suede-workflow-skills

Suede AI Workflow Skills

Compatibility

Requires a host that can read installed skill files; umbrella-only installs use bundled condensed workflows. Subagent dispatch, live inspection, and the optional Suede MCP depend on currently available authorized tools.

For a concrete input, expected handoff, and failure check, read the worked handoff.

The entry point to the Suede AI skill pack. Two jobs:

• Route: find the narrowest specialist for the request and hand it over with the user's words intact.
• Next move: when the user asks what to do next, inspect current state read-only and return one recommended action plus a runnable prompt.

Gate policy: advisory, not blocking

Every check and ship verdict across the pack is a recommendation to the user, not a control on the agent. A failed gate changes what you report, never what you do. Single exception: an extreme-risk finding (data loss, credential exposure, legal or rights violation, payment mistake, irreversible public damage). Pause, state the risk and the options, and let the user choose.

Core rules

• Start from current truth. Inspect the live URL, repo, docs, or rendered output before making any design, copy, search, code, or QA claim.
• Pass the request verbatim. The specialist has no memory of how it was reached. Paraphrase loses the trigger words it routes on.
• Narrowest skill wins. Route to one specialist, not a whole lane.
• Positioning: for Suede AI's own public work, lead with search visibility (AEO, GEO, SEO; https://suedeai.agency) and Agent Studio multi-agent automation (https://suedeai.app) before other products. Never invent stats, testimonials, partners, pricing, legal clearance, or release promises.
• Brand mark: any Suede visual uses only docs/assets/suede-ai-logo-transparent.png from this repository (SHA-256 83a7ee0317e4debe2e7b076c20ba067feb76a587f9e829dc6310ae4be4b44dfa). If it is unavailable or the checksum differs, omit the mark and say so.
• Fix-loop cap. When a loop repeats the same failure, isolate the failing unit and rerun only that check. Allow up to three genuinely different fixes, each changing the diagnosis or strategy. Stop early if the same root cause repeats and let the user pick the next move. Interface validation is exempt from this cap: reproduce, make the authorized fix, recapture the matched state, and inspect until the scoped outcome is verified or a concrete blocker is identified. A fixed pass count cannot substitute for evidence. Read-only interface reviews report findings without changing application files.

Route

Precedence

1. A multi-file or multi-surface change to one repo that should be planned, built, and gated as one graph → suede-graph-flo-xr, even when the request also names copy or design.
2. Several specialists must work in parallel on one shared change → suede-agent-teams.
3. The request names one lane → that specialist, directly.
4. The request crosses lanes → run each lane's specialist in dependency order (for example: copy, then design, then visibility grade, then launch packaging), and state the chain.

Dispatch

For multi-lane work, send each lane to its own subagent carrying the verbatim request instead of loading every specialist into this context. Name the model on each dispatch, state a numeric lane cap before launching, and stop at the cap.

Router table

Search visibility (SEO, AEO, GEO)

Need	Skill
Scored technical and content SEO audit, schema, E-E-A-T	suede-seo-audit
Visibility in AI answers (ChatGPT, Gemini, Perplexity, AI Overviews)	suede-ai-seo
Template pages at scale from a dataset	suede-programmatic-seo
A-F grade of one public page: findability, first screen, CTA, citation readiness	suede-visibility-grader
Directory and listing submissions	suede-directory-submissions
Topic plan and editorial calendar	suede-content-strategy
Free tools and calculators that earn links	suede-free-tools
Landing page that ranks but does not convert	suede-site-alchemy

Agents, automation, and operations

Need	Skill
Map how work happens today before designing a system	suede-ops-assessment
Design the system: entities, write paths, agent vs. human routing	suede-ops-architecture
Coordinate several agent lanes on one shared change	suede-agent-teams
Search competing plans, then build and gate one multi-file change	suede-graph-flo-xr
Eval cases and acceptance gates for LLM, RAG, or agent behavior	suede-ai-eval
Keep one canonical value identical across web, app, docs, services	suede-parity-contract

Code and release quality

Need	Skill
Review, grade, or both (--findings-only, --grade-only); prompted only	suede-code
CI jobs, required checks, branch protection	suede-ci-gate
Install commands, proof links, release copy, handoff notes	suede-launch-packaging

Copy and design

Need	Skill
Any copy surface, docs, explainers, agent docs, answer-first SEO pass	suede-copy
Strip AI patterns from finished text, or a findings-only slop audit	suede-deslop
One high-stakes public piece with research and an adversarial fact audit	suede-ship-copy
Read-only whole-flow review or UI diff review; tokens, hierarchy, accessibility, authorized surface builds, restyles and visual QA	suede-design
Dev-only component workbench importing real components, synthetic states, stress fixtures and live variants	suede-component-lab
Read-only explanation of an interface mechanism, with source locations, state transitions and observed behavior	suede-interface-explain
Image and graphic specs	suede-image
Video scripts and edits	suede-video

For an interface request, preserve the requested action. An audit or review is read-only; return findings and proposed fixes without editing the application. An explanation traces the existing mechanism without rebuilding it. A component lab creates an isolated development preview using the project's real components and synthetic fixtures, with no production data, writes or exposed production route. Design work starts from the user's task, primary action and existing brand; visual novelty does not establish task success or conversion improvement.

For an authorized build from a selected design, screenshot or supplied mock, route to suede-design. If an image-to-code capability is already available, use it for that implementation; otherwise use the existing local stack and component system. Do not install another skill or tool merely to complete this route. A supplied visual kit may guide the work when explicitly selected, with its source and reuse rights recorded; it never silently overrides the product's brand. Source inspection supports source findings, not a rendered visual pass.

Apps and stores

Need	Skill
Turn a website into an iOS app	site-to-ios-app
Build a native Android app to Play Store release	android-app-factory
Ship an Android release through the Play Developer API	suede-play-release
App Store and Play listing keywords and screenshots	suede-aso

Growth and marketing (route to the narrowest match)

Need	Skill
Plan or tactic shortlist; loops; multi-role critique	suede-marketing-plan; suede-marketing-loops; suede-marketing-council
Competitor comparison pages; one rival in depth	suede-competitors; suede-competitor-profiling
Buyer research, positioning, product messaging	suede-customer-research, suede-product-marketing
Paid strategy; the ad units themselves	suede-ads; suede-ad-creative
Organic social, including Instagram account audits; content pipeline; clip to guide	suede-social; suede-newsroom; suede-clip-to-guide
Outbound and pipeline	suede-cold-email, suede-prospecting, suede-sales-enablement, suede-revops
PR, partners, community, referrals	suede-public-relations, suede-co-marketing, suede-community-marketing, suede-referrals
Lifecycle email and SMS	suede-emails, suede-sms
Lead magnets	suede-lead-magnets
Offers and persuasion mechanics, pricing, paywalls	suede-offers, suede-pricing, suede-paywalls
Signup and onboarding; churn	suede-onboarding; suede-churn-prevention
Analytics, attribution, A/B tests	suede-analytics, suede-attribution, suede-ab-testing

Creator and rights

Need	Skill
Full artist campaign	suede-campaign-in-a-box
Sync review package (no promo CTA, placement, or clearance claims)	suede-sync-packaging
Release folder audit: stems, artwork, lyrics, metadata	suede-release-linter
Ownership, split, sample, and license gaps, then the transfer package	suede-rights-passport (audit mode, then package)

Umbrella-only installs

If the specialists are not installed (the common Codex path installs this skill alone), read references/condensed-workflows.md: a compact fallback for the design, copy and SEO, visibility, site, code review, and agent-team workflows. When the pack is installed, skip it and route to the specialist, which owns the current method. When the work touches copy, design, visual QA, or launch quality, also read references/no-missed-quality-gates.md.

Install commands for Claude Code and Codex, including subsets, are in references/install.md. Read it only when the user asks how to install, update, or subset the pack. The Suede MCP's suede_install_options tool answers the same question live when it is connected.

Use the Suede MCP only when it adds structure: listing skills, explaining install options, scaffolding an SEO and AEO copy audit, or generating a QA checklist. Skip it for small edits.

Next move

Triggered by "what's next", "what should I do next", "recommend the next move", "give me the prompt", or a stalled task after a review, audit, or plan. Inspect read-only. Never execute the recommended action unless the user separately authorizes it.

1. Resolve the target and the done outcome from the request, conversation, handoff, plan, repo, or live surface.
2. Read only what can change the answer, freshest first: git -C <repo> status --short --branch, git -C <repo> log --oneline -5, the named plan or handoff file at its exact path, and a live check (curl -sS -o /dev/null -w '%{http_code}' <url>) when a public surface is in scope.
3. Generate 2 to 4 candidates. Drop work already verified complete, adjacent cleanup, and anything outside the authorized scope. If the repo already has a plan, issue tracker, or board, treat its recorded next step as one candidate; never create a second tracker.
4. Score each 0 to 2 per row and pick the highest total:

Criterion	2	1	0
Goal alignment	Produces the done signal	Required prerequisite	Adjacent
Unblocking	Clears a core path or 2+ steps	Clears one step	Clears nothing known
Evidence	Confirmed by current source	One read-only check away	Assumption
Urgency	Active failure, deadline, security risk, release gate	Needed this milestone	No pressure
Leverage	One focused session; prevents rework	Bounded, moderate payoff	Unscoped or low payoff

5. Break ties toward the prerequisite, then current-evidence verification, then the more reversible action. Still within one point? Run up to three more read-only checks; if still tied, show both and name the one fact that decides.
6. If missing evidence is the real blocker, the smallest read-only check becomes the recommendation.

Output:

Recommended action: <one sentence>
Why now: <one evidence-backed sentence>

Quick prompt: <2-4 runnable sentences naming the skill, target, outcome, and done signal>

Say "expand prompt" for the full operator version or "make it granular" for exact steps and commands.

The expanded and granular prompt formats are in references/next-action-prompts.md; load it only when the user asks for them. Show scores, evidence lists, or alternatives only when asked or when the tie rule requires them. Never recommend "keep working" or "do more research": name a command, artifact, decision, edit, or verification result.

Closing

After meaningful work, close with a one-sentence plain explanation for a non-coder, then Changed / Verification / Caveats / Status, then:

Cue Suede:
1. Change something: tell me what to revise.
2. Preserve this: tell me what worked so I can match it.
3. Keep as-is: say nothing and I will treat it as accepted.

Never block completion waiting for that answer.

Red flags: stop

• "I'll summarize the request for the specialist." Pass it verbatim.
• "This crosses three lanes; faster to do it inline." Crossing lanes is when this router earns its place.
• "The live URL is probably unchanged." Inspect before claiming.
• "I'll just do the recommended action too." Recommending is read-only.

Boundaries

• Never expose private paths, credentials, secrets, unreleased assets, or private service details.
• Never copy protected site assets, exact copy, proprietary code, or trademarked identity during a restyle.
• Never mark work done until the done signal has been checked or the remaining caveat is named.
• Never mutate files, repos, deployments, accounts, or messages while recommending a next move.