A growing collection of high-quality, installable Claude Code skills.
Flagship plugin: prompt-crafting — it interviews you to nail the spec, then crafts a production-grade prompt for Claude, grounded in Anthropic's official, sourced-and-dated guidance. Newest: mac-cleanup — safely reclaim SSD space on an Apple Silicon Mac (read-only scan → per-batch approval, zero data-loss surprises).
Ask it to delete your files and it won't — it crafts you a prompt for that goal instead, then aligns on the details.
Most prompts fail before a word is written — not because people can't phrase things, but because they haven't pinned down what they actually want. Tools that generate a prompt from a one-line idea skip that step. skill-forge is built around it: it interviews you first.
A short, sharp alignment dialogue nails the spec — goal, output, success bar, failure modes, target model — and only then does it craft the prompt, in Claude's idiom, grounded in Anthropic's official guidance. You bring a messy idea; you leave with a prompt that works.
What makes it different — and hard to copy:
- It nails the spec first. The align-first interview targets the real failure point — a confident prompt for the wrong goal — instead of dressing up an under-specified idea. That's the wedge a one-shot generator structurally can't cover.
- Sourced, dated, and kept current. Every technique traces to an official Anthropic doc with a
last-verifieddate, and an automated job re-checks those docs and opens a PR when they change (that's how it picked up Sonnet 5). You always see where each claim came from and how fresh it is. - Runs where your code lives. It's a Claude Code skill — in-repo and read-only — and it can drop the finished prompt straight into your project as a file you version like any other. It never runs your task, edits your code, or touches your shell.
- Tuned per model. Loads the target model's own official guidance (Opus 5 · Sonnet 5 · Haiku 4.5 · Fable 5 · Mythos 5) via
--modelor auto-detect — effort defaults, instruction style, model-specific gotchas. - One-off or reusable. A clean, ready-to-use prompt by default; add
--templatefor a parameterized system/user template with{{variables}}for your app or pipeline.
- Quick start
- What you get
- See it work
- How it works
- Flags & modes
- Provenance & freshness
- Repo layout
- Contributing
- License
Inside Claude Code:
/plugin marketplace add Mediacom99/skill-forge
/plugin install prompt-crafting@skill-forge
That's it — now invoke it:
/claude-prompt-crafting I want something that reads support emails and tells me how angry the customer is
/claude-prompt-crafting --template a system prompt for an agent that triages our failing tests
Also available:
/plugin install mac-cleanup@skill-forge— safely reclaim SSD space on an Apple Silicon Mac (scans read-only, then removes only what you approve)./plugin install session-handoff@skill-forge—/handoffsaves a long session's state and hands you a resume prompt, so you can/clearand carry on. Maintainers can install the upkeep tooling:/plugin install maintenance@skill-forge
Alternative: install without the plugin marketplace (git clone)
You can drop the skills straight into your skills folder (no auto-update, no /plugin UI):
git clone https://github.com/Mediacom99/skill-forge.git
cp -r skill-forge/plugins/prompt-crafting/skills/* ~/.claude/skills/The plugin-marketplace path above is recommended — it gives you discovery and automatic updates.
| Skill | Invoke | What it does |
|---|---|---|
| claude-prompt-crafting | /claude-prompt-crafting |
Crafts or improves a production-grade prompt for Claude in Claude's idiom (XML structure, multishot, effort/budget), tuned to the target model (Opus 5 / Sonnet 5 / Haiku 4.5 / Fable 5 · Mythos 5) via --model or auto-detect; --template adds a reusable system+user split with variables. Grounded in Anthropic's official docs. |
| reclaim-disk-space (mac-cleanup) | /reclaim-disk-space |
Safely frees SSD space on an Apple Silicon Mac: scans read-only for reclaimable space (caches, build artifacts, dev/package caches, Xcode, Docker), presents a ranked report, then removes only what you approve — batch by batch, Trash-first. Built for zero data-loss surprises. |
| handoff (session-handoff) | /handoff |
Ends a long session cleanly before /clear: writes everything worth keeping into the project's own state files (progress, TODO, decisions, session log — following its CLAUDE.md / AGENTS.md rules), then gives you a short, copy-ready prompt for the next session, already on your clipboard. No questions, no re-exploring — built to be fast when the context is full. |
| refresh-references (maintenance) | /refresh-references |
Maintainer tool: re-fetches the official source docs behind a skill's references, diffs them, and proposes updates. |
It also refines existing prompts — paste one and ask to improve it. By default it returns an improved, ready-to-use prompt; add --template for a reusable, parameterized version.
A real run — paste a rough prompt, and it diagnoses the gaps, crafts a better one, and hands it over.
1. Align — it maps your prompt onto the nine dimensions and asks only what's missing (including whether you want a one-off prompt or a reusable template):
2. Craft & deliver — it rebuilds the prompt, explains every change, and always asks how you want it — it never saves silently:
The skill runs an align first, then craft engine:
flowchart LR
A["Messy idea"] --> B["Intake:<br/>map to 9 dimensions"]
B --> C["Adaptive dialogue:<br/>ask only the gaps"]
C --> D{"Alignment<br/>checkpoint"}
D -- confirm --> E["Load references +<br/>target-model guide"]
E --> F["Craft in<br/>Claude's idiom"]
F --> G["Self-critique vs<br/>success + failure modes"]
G --> H["Deliver:<br/>inline / save / clipboard"]
D -- edit --> C
A prompt spec is "ready to craft" once these nine dimensions are pinned — the skill asks only about the ones your idea leaves open:
goal · output · audience · success criteria · failure modes · context the model needs · constraints · target model + how it's used · examples available
Heavy technique libraries — plus the target model's own tuning file (models/opus.md, sonnet.md, haiku.md, or fable.md) — load only at the craft step (progressive disclosure), so the dialogue stays cheap.
Safe by construction — with hard boundaries. The skill is read-only: it reads, asks, copies to your
clipboard, and writes exactly one file — the finished prompt, and only when you choose save. No edits to
your code and no arbitrary shell, even in auto mode (enforced via scoped allowed-tools: read tools +
Write + a fixed set of clipboard commands). Each run is bounded: it opens by confirming the single
prompt it's crafting and closes by asking where you want it — it crafts that one prompt and nothing else.
| Flag | Effect |
|---|---|
| (none) | Standard: one to two focused rounds of questions. Returns an improved, ready-to-use prompt by default. |
--quick |
One short round max; fills gaps with sensible, stated assumptions. |
--deep |
Exhaustive alignment, loads the advanced reference appendix, and offers a dry test-run (paper simulation) before delivery. |
--refine |
Treat the input as an existing prompt to diagnose and upgrade (also auto-detected when you paste one). |
--template |
Output a reusable, parameterized template (system/user split + {{variables}}) instead of a one-off prompt — also auto-detected when reuse is clearly intended. |
--model <opus|sonnet|haiku|fable> |
Set the target model your prompt will run on (its destination — e.g. a Fable 5 agent), not your current session — so you can craft a Fable/Haiku/Sonnet prompt while running on Opus. Loads that model's tuning at craft. If omitted, it defaults the target to the model you're running as and shows it as a flippable assumption at the checkpoint. |
Every technique in the references is sourced and dated. The skill's references/_sources.md lists the exact official URLs it was distilled from, a last-verified date, and a "volatile items" list (model IDs, reasoning settings — the things that change). Four layers keep it current:
- Sourced + dated references, with stable principles separated from clearly-flagged volatile facts.
/refresh-references— one command re-fetches the sources, diffs them, and proposes updates.check-sources.yml— a weekly GitHub Action that detects when a source doc changes and opens an issue telling the maintainer to refresh. No LLM, no secrets — just fetch + hash.- Automated reconcile (optional) — a scheduled claude.ai routine re-checks the sources and, when they've drifted, reconciles the references and opens a pull request for review (subscription-billed; the PR is the review gate). The update that added Sonnet 5 landed exactly this way.
This is a prompt-engineering tool, so trustworthiness matters: you can always see where every claim came from and how fresh it is.
skill-forge/ # this repo IS the marketplace
├── .claude-plugin/marketplace.json # lists the plugins
├── plugins/
│ ├── prompt-crafting/ # the prompt-crafting skill
│ │ └── skills/claude-prompt-crafting/
│ │ ├── SKILL.md # the align-then-craft engine
│ │ └── references/ # techniques + per-model tuning (models/) + _sources
│ ├── mac-cleanup/ # the reclaim-disk-space skill (Apple Silicon)
│ │ └── skills/reclaim-disk-space/
│ │ ├── SKILL.md # two-phase safety spine (scan → per-batch approval)
│ │ └── references/ # scan-catalog + report-format + reclaim-commands
│ ├── session-handoff/ # the handoff skill
│ │ └── skills/handoff/SKILL.md # persist state → resume prompt → /clear
│ └── maintenance/ # refresh-references (upkeep)
├── .github/
│ ├── workflows/{validate,check-sources,notify-pr}.yml
│ └── scripts/{validate,check_sources}.py
├── docs/ # template spec + auto-reconcile routine setup
├── README.md · MAINTAINING.md · CHANGELOG.md · LICENSE
New skills are welcome — the repo is built to grow. See MAINTAINING.md for the
"add a skill" and "add a plugin" recipes and how the freshness system works. The validate workflow keeps main installable; please make sure it passes.
MIT © Mediacom99. Built on the official prompt-engineering guidance of
Anthropic — see the skill's
references/_sources.md for exact citations.