prompt-builder

A Claude Code plugin that wraps the prompt-builder skill — a Prompt Policy Engine that classifies, diagnoses, rewrites, and scores prompts calibrated to model tier and deployment context.

Designed to be callable by agents, tools, and humans through a shared input/output contract.

What it does

Given a raw prompt, prompt-builder returns:

• A 6-Part-Stack optimized prompt (Role / Task / Constraints / Context / Output Format / Acceptance Criteria)
• A 5-dimension quality score (Accuracy / Clarity / Constraint Strength / Output Determinism / Completeness)
• Diagnosis of the top 3 issues in the original prompt
• Tagged assumptions ([ASSUMED: ...]) the caller should review
• A suggested runtime TEMPERATURE_HINT
• Regression detection when iterating from a prior version

Calibrates the rewrite to:

• Model tier: T1 (frontier) · T2 (mid) · T3 (small/fast)
• Deployment: interactive · backend · rag_pipeline · agent · plugin · eval_judge · personal_mobile

Install

Standalone repo with symlink install (matches bookmark/build-loop pattern):

cd ~/.claude/plugins
ln -s ~/dev/git-folder/prompt-builder prompt-builder

Verify:

claude --print "List my installed plugins"
# should include prompt-builder

Usage

Slash commands (humans)

| Command | Purpose | |---------|---------| | /prompt-builder:optimize [prompt or path] | Optimize a prompt end-to-end | | /prompt-builder:score [prompt or path] | Score-only; returns dimensional breakdown + diagnosis | | /prompt-builder:compare <a> <b> | A/B two prompt versions with regression detection | | /prompt-builder:save <id> | Persist the last optimized prompt to the library | | /prompt-builder:list [filter] | Show saved prompts and their latest scores |

Skill invocation (agents and tools)

Any caller can invoke the skill directly with labeled inputs. The skill returns a structured response per skills/prompt-builder/references/caller-contract.md.

Example invoking prompt (sent by an agent):

Use the `prompt-builder` skill.

raw_prompt: You are a helpful assistant. Analyze the CSV and tell me what's interesting.
model_tier: T2
deployment: plugin
output_consumer: human
risk_level: medium

Example response (abridged):

CONFIG: v1.0 | T2 | plugin | Instructional+Analytical | SCORE: 21/25 [A:4|C:5|Cs:4|D:4|Cp:4]

DIAGNOSIS:
- Ambiguity: "what's interesting" is not a deliverable
- Missing constraints: no null/type handling
- Missing context: no column descriptions, no audience

OPTIMIZED_PROMPT:
...

TEMPERATURE_HINT: 0.2

Full contract: skills/prompt-builder/references/caller-contract.md

Library (project-local storage)

Saved prompts live under .prompt-builder/prompts/<id>/ in the consuming project:

.prompt-builder/
├── prompts/
│   ├── summarizer/
│   │   ├── v1.md           # The optimized prompt (copy-ready)
│   │   ├── v1.meta.json    # tier, deployment, score, assumptions, timestamp
│   │   ├── v2.md
│   │   └── v2.meta.json
│   └── rag-reranker/
│       └── ...
└── memory/
    └── lessons.md          # Append-only cross-prompt insights (optional)

Versions are never overwritten — /prompt-builder:save always increments.

Regression testing

The plugin ships with a zero-dependency eval runner:

cd ~/dev/git-folder/prompt-builder
node evals/run-evals.mjs              # run all cases
node evals/run-evals.mjs reranker-t3  # run one case
node evals/run-evals.mjs --dry        # preview invoking prompts without calling claude

The runner spawns claude -p per case, parses the CONFIG line, and asserts on must_contain, must_not_contain, required_sections, and min_score. Exit code is non-zero on any failure.

When to use this plugin

• You're hand-authoring a system prompt for a product and want it scored before ship
• An agent framework (build-loop, research agents, custom orchestration) needs a sub-prompt for a downstream LLM call and wants to delegate the prompt work
• You're comparing two prompt versions and want a regression guard
• You're building a RAG pipeline, agent, or plugin and need deployment-aware optimization

When NOT to use this plugin

• The task is purely creative and you want stylistic looseness
• You need one-shot creative writing with no reusability — just write the prompt
• You want to execute the prompt (this plugin optimizes prompts; executing them is the caller's job)

Architecture

• 1 skill (prompt-builder) — the engine. Progressive disclosure via 4 reference files + 5 worked examples.
• 5 slash commands — thin wrappers. Humans use these; agents invoke the skill directly.
• No agents, no hooks, no MCP server — the skill is a stateless transform. Callers get better results invoking it directly than orchestrating sub-agents.
• Zero runtime dependencies — plain Markdown + JSON + Node's built-in stdlib.

License

MIT

Codex

This package now ships an additive Codex plugin surface alongside the existing Claude Code package. The Claude package remains authoritative for Claude behavior; the Codex package adds a parallel .codex-plugin/plugin.json install surface without changing the Claude runtime.

Package root for Codex installs:

• the repository root (.)

Primary Codex surface:

• skills from ./skills when present
• MCP config from (none) when present

Install the package from this package root using your current Codex plugin install flow. The Codex package is additive only: Claude-specific hooks, slash commands, and agent wiring remain unchanged for Claude Code.