One author. Every line on disk.

You describe the outcome. One folder opens.

A prompt is a wish you retype. A skill is a method that runs. Every capability is an inspectable SKILL.md folder the agent loads on demand, and Claude Code and Codex read the same files. Open the folder and read the file: plain Markdown, every line of it on disk, one author, MIT for the original work with upstream notices kept.

Start with a worked run Browse all skills Install paths

A run, start to finish.

You do not call skills by name. You describe the outcome, the agent matches your request against the one-line descriptions, and it loads the body of the skill it picks. Here is that loop with suede-visibility-grader, using the real lanes defined in its SKILL.md.

claude code · what you type 
> grade my landing page before I launch it
> https://example.com/pricing
1 · RouteNo skill name required. The description “grade a public page for launch appeal” matches, so the agent loads that one folder. Every other skill stays on disk.
2 · InspectThe skill forbids grading from memory. It pulls the live URL, status code, canonical, robots, sitemap, rendered first viewport, headings, CTAs, Open Graph, and JSON-LD before scoring anything.
3 · ScoreFive lanes, each graded A–F from inspection evidence and mechanical caps, never impression or generosity, then one overall grade.
4 · Hand backA grade per lane, the evidence behind each, and the ranked fixes. The verdict is advice: a failed gate changes what the agent reports, never what it does.

What comes back

The five lanes the grader scores, and the question each one answers.

A–FFindabilityStatus, canonical, robots, sitemap, title, description, durable keywords, duplicate URL risk.
A–FFirst-screen clarityDoes the rendered first viewport answer who this is for, what changes for them, and what to do now, without scrolling?
A–FCTA pullPrimary action, secondary proof action, button text, link targets, and whether there is a reason to click now.
A–FProof and trustScreenshots, commands, docs, live routes, source files, receipts, authorship, and evidence boundaries.
A–FAI readabilityCan a model summarize and cite the page without hallucinating? Structured lede, citation-ready headings, sourceable claims, entity schema, internal links.
Read the full skill

Same loop for every other lane: describe the outcome, let routing pick, read the evidence it hands back. Fixes that are conversion-shaped route onward to site alchemy; a page shipping as part of a release routes to launch packaging.

The folders behind 30+ live sites and 8 iOS apps.

Here is the first thing that changes. Your next staged diff comes back with a letter instead of a compliment.

The ship DAGsuede-graph-flo-xr writes competing plans for one repo change, scores them, prunes the losers, then refutes the survivors and builds only the winner. Call budgets stop at 55, 110, or 200. It reads production and never deploys.
Inspectable by designThe file the agent reads is the file you read. Open skills/suede-code/SKILL.md and that is the whole method, in Markdown, with nothing compiled behind it.
One-name modesJohnny Suede Write and Johnny Suede Design load the whole copy or design stack under one name. Both run on a brief that lists the allowed claims and the forbidden ones. A vague claim gets swapped for the artifact behind it before the copy ships.
SuedifyStudies a reference URL and pushes a target URL toward that style without copying protected assets, claims, or brand identity.
CI that holdsThe CI lane writes GitHub Actions that build only what changed, one required check that cannot deadlock, and branch protection that gates the merge.
A public-workflow spineThe same discipline runs on a page and on a diff. A public page passes an evidence-based audit before it goes out. A code change comes back with a letter, and the weakest of seven lanes caps it.
Proof in publicThe docs ship from the same repo as the skills, one author on both. Every skill folder carries a skill card naming its purpose, owner, license, risks, and the evidence behind the release. Check any claim on this page against the folder it names.
Cue SuedeThree options close a run: change something, preserve this, keep as-is. Name what worked and the agent carries it forward instead of guessing. That is how a folder accumulates judgment: what to do, what to avoid, and when to stop.

Installing every skill does not cost you every skill.

The objection to a big pack is always context. It does not apply here: the agent reads a one-line description per skill and loads the body of exactly the one it picks. Everything else stays on disk.

Progressive disclosure Every skill sits on disk. The agent reads one short description each, matches the request to a single skill, and loads only that skill's full instructions into the context window. 73 folders on disk costs nothing to have One-line descriptions the only thing always in context match 1 skill loaded full instructions, on demand The rest stay unread no tokens, no interference, still one prompt away

This is why breadth is worth having: the marginal cost of the next skill is one line of description. Why breadth is free now ->

Suede AI has zero employees. This is the staff.

Six specialties, from a build DAG that refuses to deploy to an Amazon chat window that argues a refund back. Every description ends with a NOT FOR line naming the sibling that should win instead. That line is what stops a large pack from arguing with itself inside one prompt. The skill catalog lists every folder with its docs.

A studio workstation with a large screen for design and media review.

Suedify

Give the agent a reference URL and a target URL. It reads the layout and the type scale, then rebuilds that grammar with content and assets the target already owns. Your logo, photography and claims stay your own.

Workflow docs Skill folder
A laptop and studio desk used for reviewing copy, visuals, and campaign material.

The page grade, before launch

The page gets an A to F launch grade before it ships. A design-signal lane at D or F holds it, whatever the other lanes score. A published claim the page cannot support caps the grade at D.

Visibility grader Code grade Copy bank
Suede AI Skills resource map for public docs, install paths, Codex, Claude, and workflow evidence.

Campaign in a Box + Music/IP Suite

Point it at a release folder and it comes back with the gaps: a split nobody confirmed, a sample with no source named, a master with no artwork beside it. The evidence lands in a local package on your machine.

Campaign skill folder
Workflow SkillsThe umbrella. Two commands install it, and after that the agent picks its own lane. Narrower plugins exist if you want the marketing set or the code set on their own.
Johnny Suede WriteOne loadable writing mode: page copy, repo copy, CTAs, FAQs, social snippets, persuasion frameworks, and an anti-slop gate before anything ships.
Agent TeamsBefore any lane opens a file, it maps file ownership and flags a collision when two lanes claim the same path. In production, data loss or an auth bypass goes straight to rollback, before anyone investigates.
Code Review + GraderA hardcoded secret, or a permission check a request param can walk around, locks the grade at F with the exact file and line. No other lane lifts it. Seven lanes run from correctness through deploy readiness, and the weakest one caps the letter.
Ship GateWire any repo so CI gates the merge for real: stack and lockfile detection, path-aware builds, one required check, and branch protection that holds.
AI EvalsTurn LLM, RAG, classifier, and agent surfaces into AI-SPECs with failure-mode rubrics, eval cases, and acceptance gates you can fail a release on.
SEO and AI EO AuditNine scored lanes stand between a draft and a public page. A claim the page cannot source comes back as an exact rewrite, with the sourced version already written.
Visibility GraderOne page in, one letter out. The grade comes from what was inspected, so no live inspection caps it at C and a broken primary CTA caps it at D.
Conversion CopyWhatever it writes, it scores. Every draft comes back out of 70 across seven dimensions, directness through search readability, and below 58 it goes back for a rewrite, below 62 when the surface is a launch page.
Design SystemBody text on a dark surface has to clear 7 to 1. Secondary text, 4.5 to 1. Tokens, type scale, and spacing come from the same table, so a review cites the rule it failed instead of a preference.
Site AlchemyIt reads the page as a conversion path and names where the path breaks: the friction sitting in front of the CTA, the proof that never arrives, the mobile view nobody opened.
Launch PackagingNothing is called live until the exact install command has run from a clean temporary directory and the live URL has answered. The completion claim leaves the thread after that, and not before.
MCP QAThe MCP server is dependency-free stdio: 12 tools, 7 resources, 5 prompts. QA drives the full JSON-RPC lifecycle against a running instance, from initialize through prompts/get. A malformed request has to fail the way the protocol says it fails.
Campaign in a BoxRun an artist campaign from one skill: brief, assets, copy, release packaging, and rights package context, plus sync packaging, release linting, and rights audit.
Release LinterAudit a release folder for missing metadata, artwork, masters, stems, credits, splits, and provenance gaps before the work leaves your machine.
Rights Passport + AuditOrganize ownership, contributor, sample, and split evidence into a local transfer package.
Amazon Returns RecoveryAudit an Amazon account for restocking fees, short refunds, and forgotten Amazon-billed subscriptions, then drive live chat until the fee is waived or the refund lands. Uses synthetic examples and requires per-item confirmation.
Subscription RecoveryThe same negotiation discipline pointed at everything billed outside Amazon: find forgotten streaming, gym, SaaS, and app-store subscriptions, cancel the dead ones, and negotiate refunds through each service's own support channel.
Cue SuedeTell the agent what you like or dislike during the run, then let the skill preserve the correction so the next pass is sharper. Final handoffs explain the result twice: simple version first, then the breakdown.

Use it before the page goes public.

In the minute before you publish, the page gets a letter. The first-screen lane asks three questions: who this is for, what changes for them, and what to do now. A first viewport that answers none of them takes the whole grade down with it.

Before redesignPoint Suedify at the reference site and it maps that design grammar onto the target. The reference lends its structure, and the target keeps its own logo and claims.
Before publishing docsRun the install command yourself, from a clean directory, and open the page on a phone. A doc that publishes a command nobody ran is a guess with good typography.
Before launchGrade the page first, then fix only the lane that failed. The ranked fixes arrive with the evidence behind each one, so you are not guessing at which change moved the letter.
Before the next passTell the agent what looked right and what missed, or say Cue Suede to preserve the preference, revise the output, or keep the pass as-is.
Before handoffAgent-team loops scout first, split lanes that do not touch the same files, and review each other adversarially. The record at the end names the files, the commands, and every caveat that was not dropped to make it look clean.

Every claim on this page points at a file.

The license is MIT for the original work, with upstream notices kept where a component was adapted. There are no binaries in the repo, no telemetry in the skills, and no account to create. Every link below opens a file or a page you can read without installing anything.

GitHubInspect the source, license, skills, scripts, docs, tests, and MCP catalog. Skill DocsBrowse every installable skill folder and their public descriptions. Suede Graph Flo XRRead the shipping DAG for broad, authorized, outcome-bound work. Installs + MCPUse public install commands and optional structured MCP discovery. Copy BankReuse public-safe wording for docs, launches, FAQs, and evidence boundaries. Visibility GraderGrade public pages for findability, clarity, proof, and AI readability. Code Review + GradeReview ship risk with evidence-backed A-F grading lanes. AI EvalDesign eval cases, failure rubrics, and acceptance gates for AI features. CampaignPackage artist campaign briefs, assets, copy, content plans, and next moves. Rights PackageCreate a local creator rights package for downstream review. Release LinterGenerate a release-readiness report before handoff. OpenAI SkillsCompare the standard public skill format used by Codex. Claude SkillsReview the Claude Code skill documentation path.

Nothing leaves your machine unless you say so. Local workflows read and organize files. Browser, deployment, publishing, and support workflows use access you already have, and only on a run you approved.

From folder to evidence package.

The five creator skills start from a local folder and end with an organized evidence package for review. Reports stay local and nothing is uploaded.

1 · InspectWalk the folder as it stands and write down what is in it, down to the stems and the loose documents.
2 · LintA bundled offline script scores the folder and marks every fact present, missing, or unknown. Unknown never gets upgraded to confirmed.
3 · ConfirmOnly the creator moves a fact from unknown to confirmed. The skill asks the question and records the answer.
4 · PackageThe package is written to your machine.

Two commands for Claude Code. Two for Codex.

These are standard SKILL.md folders with bundled references, templates, and Python scripts.

Claude Code: first install

First install takes two commands: add the marketplace, then install the full pack. After the marketplace is present, install or update is one command. The first thing worth running: use suede-code to review my staged diff and give it an A-F ship grade.

/plugin marketplace add JasonColapietro/suede-creator-skills
/plugin install suede-skills@suede
Install and MCP docs Public explainer

Codex: full pack

Add the Codex-native marketplace, then install every skill and 1 read-only MCP server with 12 tools. Use the direct skill installer only when you want one selected skill.

codex plugin marketplace add JasonColapietro/suede-creator-skills --ref main
codex plugin add suede-skills@suede-codex

Hermes Agent: tap and install

The repo is a Hermes skills tap. Subscribe to search the pack from inside Hermes, or drop the first line and install one skill straight from GitHub.

hermes skills tap add JasonColapietro/suede-creator-skills
hermes skills install JasonColapietro/suede-creator-skills/skills/suede-seo-audit

Claude Code: manual

Install into a project .claude/skills folder or your user-level ~/.claude/skills folder.

git clone https://github.com/JasonColapietro/suede-creator-skills.git /tmp/suede-creator-skills
cd /path/to/your-project
mkdir -p .claude/skills
cp -R /tmp/suede-creator-skills/skills/suede-workflow-skills .claude/skills/

Books by Jason Colapietro.

Jason has published five books. Same author, same insistence on receipts: guitar tone, AI ownership, proof systems, and creator rights.

The Signal ChainElectric-guitar tone history, memoir, method, and workbook editions. The Guitar Without a NumberSelf-taught guitar method with theory, tone, songbooks, and music IP rights. The Human Authenticity LayerOwnership, origin, and AI in the creative map. Proof as InfrastructureDurable systems built around proof instead of trust assumptions. Stake Your ClaimCreator ownership, authorship, provenance, agents, and durable rights.

Field notes.

The design decisions behind the pack, written up in full: what a skill costs, how routing avoids collisions, and where memory belongs.

Released Oct 24 · 89754d4 Five offline SEO tools in the public pack 535 commits · 24 weeks · read the full ship log ->
Aug 2026 71 skills installed. Your agent reads almost none of them. Progressive disclosure, 55/110/200-call Suede Thought Graph shipping (winner-only, read-only, never deploys), and the MCP layer Aug 2026 Why breadth is free now, and what that changes The economics of a big pack Aug 2026 NOT FOR: the two words that make a big skill pack work How skills route without colliding Aug 2026 Your memory file is a tax you pay on every prompt Why memory belongs in skills Archive All essays The full writing archive

Suede workflow FAQ.

Four answers for the developer deciding whether to install this before the next diff.

What are Suede AI Skills?

Suede AI Skills are free, open-source SKILL.md folders for Claude Code and OpenAI Codex. Original work is MIT licensed, and adapted components retain their upstream notices. The range runs from an A to F review of a staged diff to an account-owner-approved Amazon refund review.

Does this work with my agent?

Claude Code and Codex each install with two commands. For anything else, npx skills add JasonColapietro/suede-creator-skills drops the same folders into Cursor, Copilot, or Windsurf. The files are identical on every surface, because there is only one copy of them.

When should I use Agent Teams?

When two agents are about to edit the same file and neither of them knows it. Agent Teams maps file ownership before anyone opens anything, and sequences the lanes that collide. The handoff is held until the evidence is attached, so done means something.

How does Suedify use a reference site?

Suedify studies layout, rhythm, typography, color roles, motion, and proof structure, then recreates the design grammar with target-owned content and assets.

Proof is part of the release.

First install takes two commands. After the marketplace is added, install is one. Every skill is plain Markdown.

$ /plugin marketplace add JasonColapietro/suede-creator-skills → /plugin install suede-skills@suede All install paths