agent-prompt-engineering 0.2.0

Deterministic prompt builder on pydantic-stack-core (every prompt = a MetaStack of RenderablePieces) + cognition chains + optional grammar gate — the shared base of the chaincompiler ecosystem.

---

name: prompt-engineering description: The Prompt Engineering System (APE) — a deterministic prompt builder on pydantic-stack-core. A prompt is a MetaStack of RenderablePiece "blocks" (typed, self-rendering pydantic models); the same blocks render the SAME prompt every time. Compose blocks into stored "prompt templates". Use it to make system prompts, personas, rule blocks, few-shot blocks, cognition chains, or any prompt. No model call, no randomness. when_to_use: Building or reusing prompts/personas/templates programmatically; you want the same spec to render the same prompt every time; making a new prompt template (folder of block specs).

Prompt Engineering System (APE)

A skill (this SKILL.md + the code beside it) that turns data → a prompt, deterministically, on the pydantic-stack-core foundation. The model decides what blocks to write; this engine guarantees how they render — same inputs, same output, every time. That reliability is the point: an agent reads this skill and uses it to make prompts/personas the same way each time.

The foundation rule: every prompt is a MetaStack of RenderablePieces. A block is a real, typed, self-rendering pydantic model (RenderablePiece subclass) — not an exec'd code string. (This supersedes the old {values, methods:[code-string]} mechanism; see META-APE-ARCHITECTURE.md §1.)

Vocabulary

• block — a RenderablePiece: typed fields + a pure render() → a chunk of prompt. Concrete blocks ship in blocks.py (Text, Heading, Section, Role, BulletList, KeyValue, Fenced, Template, Group). Declarative form: {"type": <BlockName>, **fields}.
• prompt — a MetaStack of blocks, built with prompt(...) / render_prompt(...).
• prompt template — a stored, ordered set of block files (a folder under templates/) → one prompt.

There is no baked-in structure and no "levels." You bring your own via blocks and templates.

Use it (the package prompt_engineering)

# pip install -e .   (deps: pydantic-stack-core, rulecatcher, agent-skilltree)
from prompt_engineering import Role, BulletList, render_prompt, render, render_template, list_templates

# 1) compose blocks (RenderablePieces) into one prompt — a MetaStack
render_prompt(
    Role(role="a senior Python code reviewer", mission="find correctness bugs"),
    BulletList(title="Rules", items=["Cite file:line.", "Never invent an API."]),
)  # → "You are a senior Python code reviewer. Your mission is to find correctness bugs.\n\n## Rules\n- ..."

# 2) render a single block from a declarative spec (data → prompt, no exec)
render({"type": "Template",
        "template": "# {role}\n\nYou are the {role} of {domain}.",
        "values": {"role": "Worldsmith", "domain": "otters"}})   # same spec → same prompt, always

# 3) render a stored prompt template (its blocks composed in order)
list_templates()               # ['five_level', 'persona', 'simple_agent']
render_template("five_level")  # → one prompt with WORLD/EGREGORE/DEITY/PROGENITOR/AGENT sections

Define your own block by subclassing RenderablePiece and (optionally) registering it for declarative use:

from prompt_engineering import RenderablePiece, register_block

@register_block
class Callout(RenderablePiece):
    emoji: str
    text: str
    def render(self) -> str:
        return f"> {self.emoji} {self.text}"
# now usable directly or as {"type": "Callout", "emoji": "⚠️", "text": "..."}

Add a prompt template — DROP A FOLDER

templates/
  {template_name}/
    template.json   # {name, description, blocks: ["block1.json", "block2.json", ...]}
    {block}.json    # a declarative block spec: {"type": <BlockName>, **fields}

Create the folder, write template.json, drop in the block specs. Nothing in the engine changes. list_templates() / render_template(name) pick it up by convention.

Resource files (referenced by relative path)

• prompt_engineering/blocks.py — the block library (RenderablePiece subclasses on pydantic-stack-core)
  • prompt() / render_prompt() + declarative block_from_dict() / register_block() / REGISTRY
• prompt_engineering/prompt_engine.py — render, render_file, list_templates, load_template, render_template, render_template_blocks (single block + stored templates)
• examples/ — single-block specs for different block types (system/role, persona, rules, few-shot, constraints, output-format) and full personas — read these to learn the declarative block format
• templates/five_level/, templates/simple_agent/, templates/persona/ — worked prompt templates
• prompt_engineering/cognition.py — the cognition-chain blocks (AttentionChain / ChainOfReasoning / SkillChain, all RenderablePieces; a CoR nests an AC) + validate_chain (a no-dep broken-detector)
• prompt_engineering/gate.py — the REAL gate: gate(text, exemplars) LEARNS the grammar from exemplar chains → violations + fixes + the orthogonal/syntax_break (steerable/fatal) verdict
• prompt_engineering/grammar.py — the persistent rulecatcher seam (learn/gate/render on a Connection)
• prompt_engineering/skillpack.py — write_skill / cohere_skills: a markdown body → a publishable <name>/SKILL.md (skilltree-aware)
• pydantic-stack-core — THE foundation (RenderablePiece + MetaStack); rulecatcher — the grammar engine the gate runs on; agent-skilltree — coordinate-addressed skill placement
• prompt_engineering/persona.py — make_persona(spec) / make_persona_file(path): a full persona prompt (block types + a cognition DSL block) from a small spec
• scripts/ — make_prompt.py, make_persona.py, new_template_skill.py (scaffold a skill for a template)
• test_*.py — run to verify (all NO-API): prompt engine · cognition · persona · gate · skillpack

Cognition chains & personas

cognition.py gives the three chain primitives as RenderablePiece blocks: an attention chain, a chain-of-reasoning (which nests an AC = block-nesting), and a skill chain. Rendering never touches a grammar engine or a DB. The gate is a separate, optional layer (gate.py, on rulecatcher): gate(text, exemplars) LEARNS the grammar from exemplar chains (next_kind = shape, next_token = vocabulary), then lints a candidate → violations + fixes + the orthogonal (in the language, wrong slot → steerable) vs syntax_break (foreign → fatal) verdict.

from prompt_engineering import render_cor, gate, make_persona

render_cor("Debugger", ["Symptom", "Repro", "Hypothesis"], "Localize", "the root cause")  # a CoR persona
gate("[Goal] ⇒ [Repro] ⇒ |Localize|",                                                     # learn, then lint:
     exemplars=["[Symptom] ⇒ [Repro] ⇒ |Localize|", "[Symptom] ⇒ [Bisect] ⇒ |RootCause|"])
# → {ok: False, verdict: 'orthogonal', violations: [{found:'Goal', candidates:['Symptom'], ...}]}

make_persona({                                   # a whole persona with a complex reasoning DSL, in one call
  "name": "Ada", "role": "a senior debugger",
  "reasoning": {"foci": ["Symptom", "Repro", "Bisect"], "held": "RootCause", "decision": "the minimal fix"},
  "rules": ["Reproduce before theorizing."], "output": "A markdown report.",
})

Why deterministic matters

Each block's render() is a pure function of its typed fields; a MetaStack joins the rendered pieces in order with a separator — pure composition, no model call. That's what lets an agent rely on it to produce prompts consistently across runs. This is the base layer of the chaincompiler ecosystem (CC → ACCC → CORCC → SCCC all import DOWN to here; see META-APE-ARCHITECTURE.md).