git:20260828.e94eeab to git:20260828.30e6794

# ai-helpers Plugin

Skills and output styles for engineering AI artifacts: prompts, skills, agents, and output styles.

## Skills

- **`prompt-engineering`** — writing instruction text models read, organized by the surface it lands on: one-shot

requests, persistent context, tool and schema descriptions, delegated prompts. Covers the instruction budget, wording

and timelessness rules, formatting choice, and the two behaviors that make reasonable instructions fail — literal

following and instructions compounding with defaults

- **`skill-engineering`** — author and maintain Agent Skills: the routing metadata that decides activation, proscriptive

body content over worked examples, splitting by load condition, and evaluation against a no-skill baseline. Carries

the archetypes, the portable-versus-Claude-Code frontmatter split, host behavior including the listing budget and

compaction limits, and the supply-chain rules for publishing or installing a skill

- **`subagent-engineering`** — design and maintain subagents: when isolation pays over the main conversation, the

description that decides delegation, the frontmatter fields that govern a run, and the system prompt the agent wakes

up with. Carries the false-positive calibration for reviewing agents, the mutating-agent isolation rule, the

orchestration patterns, and the evaluation and troubleshooting depth

- **`output-style-engineering`** — design output styles injected into the system prompt: choosing the style mechanism

over CLAUDE.md or a hook, setting `keep-coding-instructions`, verifying injection with a canary, role framed as

outcome instead of credentials, intent-carrying rules, and voice in rules instead of examples. Carries the file spec

and built-in catalog, the creation patterns, the scoring dimensions, and the iteration diagnostics

- **`prompt-terser`** — adherence-driven terseness audit for iteratively-edited prompts: same-thought-fewer-words cuts

with U-curve discipline, narrative-vs-structural distinction, three-check falsification gate (verbosity-type →

terseness → behavior preservation), and a bracketing preservation inventory that verifies load-bearing literals

survive the cuts. Surviving narrative is rewritten rather than kept; borderline cuts go by default onto a restore list

the caller draws from. User-invoked only (`disable-model-invocation`) — its triggers sit inside

`prompt-engineering`'s, so leaving it in the routing pool only produced a contest it could not win

## Output Styles

## Skill Dependencies

```

prompt-engineering (foundation)

↑

├── skill-engineering (skills are prompts)

├── subagent-engineering (agent prompts are system prompts)

├── output-style-engineering (styles are system prompts)

└── prompt-terser (audits existing prompts for drift)

```

When creating skills, subagents, or output styles, invoke `prompt-engineering` for instruction design.

## Plugin Scope

This plugin covers AI artifact engineering within Claude Code:

- Creating and improving prompts for any AI context

- Building Claude Code skills, subagents, and output styles

It does not cover:

- General software engineering workflows (see `the-coder`)

- Language-specific conventions (see language plugins)

- Git workflows (see `git-commit`)

## Conventions

- Skills reference `prompt-engineering` via `<prerequisite>` blocks because all AI artifacts are fundamentally prompts

- Skills with external documentation dependencies maintain `.dev/reference-inventory.json` for doc fetching via CLI

tools

- For skill structure, routing metadata, packaging, and evaluation — invoke `skill-engineering`