cometweb-io/cometweb-agent-skills

Run an always-current, evidence-governed AI decision council for material, high-stakes, multi-domain decisions where options, risk gates, forecasts, and living memory matter. Do not use for routine weekly product prioritization, whole-repo roadmapping, release-candidate gates, customer support triage, or single-domain specialist work when a dedicated skill exists. Use when the user explicitly asks to "przepuść przez Radę", "zapytaj Radę", compare consequential options, challenge a major plan, decide GO/NO-GO/TEST/DEFER on strategic choices, verify whether a prior decision is still current, inspect Council health, or improve the Council itself. Orchestrate blind advisers, conditional specialists, binding risk gates, live/fresh evidence, temporal truth, contradiction testing, forecasts, living Decision Memory, watch dependencies, human escalation, and champion/challenger evaluation.

Naturalize, edit, or substantially rewrite English and Polish prose while preserving meaning, factual constraints, and the author's intentional voice. Do not use merely for generic proofreading, unrelated copywriting, or AI-authorship detection. Use specifically when the user asks to humanize text, sound less like generic AI writing, remove AI tells, perform a strong/deep/robust rewrite, clean suspicious invisible Unicode in prose or Markdown, or re-express text with provenance-aware caution. Supports blogs, articles, LinkedIn posts, emails, proposals, documentation, release notes, and READMEs. Never claim human authorship, detector defeat, or watermark removal without an appropriate supported test.

Run a final evidence-backed acceptance gate on a content, research, documentation, sales, or other knowledge artifact against an explicit brief, required evidence, unresolved findings, and format/QA criteria. Do not use as the final production-release gate for software, to choose among consequential strategic options, to perform the initial review, or to invent acceptance criteria after seeing the result; use release-readiness, ai-council, or the relevant reviewer instead. Use when the user asks whether an artifact is actually ready, complete, publishable as a deliverable, or has passed its defined acceptance contract.

Curate and version eval corpora for Agent Skills: case taxonomies, discovery/forced/negative controls, adversarial and regression cases, difficulty strata, holdout isolation, provenance, duplication, contamination, coverage balance, and benchmark hashes. Do not use to run the experiment or claim lift (skill-evaluator), to design the grading rubric (rubric-designer), to edit the skill (skill-creator), or to call leaked cases a clean holdout. Use to build an eval dataset, golden set, benchmark suite, regression corpus, or holdout.

Create or revise a durable ArtifactBrief for a formal handoff, repeatable acceptance contract, multi-specialist artifact or explicitly requested PRD. Preserve evidence policy, protected invariants, unresolved material choices and brief/rubric identity. Do not use for ordinary clarification, a clear request to produce the artifact, general planning, product discovery, candidate scoring or a consequential decision. Use when the user requests a brief before execution or a downstream receiver requires this contract.

Continuous competitive intelligence and competitor change detection. Do not use for a one-time initial deep profile with no monitoring intent (prefer a dedicated competitor-profiling skill when one is installed), for extracting a competitor's mechanisms to adapt into your own product (use product-teardown), for writing comparison pages or sales battlecards, for setting your own prices, or for a full search-visibility audit (use seo-geo-aeo-maxxing). Use when the user asks to monitor competitors over time, refresh existing competitor profiles, detect what changed since a prior scan, track pricing/product/positioning/SEO/ads/reviews/company changes, maintain a competitor watchlist, produce recurring competitor digests, verify competitor claims, analyze cross-competitor trends, or turn observed deltas into product/GTM/sales implications. Use it when temporal state, snapshots, deltas, alerts, freshness, evidence provenance, or recurring intelligence operations matter.

Build an evidence-backed initial competitor profile from public and supplied sources, covering company, ICP, positioning, product surface, pricing, proof, discovery channels, uncertainties, contradictions, and a normalized handoff to Competitive Intelligence. Use for a first deep baseline or dossier. Do not use for recurring competitor monitoring, transferable product mechanism teardown, comparison-page copy, pricing decisions, or unsupported claims.

Review editorial, marketing and sales content against its brief, reader and evidence, using editorial or adversarial checks. Use for actionable review, pre-publication QA, roasting or red-teaming claims and offers. Do not use for scientific peer review, code critique, primary research, full rewrites or final release acceptance; use the relevant specialist. content-roaster is the compatibility entry for the adversarial profile and legacy v6 tools.

Compatibility alias for content-reviewer with the ADVERSARIAL profile and retained cometweb.content-roaster/v6 tools. Use to roast, tear apart, red-team or brutally critique marketing, sales, product, editorial and long-form content with evidence and a repair. Do not use for scientific peer review, code critique, rewrite-only editing, primary fact research or SEO audits. This is one content review, not an extra reviewer or a multiagent mode.

Write evidence-aware informational or editorial prose (articles, guides, reports, documentation, research-backed explainers) from a brief or clear request. Do not use for persuasion-first landing copy, ads, or cold email, copy-editing or humanization-only rewrites (ai-humanize), long-form publication (longform-publisher), primary research (evidence-researcher), or critique (content-reviewer, content-roaster). Use when the user wants the artifact itself and factual integrity, reader utility, structure, and claim discipline matter.

Run the operational customer-to-resolution loop across support conversations, customer cases, incidents, account risk, commitments, internal handoffs, feedback clusters, and GitHub engineering work. Do not use for broad VOC/persona research, retention-program design, general CRM architecture, analytics implementation, product roadmap prioritization (Product Operator), release-candidate GO/NO_GO (Release Readiness), or security exploitation; hand those workflows to the specialist skill. Use when asked to triage an inbox or ticket queue, investigate a customer problem, detect or coordinate an incident, build an operational account 360, watch churn/non-renewal signals, find overdue promises or stalled escalations, dedupe customer-reported bugs into GitHub, produce customer-ops briefs, or verify that a fix actually resolved the customer-visible symptom. Composes with Gmail/support tools, HubSpot/CRM, billing, product analytics, GitHub, Notion/incident records, and connected files.

Find, verify, qualify, compare, and manage companies as design partners, co-development partners, beta partners, paid pilots, lighthouse customers, or early adopters. Do not use for broad outbound lead generation, interview/VOC synthesis, or writing outreach copy; hand those to a prospecting, customer-research, or cold-email skill when one is installed. Use for design-partner discovery, early-adopter shortlists, qualification of an existing candidate list, cohort selection, partner-readiness validation, active-partner review, or revalidation of prior candidates. Optimize for learning value, urgency, representativeness, implementation feasibility, real user/champion access, mutual commitment, evidence quality, transferability, cohort coverage, and cost-to-learn; separate desk-research fit from live readiness, preserve evidence lineage/freshness, and keep prestige or contract size from replacing product-learning quality.

Create, research, validate, edit, redesign, and prepare evidence-backed ebooks, white papers, and practical PDF workbooks, especially the CometWeb ebook series. Do not trigger for a standalone blog post, a generic website audit, simple PDF extraction, or merely a request to install a skill; to refresh, reconcile, or release an existing long-form publication through a canonical manuscript with derived DOCX/PDF lineage, use longform-publisher. Use for requests such as zrob ebook, napisz poradnik PDF, deep research do ebooka, zweryfikuj ebook, popraw okladke, or prepare a publication from research through reviewed PDF. Own publication scope, research coverage, claim-to-source mapping, editorial structure, meaning-preserving Polish/English editing, series design, revision-aware QA, and publication handoff. Reuse evidence-researcher and ai-humanize when available. Research-only and redesign-only requests must stay within their requested scope.

Build auditable Evidence Packs for consequential research, fact-checking, due diligence, verification, and cross-skill evidence handoff. Do not use for casual single-fact lookup or as the final decision-maker when a dedicated decision skill exists. Use when a question must be decomposed into material claims and checked against primary or system-of-record sources: verify currentness/effective dates/versions, distinguish source artifacts from claim-specific evidence, search for falsifiers and negative evidence, resolve contradictions, detect derivative or non-independent sources, identify evidence gaps, compare evidence deltas, or prepare reusable evidence for AI Council, product, technical, audit, SEO/GEO/AEO, sales, or customer workflows.

Aggregate repeated findings, acceptance failures, eval results, user corrections, and workflow incidents across skill runs to identify recurring failure patterns and propose evidence-backed improvements to skill instructions, routing, references, kernels, evals, or host integration. Do not use to silently self-modify skills, learn from a single weak signal, replace product/research analytics, or apply repository changes without explicit authorization. Use when the user wants the skill system to learn from repeated mistakes, run a retrospective, improve reliability over time, or convert observed failures into regression tests.

Run the publication control plane for long-form work (ebook, report, playbook, white paper, guide, handbook): one canonical manuscript, claim traceability, derived DOCX/PDF/HTML lineage, release stages. Do not use to create a new CometWeb ebook from research to design (ebook-publisher), for primary research (evidence-researcher), humanization (ai-humanize), an article draft (content-writer), or PDF/DOCX file edits. Use to refresh a publication into a new edition, reconcile a manuscript with its derived formats, or gate RELEASE_READY.

Evidence-grounded cross-domain portfolio control plane for deciding what a user or small team should focus on across products, client work, research/academic commitments, growth, operations, and side projects. Do not use for deep prioritization inside one product/repo, release GO/NO-GO, or executing a multi-skill workflow when a specialist/control-plane skill owns that task. Use when the user asks what to do this week/next 7-30 days across multiple projects, how to allocate limited capacity, which commitments conflict, what to pause/drop/delegate, how to reconcile client deadlines with product/research work, or for a portfolio-wide review.

Evidence-governed product operating system that reconciles GitHub implementation and release state, Notion roadmap/tasks/decision docs, product context, and available outcome signals to answer what one product team should do next. Do not use for a first-time whole-project roadmap baseline (repo-to-roadmap), a release-candidate GO/NO_GO (release-readiness), allocating capacity across several products and commitments (portfolio-operator), or customer support triage (customer-ops). Use for questions such as what remains, what to build/fix/verify next, how to finish or unstick a product, what is actually done, whether roadmap and repo agree, how to plan the next product cycle, what changed since the last review, or what should wait/stop. Produces bounded BLOCKER/VERIFY NOW/DECISION NOW/NOW/NEXT/LATER/STOP actions, dependency-aware sequencing, state drift, readiness, immutable snapshots/deltas, and specialist handoffs. Read-only by default; delegates deep audits and consequential decisions.

Analyze external products, websites, apps, APIs, documentation, and source repositories to extract evidence-backed, transferable, implementable product and engineering patterns. Do not use for ongoing competitor monitoring/delta analysis, broad competitor dossiers, external-facing comparison pages, or a roadmap of the user's own repo. Use for teardown, product/architecture reverse-engineering, benchmarking, clean-room reconstruction maps, evidence-linked feature parity, multi-product synthesis, and source-to-target adaptation; including Polish requests such as "przeanalizuj produkt/repo", "wyciagnij wzorce", and "co warto wdrozyc".

Run the CometWeb quality loop for an artifact, repository, or skill (brief, create, review, roast, repair, accept, measure, revalidate, roll out) while keeping candidate identity, policy locks, finding lineage, disagreements, and quality debt. Do not use for general multi-skill orchestration (skill-orchestrator), specialist analysis, strategy decisions (ai-council), or software production readiness (release-readiness). Use to run or resume the quality workflow, a batch campaign, finding reconciliation, or a staged skill rollout.

Decide whether a named release candidate (version, build, commit, or digest) of an app, service, or library is ready for production in a named environment: an evidence-backed GO / GO_WITH_CONTROLS / NO_GO / DEFER verdict across product, QA, security, ops, docs, billing, and support. Do not use for whole-repo baselines (repo-to-roadmap), weekly prioritization (product-operator), QA with no named candidate (web-app-auditor), or gating a report or ebook (artifact-acceptance). Use for "is this build ready to ship?" and pre-deploy gates.

Convert findings from reviewers, roasters, audits, tests, or acceptance gates into a minimal dependency-aware repair set, optionally apply authorized edits, and verify that fixes close root causes without introducing regressions. Do not use to invent new requirements, perform the initial broad audit, make consequential strategy decisions, or claim a finding is fixed without fresh verification evidence. Use when the user wants issues actually fixed rather than merely analyzed, especially after content-reviewer, content-roaster, science-roaster, repo-roaster, or artifact-acceptance.

Run evidence-anchored adversarial reviews of software repositories, codebases, monorepos, pull requests, branches, modules, architecture, tests, configuration, CI/CD, migrations, data pipelines, infrastructure, and engineering hygiene. Do not use as the primary skill for whole-project roadmapping, external-product pattern extraction, runtime web QA, or final release GO/NO_GO; route those to repo-to-roadmap, product-teardown, web-app-auditor, or release-readiness. Use when the user asks to roast, tear apart, red-team, stress-test, re-check a repaired repo, or brutally critique a repo/codebase and wants concrete file/symbol evidence, critical-invariant reasoning, trust-boundary and state-transition analysis, execution-path reachability, blast radius, false-positive checks, repairs, and executable verification steps.

Analyze an entire software/product project and turn verified project truth into an evidence-based, dependency-aware, reusable roadmap. Do not use for weekly sprint control, release-candidate GO/NO_GO gates, or "is build v1.2.3 ready to ship?" — use Product Operator or Release Readiness instead. Do not use as an implementation agent, narrow code review, or specialist security/SEO/CRO audit. Use for first-time or delta whole-project baselines such as "analyze the whole repo/project", "what is left to build before target state", "create/update a roadmap from GitHub/Notion", "przeanalizuj całe repo", or "update the roadmap after recent changes". Inventory topology before searching, separate intent/presence/behavior/release/outcome truth, prove material absence instead of inferring it from search misses, expose coverage and contradictions, model capabilities/critical journeys, prioritize target blockers and gates, validate hard dependencies, and create immutable baseline/delta roadmap snapshots.

Plan and govern research programs from methodology through study execution, analysis, manuscript readiness, and next-study decisions. Use for research stage gates, follow-up studies, estimands, hypotheses, falsifiers, stop or continue criteria, dependencies, evidence requirements, and unresolved governance. Do not use for one-off claim verification, adversarial manuscript review, final publication formatting, portfolio allocation, or consequential decisions.

Design, normalize, lint, and freeze evaluation rubrics before substantive review or benchmarking, with observable criteria, explicit pass/fail semantics, evidence floors, materiality, blocker rules, scope boundaries, and immutable rubric hashes. Do not use to judge the candidate itself, to write the artifact, to run empirical skill benchmarks, or to move criteria after seeing the result; use the relevant reviewer, artifact-acceptance, skill-evaluator, or quality-loop-operator for those tasks. Use when the user asks to create grading criteria, an evaluation rubric, acceptance scoring framework, reviewer scorecard, red-team rubric, or repeatable assessment contract.

Run evidence-anchored critiques of scientific manuscripts, papers, protocols, theses, methods, analyses, reviewer responses, validation studies, and research drafts. Do not use for generic content critique, repository/code review, one-off claim verification, or research-program planning; route those to content-roaster, repo-roaster, evidence-researcher, or research-program-operator. Use when the user asks to roast, peer-review, red-team, stress-test, re-review a revision of, or challenge scientific work and wants criticism anchored to source evidence, inferential validity, counterevidence, minimum repair burden, and an observable verification condition.

Run evidence-governed, multi-pillar website visibility audits across technical SEO, relevance, authority/trust, GEO/AI citation readiness, and AEO/answer extraction. Do not use for isolated schema/meta tasks, ongoing competitor monitoring, whole-product roadmapping, or release-candidate GO/NO-GO; hand accepted findings to the corresponding specialist. Use when the user explicitly asks for SEO+GEO+AEO or "maxxing", a broad end-to-end search/AI visibility audit, cross-pillar diagnosis, same-rubric competitor comparison, or repeat/delta audit. Use PILLAR only when this skill is explicitly requested for one pillar. Diagnosis only; never mutate live sites.

Audit an Agent Skill or skill library for trigger quality, scope overlap, instruction conflicts, context cost, broken references, eval blind spots, false-green paths, version and contract compatibility, host-support overclaims, package hygiene, and registry/docs/archive drift. Do not use to create or edit the skill (skill-creator), to benchmark behavior with vs without it (skill-evaluator), to audit ordinary software (repo-roaster), or for a release verdict. Use to audit, review, roast, or harden a skill itself.

Design and evaluate Agent Skill experiments that measure whether a skill improves model behavior, discovery, task success, reliability, cost, or latency relative to a no-skill/prior-version baseline, including host/model comparisons, judge agreement, quality-cost Pareto trade-offs, and runtime drift under frozen measurement identity. Do not use for static package/routing audits (use skill-auditor), to create/edit a skill (use skill-creator), to fabricate real-host results, or to claim universal superiority from one configuration. Use when the user asks to benchmark, A/B test, evaluate, compare, prove, regress-test, or measure a skill across supported harnesses, or to prepare executable real-host suites when execution is unavailable.

Plan and execute multi-skill CometWeb workflows with CW-AIP handoffs when the user wants one entry point instead of tagging each specialist. Do not use for a single-domain task when one specialist skill is enough, for routing-only questions without execution, or to bypass AI Council decision policy or Release Readiness gate rules. Use when the user asks to orchestrate, sequence, or run end-to-end flows (e.g. evidence then Council, audit then release, research then weekly ops), run everything needed for a goal, or chain Evidence Researcher with ai-council or other skills in order. Supports execution_mode auto|single_thread|isolated_subagents. Loads each step's SKILL.md, runs it fully, and passes envelopes between steps.

Alias for skill-orchestrator with execution_mode=isolated_subagents. Do not use when a single specialist suffices, when the host has no subagent/Task API (use skill-orchestrator single_thread), for routing-only questions, or to bypass AI Council or Release Readiness policy. Use when the user asks for multiagent orchestration, separate agents per skill, Task/subagent per step, or strict isolation between Evidence Researcher, AI Council, auditors, and other skills.

Evidence-driven QA and product audit for user-facing websites and web applications. Do not use for backend-only review, greenfield implementation, whole-repo roadmapping, critique of page copy or offers (content-roaster), search/AI-visibility audits (seo-geo-aeo-maxxing), or as the final production-release gate; provide candidate-bound QA evidence to Release Readiness when relevant. Do not perform penetration testing/exploitation with no product-audit goal. Use when asked to audit, review, inspect, QA, verify, click through, test, or "roast" a site/app/page/dashboard/ checkout/form, including UI/UX, accessibility, regression, data integrity, and critical flows. Supports browser, browser+source, screenshot-only, source-only, and fetch-only environments.