nothing load-bearing was dropped in translation.
3. If the original is underspecified or low-quality, improve the content as you convert (apply core-principles.md).
4. If the original's intent is ambiguous, do not guess — confirm with the user first.
5. Present as a single whole before/after, plus a short conversion note explaining which philosophy shifts you applied:

❌ Original ([source style]):
[full original prompt]

🤔 Diagnosis + conversion notes:
[axis scores, principles violated, and how the target philosophy was applied]

✅ Rewritten ([target style]):
[full rewritten prompt]

Self-check (both paths)

6. Self-check with anti-patterns.md

---

Mode 3 — Eval Loop

1. Define the baseline:
   • New prompt → baseline = no system prompt (raw model)
   • Refined prompt → baseline = the version before the change
2. Design the input set: minimum 3 scenarios — typical / edge / adversarial
3. Score each scenario on the 4-axis rubric from eval-rubric.md
4. Check for regressions: re-run all previously-passing scenarios
5. Report: trigger-fit accuracy + output quality scores + regression status

Fresh session rule: always evaluate in a session that has not seen the prompt in conversation history.

---

Self-Check Before Delivery

After drafting, always run through anti-patterns.md. Key questions:

• Is every rule verifiable (binary or measurable)?
• Is any rule here to prevent a failure mode I've never observed?
• Are volatile values at the end of the system prompt or in the user turn?
• Do few-shot examples cover edge cases, not just the happy path?
• Am I using the correct structural markers for the chosen style — not just "has XML tags" (all three styles can contain XML tags; the real signals are a single system-prompt-wide wrapper + RULE N for Google, bare H1 headers with no wrapper and only supplementary plain XML for OpenAI, and many independent sibling tags with headers nested inside them for Anthropic)?
• Does the prompt contain the instructional content its use case requires (e.g. a stop condition for agentic/deep-research, a null-case rule for single-shot transform, an insufficient-context rule for RAG grounding), each placed in the container its style's placement table specifies?
• Are tool/schema descriptions synchronized with the prompt?
• If writing to a file, does the format match the project's existing convention — wrapped as a function/render call when parameterized?
• Do the system and user constants/functions literally contain system_prompt / user_prompt?