Lesile-Yin/agent-skills

# Audit Swarm

One reviewer sees one lens. audit-swarm sees many at once by spawning a small
swarm, each constrained to its lane, then merging without duplicating noise.

## When to use

- After a non-trivial change, before merge.
- When prism-review's four lenses need more coverage (e.g. a dedicated coverage
  or style lane).
- When a single model's review keeps missing a category of defect.

## Workflow

# Caveman

Turn padded text into its load-bearing skeleton. Caveman mode is not rudeness —
it is *lossless compression of intent*. The goal: every surviving token changes
what the reader/agent does next; every removed token was noise.

## When to use

- A chat transcript or status update restates the same point three ways.
- A tool output or log dump is about to be pasted back into an agent context.
- A PR description, commit message, or design doc is 80% framing and 20% content.
- You are about to pay token cost to re-read something you already said.

## Workflow

# Grill Me

A plan that survives only its author's scrutiny is untested. grill-me puts the
plan in front of a *different* model whose job is to break it, not bless it.

## When to use

- An architecture or design proposal is internally consistent but unchallenged.
- You have only one model's context on a high-stakes decision.
- A plan keeps "working" in tests but you can't name its failure modes.

## Workflow

### Step 1 — Planner emits the design

# Prism Review

One change, four lenses, one merged verdict. A white beam through a prism
splits into distinct bands; a code change through this skill splits into four
independent examinations that never contaminate each other, then recombines
into a single ranked report.

## Why parallel lenses

A single reviewer carrying four concerns in their head produces two failure
modes: anchoring (the first bug found colors everything after) and fatigue
(lens #4 gets a skim). Independent passes avoid both, and their disagreement
is signal — a finding confirmed by two lenses is almost never noise.

# Prompt Forge

Turn fuzzy intent into a prompt that an agent (or a human) can execute on the
first try. A forged prompt is not longer for the sake of length — it is dense
with constraints that remove rework.

## When to use

- The request contains vague quantifiers ("better", "clean", "a few", "optimize").
- The request hides assumptions (stack, environment, audience, deadline).
- The output will be consumed by another agent/session with no shared context.
- A previous attempt at the same task produced a wrong-shaped result.

## Workflow

# Slide Smith

Batch-forging decks: audit → standardize → verify. Built for courseware
pipelines where dozens of .pptx files must share one visual system.

## Safety rule (non-negotiable)

Never modify originals in place. Every script reads a source path and writes a
new file (`*_standardized.pptx` or an explicit `--out`). If the user asks to
overwrite originals, produce the outputs first, then let *them* swap.

## Capabilities

### 1. Inventory — know what you're dealing with

# Token Saviour

Token discipline is a routing problem, not a willpower problem. token-saviour
picks the lightest tool that can still do the job, then compresses what must be
sent. It is the operational layer beneath token-warden's budget accounting.

## When to use

- Before spinning up an expensive coding agent for a small, well-scoped edit.
- When a conversation history exceeds a token threshold and must be trimmed.
- When deciding which model/agent should own a sub-task.
- When a task can be done by a formatter, a linter, or a grep instead of a chat.

## Workflow

# Token Warden

Tokens are spent in four independent layers. Optimizing one layer does
nothing for the other three — and most "save tokens" advice only touches one.
This skill audits all four and applies the cheapest adequate tool per layer.

## The four-layer model

| Layer | What burns tokens | Symptom of waste |
|---|---|---|
| R — Reading | Files, search results, docs pulled into context | Full-file reads to find one symbol |
| T — Tool output | Command stdout, test logs, build output | Raw 5k-line logs pasted into context |
| C — Conversation | The agent's prose, explanations, summaries | Preamble, restating, apology loops |
| G — Generation | Code written/rewritten | Whole-file rewrites for 3-line fixes |