Core / engine room skills

Design and run the certification's adaptive proctored exam as a defensible measurement instrument (IRT-calibrated item bank, max-information adaptive selection, theta (ability) scoring with standard-error stopping, exposure control + content balancing, and aberrance analytics that flag for a human) making a static quiz employer-trustable.

Produce certification course content: module outlines, lesson scripts, labs, supplements, and assessment items.

Manage the credential lifecycle: issuance, public verification, revalidation cadence, expiry reminders, and revocation.

Assist human reviewers grading certification submissions against calibrated rubrics, pre-score, surface evidence, flag integrity issues, draft feedback.

Help a member or the team run one of the 5 core Scripts (Discovery, Architecture, Delivery, Optimization, Scale) and produce its deliverable.

Run WCAG 2.2 AA accessibility audits + the compliance program on any web property (ours or a client's) POUR-structured, layering screen-reader and keyboard-only passes over automated scanners (floor, not ceiling), severity tiering, remediation by user impact, VPAT/ACR prep, legal orientation (ADA, Section 508, EAA) without legal advice.

Install a deny-by-default RUNTIME firewall in front of an autonomous agent's tool calls plus a tamper-evident hash-chained ledger, then emit a signed attestation, agent X ran under constraints Y, touched Z, spent $W, verifiable against the chain hash. Adds a kill-switch to quarantine a hijacked agent mid-run.

Run a swarm of agents off a kanban board instead of a flat task list, triage each card into a runnable spec, let workers pull cards through Ready → Running → Review → Done under WIP limits, and (the part that actually matters) recover the cards that land in Blocked with a named escalation playbook so the swarm never silently jams.

Stop one-off prompting: design autonomous loops that do the work and supervise at a higher level. Picks the lightest mode on the autonomy spectrum (one-shot → bounded goal+verifier → interval loop → recurring automation → continual-learning), sets a stop condition + budget cap, gives the loop a way to verify its own work (the #1 reliability lever), and keeps a human on every irreversible action.

Design and run a standing autonomous AI organization, a business unit staffed by agents in named roles, not one prompt at a time.

Author the allow / ask / deny permission policy that defines an autonomous agent's autonomy envelope, the static guardrail every loop, org, and coding harness runs inside (e.g. a settings.json permissions block).

Run responsive + accessibility (WCAG 2.1 AA) gates on the Next.js app, breakpoints, keyboard nav, focus management on Radix/shadcn, reduced-motion, contrast, screen-reader semantics.

Audit and remediate AI-generated application sprawl across code, navigation, design systems, accessibility, and tooling. Detect the stack, add missing lint/format/test gates, consolidate duplicate UI, and ask product questions before changing menus or behavior.

Analyze a website, app, repository, screenshots, or product material into an evidence-backed dossier of audiences, features, workflows, benefits, differentiators, integrations, roles, gaps, and explanations. Not for UI remediation or developer docs.

Score a named workflow on an L0-L5 autonomy ladder (manual → fail-closed autonomous) and write the one-level climb. Not vehicle Autopilot advice.

Turn a natural-language question about the Project Brain into a portable, diffable saved query/view (Obsidian Bases-style YAML) over the Brain's frontmatter-tagged markdown notes. Then render the current result and store the.base so it is rerunnable forever. Pairs with as-core-project-brain (Mode C) and the Retrieval Protocol; loadable by any skill that needs a structured slice of the Brain.

Autonomously search connected platforms (Notion/Drive/Box/Figma/Gong/Slack) for existing brand materials and produce a discovery report, what exists, where, how authoritative, and the gaps. Discovery + audit: feeds guideline generation and the voice guard.

Generate brand voice guidelines / a style guide from source documents, call transcripts, or a discovery report. Produces the artifact the voice guard enforces, claims LAW enforced throughout.

Enforce the sender's brand voice, terminology, and claims discipline on any content. Used internally by every content skill.

The intake layer for any engine-backed job, because the brief is the interface. Validates a full brief without re-asking, asks a partial brief ONLY for its gaps, or runs the sectioned questionnaire from nothing; soft fields skip into ledgered assumptions, hard fields (identity, rights, the reference) block with a clear ask.

Query and update the 981-task Agency Script canon.

Run the self-improvement flywheel: detect a recurring manual pattern across session journals, corrections, and failures, auto-draft a candidate skill via the skill-creator, route it through the safety and curation install gates, and propose it for human approval (never auto-install).

Browse the Agency Script OS skill pack by section and pick a skill to run, the menu/directory.

Make the company get smarter over time, turn finished work, user corrections, and post-task retros into confidence-scored learnings written back to the Project Brain, and propose new/updated skills when a pattern recurs above threshold. Runs automatically when a task completes or a correction is given; promotion to a new skill is human-gated.

Strip the AI tells from any prose so it reads like a person wrote it, throat-clearing openers, emphasis crutches, jargon, hedges, and the structural clichés that survive a word swap (binary "not just X but Y" contrasts, negative listing, rule-of-three tics, fragmentation, rhetorical setups, false agency, passive voice).

Search across ALL connected sources in one query (chat (Slack), email (Gmail), docs (Notion/Drive/Box), knowledge bases (Guru), and project/issue trackers (Linear/Asana/ClickUp)) then rank and dedupe results with source attribution and a confidence level. Also manages which sources are connected.

Change existing, working work by shipping a new version ALONGSIDE it instead of overwriting it (variant registries, preview params, feature flags, snapshots) so both can be seen doing their real job before a human picks.

Capture a verified mistake as a typed anti-pattern and auto-inject its guard before a similar future task repeats the error, the loss-side counterpart to compound-learning, built to hunt the fake-success failure mode. Post-mortem the loss, type it, write a trigger-to-guard ledger entry, match its shape on the next task, surface the guard at the irreversible seam.

Stop work safely and write a resume-ready handoff when tokens, context, time, or account usage are running out. Freeze new work; capture verified progress, dirty files, tests, failures, decisions, blockers, active agents/processes, guardrails, and one exact next action.

Meta-skill that designs a domain-specific agent team and generates the agent definitions and skills they use.

Tune a running AI agent harness (Claude Code, Codex, Cursor, OpenCode) so a run costs fewer tokens, finishes faster, and is right the first time, model routing (cheap model for cheap steps, escalate on failure), token/cost optimization (slim prompt, hold pointers not bodies, deliberate compaction), context-window budgeting, parallelization (git worktrees + the cascade method).

Design AND install agent lifecycle hooks so the right action fires automatically on the right event, the "hook-able" pillar that turns a pile of skills into an operating system. Every hook is event → guard → action.

Core category hub: the pack's own engine room.

Design a durable human-approval gate, at an irreversible step (send/post/buy/transfer/sign/delete/deploy) the workflow snapshots state, notifies the owner, BLOCKS so the pending decision survives a restart, then resumes on approve/reject/edit, logging each verdict to the governance Ledger. Unattended runs STOP-and-queue; money and PII never auto-approve.

Hand it a rough prompt and what a good result looks like, and it sharpens the prompt by running it, scoring the output against your standard, rewriting, and repeating until the gains stop. Keeps the BEST version rather than the last, stops on convergence rather than a fixed count, and reports which rule fired and why.

Watches your other loops and catches what they miss, because loops drift, the rule stops being checked, inputs shift underneath, and quality degrades so gradually nobody notices until a client does.

Apply the KISS rule to anything that got too complicated, walk the Delete Ladder (delete, merge, default, defer, explain, where explain is the failure move), count the complexity budget (decisions, steps, concepts, surfaces, time-to-value), and re-architect into default/adjust/full-power layers so the surface shrinks while every capability survives underneath.

Design and build a knowledge base or documentation set for ANY situation (help center, developer docs, internal wiki, agent retrieval corpus, onboarding path, SOP library) from one source of truth that serves human readers and AI retrieval at once. Decides which builder you actually need, then owns granularity, provenance, freshness, taxonomy, and access.

Per-task model-tier routing (cheap/default/pro by role) across all client workloads, plus an LLM-COGS ledger tagged by client/project/skill with budgets, alerts, and a margin report flagging engagements where token cost eats the retainer, making AI spend a governed P&L line, not a surprise invoice. Keywords: tier routing, LLM COGS, token attribution, cost ledger, burn-rate alert, margin report.

Pick a proven, ready-to-run agent loop off the shelf instead of designing one from scratch, browse a curated library of loop recipes (docs sweep, production error sweep, flaky-test stabilizer, full product evaluation, CVE burndown, SEO/GEO visibility, pre-publish source-check, accessibility repair, and more), fill the [brackets], confirm its stop condition + verifier.

Run the periodic "sleep" pass that keeps an agent's persistent memory lean and recall sharp, read the memory store, merge duplicate notes, reconcile contradictions, prune the obsolete, compact verbose entries, and rewrite the top-level index as an index not a dump. The agent analog of sleep, sessions accumulate, the dream consolidates into durable memory.

For the rare high-stakes, one-shot decision, run a panel of diverse models instead of one, several models answer the same prompt independently, a judge model critiques and reconciles them, and the result is fused into a single answer stronger than any one model's.

Decide where large content sets (blogs, programmatic pages, docs) LIVE and keep builds fast as they scale to 10k+, database vs markdown-in-bundle vs object storage, plus the failure modes (EMFILE "too many open files", oversized ISR payloads, frozen-lockfile deploy breaks, repo/bundle bloat) and the safe files→DB migration. Sibling of as-seo-technical (which owns the rendering side).

Operate an Obsidian vault at the file/tooling layer, author valid Obsidian-Flavored-Markdown notes (wikilinks, embeds, callouts, properties/frontmatter, tags), structure the vault (MOCs, daily/atomic notes), and run bulk ops via the Obsidian CLI.

Explain and navigate how the pack works, hubs vs leaf skills, the three execution modes (on-demand / scheduled / triggered), the Project Brain context layer, the 10 guardrails, and which skill or hub to use for a given goal. Renders the operating-model diagram on request.

Turn messy saved operator language into a filled, slot-based Agency Script OS prompt that already routes to the right skill chain, continue, handoff, launch audit, homepage, logo, build loop, wow-native, and the rest of the operator shelf.

Audit, score, and improve the Agency Script OS skills pack across six independent release gates, privacy, validation, structural conformance, one-line self-sufficiency, behavioral evidence, and distribution parity. Fix trigger collisions, weak orchestration, unsafe release content, and package drift without averaging blockers away.

Consolidate overflowing repository planning folders and prepare a durable cold-start packet before ChatGPT, Claude, Cursor, Codex, or another LLM compacts its chat context.

Turn a no-API web portal behind only a login (a vendor, bank, or ad console) into a normalized, scheduled data export, for AUTHORIZED own or client properties only. The governed portal-export primitive: authenticated headless fetch, named credentials never logged, MFA and transactional steps human-gated, schema-bound with change detection.

Build and maintain the perfect AI Brain for a project, a structured, agent-readable context layer (spine + curated knowledge + decisions + references) with the best guardrails on the market baked in, so every skill runs against ultimate, governed project context. Loaded by other skills that need project context.

Turn each agent session into durable, governed memory in the Project Brain, compress the session's exhaust (tool-uses, edits, decisions, open threads) into a dated, human-readable summary, redact secrets/PII at capture, file it to the Brain, and re-inject the spine + last-N sessions at the next start so the next session begins where the last one ended.

Author, upgrade, and ship an Agency Script OS skill to the house bar, description that routes, lean SKILL.md over deep canon, the ten self-sufficiency dimensions, and the mandatory baseline-vs-with-skill gate that kills skills which only tie an expert baseline.

Point at any OpenAPI spec, API URL, or Postman collection and emit a vetted connector row plus a thin skill scaffold (named-credential auth seam, rate-limit/backoff, idempotency, pagination, typed errors) conforming to pack conventions, collapsing the long-tail-connector problem into one step.

Curate an incoming community skill, MCP, or plugin into our pack, one pipeline routes safety to as-sec-skill-supply-chain, checks license + provenance, scores stack/audience FIT, dedupes capability-overlap against what we own, returns an adopt/adapt/decline verdict + placement.

Design sub-agent delegation so a long agent session stays sharp, when the main session should spawn focused, stateless workers (fresh context, own model, restricted tools) that do one scoped job and return only the distillate, keeping the main thread clean.

Enable Agency Script OS Super mode (skill-first process rails, starter hooks, licensed MCP) and retire generic process plugins. Not ultra/swarm effort (as-core-ultra-mode), not hook design (as-core-hooks), not the session router alone (as-core-use-agency-script-os).

Restore crashed or compacted chats into planning/handoffs/CURRENT.md. Lists last N host threads (Grok; Claude/Codex/Cursor via resume-session), distills shipped/open/last-turn, and write-through a Bootspring digest.

The house lens for building a product that wins, drawn from how SpaceX, Tesla and Apple actually decide. Applies the requirement-deletion order, first-principles cost, the demo-or-it-does-not-exist rule, and the integration moat.

Agency Script OS startup/router and Superpowers replacement. Inventories local as-* skills, routes to the best Agency Script OS skill chain, loads Project Brain, canon, trust gates, and applies proof-first, test-first, root-cause-first, no-placeholder execution.

Chain skills into a named, repeatable, end-to-end workflow, the "workflow-able" pillar, so one input produces a finished outcome instead of re-typing the same prompts.