rp1-run/prompt-writer

Prompt Writer

Guide for writing maximally terse agent prompts with built-in constitutional governance and epistemic stance. Combines compression-by-default authoring (tersify), governance primitives (constitution), and epistemological foundations (epistemology) into a single progressive-disclosure skill.

When to Use

• Creating new agent specifications or skill prompts
• Writing command/skill prompts with governance built in
• Drafting instruction sets for AI agents
• Applying constitutional or epistemic review to existing prompts
• Running the six-stage prompt pipeline via /build-prompt

Core Principles

Structure-First: Start with skeleton, not prose. Choose section pattern, fill minimally.

Compression-by-Default: Every word must earn its place. Remove before adding.

Preserve Exactness: Normative language (MUST/SHOULD/MAY), literals, constraints stay verbatim.

Governed-by-Design: Constitutional primitives and epistemic stance are selected upfront, not bolted on.

Skill Files

Reference Layers

| File | Purpose | When to Load | |------|---------|--------------| | references/tersify.md | Compression discipline, section patterns, abbreviation policy, style rules, validation checklist | Direct /prompt-writer invocation; prompt-validation pipeline stage | | references/constitution.md | 10 governance primitives (anti-loop, output discipline, role, scope limits, orchestrator purity, error degradation, exploration bounds, anti-bias, truth constraints, transition guards) + four agent-type profiles | Constitutional-checklist pipeline stage; standalone constitutional review | | references/epistemology.md | Six epistemic stances (Fallibilist Empirical, Interpretivism, Phenomenology, Constructivism, Pragmatism, Compare-Mode) with composable contracts | Epistemic-stance pipeline stage; standalone epistemic review | | references/budget.md | 15% aggregate governance cap, per-stage budget allocations, line classification rules, enforcement protocol | Pipeline execution (Stage 6 budget enforcement); standalone budget review |

Companion Files

| File | Purpose | When to Load | |------|---------|--------------| | PATTERNS.md | Reusable prompt patterns: constitutional agent, interview, map-reduce, state machine, tool selection, error hierarchy, output contract, progressive disclosure, checklist gate, parameterized behavior, engineering discipline | Direct /prompt-writer invocation; prompt authoring | | TEMPLATES.md | Ready-to-use prompt templates at simple, moderate, complex, command, and tool/skill levels | Direct /prompt-writer invocation; prompt authoring |

Pipeline Stages

Six stages executed in fixed order by /build-prompt via the prompt-pipeline-runner agent. Each stage file follows a consistent structure: Purpose, Input, Process, Output.

| File | Stage | Purpose | When to Load | |------|-------|---------|--------------| | pipeline/constitutional-checklist.md | 1 | Filter governance primitives by agent-type profile; generate constitutional directives | Pipeline execution; standalone constitutional review | | pipeline/fallibilist-overlay.md | 2 | Unconditionally inject five Popper-Deutsch overlay clauses | Pipeline execution (always applied) | | pipeline/epistemic-stance.md | 3 | Select domain stance from six options; compose epistemic contract | Pipeline execution; standalone stance selection | | pipeline/popper-patterns.md | 4 | Select and inject Popper-Deutsch patterns relevant to problem domain | Pipeline execution | | pipeline/confidence-schema.md | 5 | Apply 5-level ordinal confidence scale; map existing rp1 idioms | Pipeline execution; standalone confidence normalization | | pipeline/prompt-validation.md | 6 | Run 3-axis validation (style, constitutional, epistemic); produce validation report | Pipeline execution; standalone prompt validation |

Loading Instructions

Direct Invocation (Backward-Compatible)

When invoked as /prompt-writer for prompt authoring and style guidance:

1. Read this SKILL.md for overview and core principles
2. Read references/tersify.md for the complete compression discipline and style rules
3. Read PATTERNS.md for reusable prompt patterns
4. Read TEMPLATES.md for ready-to-use templates at different complexity levels

This provides the same terse prompt authoring guidance as the original prompt-writer. The reference layers and pipeline stages below are available as additional resources but are not required for this workflow.

Pipeline Execution (via /build-prompt)

When the prompt-pipeline-runner agent executes the six-stage pipeline:

1. Read this SKILL.md to discover the file manifest
2. Execute stages in fixed order -- do NOT skip or reorder:
   • Stage 1: Read references/constitution.md then pipeline/constitutional-checklist.md
   • Stage 2: Read pipeline/fallibilist-overlay.md
   • Stage 3: Read references/epistemology.md then pipeline/epistemic-stance.md
   • Stage 4: Read pipeline/popper-patterns.md
   • Stage 5: Read pipeline/confidence-schema.md
   • Stage 6: Read references/tersify.md then pipeline/prompt-validation.md
3. Each stage's output accumulates and feeds the next stage
4. After all stages: produce two output artifacts (prompt, confidence report)

Standalone Stage Loading

Any individual stage can be loaded independently for targeted review:

• Constitutional review: Read references/constitution.md + pipeline/constitutional-checklist.md
• Epistemic review: Read references/epistemology.md + pipeline/epistemic-stance.md
• Style validation: Read references/tersify.md + pipeline/prompt-validation.md
• Confidence normalization: Read pipeline/confidence-schema.md

Pipeline Overview

constitutional-checklist -> fallibilist-overlay -> epistemic-stance -> popper-patterns -> confidence-schema -> prompt-validation

The pipeline produces two mandatory artifacts per run:

1. Ready-to-run prompt -- SKILL.md (TYPE=skill) or standalone markdown (TYPE=prompt) with constitutional directives and epistemic stance, governed by the 15% budget cap
2. Confidence report -- Structured scoring from each pipeline stage with budget compliance metrics