The most frequently loaded skills were workflow support: reporting, delegation, access, publishing and QA. The clearest opportunities are repeated setup and duplicated handoffs.
Distinct local root chats with a confirmed skill instruction read, including native children. Multiple reads in one chat count once. A load does not prove the complete workflow ran.
| # | Skill | Chats |
|---|---|---|
| 1 | wrap-up-lite | 122 |
| 2 | to-my-parent | 94 |
| 3 | wrap-up | 92 |
| 4 | launch-agent | 86 |
| 5 | look-for-access | 79 |
| 6 | comment-html-review-host | 73 |
| 7 | luna-visual-qa-section-screehoshots | 68 |
| 8 | wrp-voice-summary | 63 |
| 9 | am-i-blocked | 43 |
| 10 | tts | 43 |
| 11 | extract-gist-of-chat-jsons | 36 |
| 12 | comment-html | 35 |
Priority is a judgment based on observed friction and repeated use. These are proposals for the original seven skill files; those files were not edited. The three prioritized autonomous skills were enhanced below. No savings were measured.
launch-agentLoad the unchanged launch skill once per parent context. Each child still gets a fresh, concise assignment with inputs, ownership, completion checks and one return channel.
Observed: Two adjacent QA launches reread identical launch instructions. One mobile child tried an unsupported app return before returning natively. Bounded research and Ads extraction succeeded and should stay.
to-my-parentFor a native subagent, its final response is the parent return. Do not also send a thread message. When the parent can read the canonical report, send its path plus the decisive result, checks and limits.
Observed: One QA child used a thread message and a native final. Other sampled children returned useful summaries linked to the same saved report.
wrap-up-litePut the delegated-child exemption first, then return. Children use their assigned report. Keep the human handoff and audio rules for the parent.
Observed: All five sampled cases used the child branch. One footer final repeated its summary before the full report. These samples do not support changing the human audio policy.
wrap-upUse a small core: status/result, decisive evidence and limits, next action. Add identity, spend and recovery detail when relevant. On correction, update the same file and open the completed revision once.
Observed: A footer fix read this skill twice and opened the report twice. A provider review reused loaded instructions, updated the same report and retained broader test failures.
comment-html-review-hostGive related pages one publishing owner and one assembled deployment. Keep stable document IDs. Verify deep URLs and the manifest, since the review-domain root redirects. Use intercepted comment storage for routine tests.
Observed: A disposable page required preservation reconciliation and --force cleanup. Another task reconstructed a macOS locking workaround. Same-URL revisions worked. Live root redirect was also verified during this audit.
luna-visual-qa-section-screehoshotsCheck one screenshot and its page/viewport/state filename first. Hide the review toolbar for content screenshots. Batch already-known related fixes before fresh deployed QA. For routing-only changes with unchanged UI, use routing checks and a smoke render.
Observed: Screenshot descriptions were attached to wrong filenames, review controls obscured content, and equivalent header rules were repaired in separate rounds. A routing-only change launched a full two-worker visual suite. Actual UI QA caught real defects, so keep it.
look-for-accessEnforce the existing private-only trigger and route-reuse rule in callers. Public pages need no access-skill load. Add one 401 recovery example: verify a saved alternate with a harmless read, then rerun without the rejected process credential.
Observed: One public-comment step loaded credential-discovery instructions unnecessarily. Another task reread them three times in a short span. A failed process credential was replaced by a saved alternate that passed a real read.
Updated canonical sources: LLL now states the exact visible change and continues ordinary recovery within the approved goal after Launch. Unhinged proves one normal workflow before scaling, uses compact evidence and avoids equivalent retries or stacked voice recaps. CCC preserves explicit model choices, stores checkpoint timing/ownership, checks original requirements and verifies cleanup receipts.
The main SKILL.md files total 4,042 → 1,449 words. Conditional detail moved into shared references; this is not a measured token-saving claim. The board template is bundled unchanged for portability. Archived prior packages have verified SHA256 checksums.
Manifest validators and relative-reference checks pass. Independent scenario review is recorded in the local change report. Live unattended scheduling and Claude Code execution remain untested. These were outside the top 12 by frequency and have five source examples each.
lll · 23 chatsPrepare a concrete review contract using the canonical board. For every deliverable show the current problem, exact change, visible location and a representative test before the detailed proof. Preserve settled user scope, name unresolved targets and keep all unproved checks unchecked. Publish one stable board and stop for review. Only a later explicit Launch starts the single executor, with any explicit user model override preserved.
Five distinct root chats selected from actual skill loads during the audited week. This is a qualitative sample, not a failure-rate estimate. Current skill read September 30. No savings measured. This evidence audit describes the pre-enhancement instructions; no Claude copies made.
lll-unhinged · 21 chatsInfer the useful end result and execute within the user’s authority. Use one compact contract and stable public board. Prove one representative normal workflow first, including content/design acceptance when relevant, then expand the batch. Wait passively for the sole executor’s report, recover material blockers with a new permitted route, and verify actual artifacts/provider readbacks. Separate tested, deployed, sent, serving and business outcomes. Honor later user stops or scope changes immediately.
Five distinct root chats with verified skill reads and relevant operational work. Missive is explicitly a simplified planning request followed by Multi LLL execution, so it is not a clean standalone Unhinged run. No causal claims about model superiority or measured token savings. The reported tests/outcomes were source-checked, not re-executed.
ccc · 5 chatsApply CCC to one authorized executor, preserving explicit user model/effort and scope. Record the task, host, board, report, schedule and next permitted checkpoint in a small state file. Stay idle until launch+20 minutes, then inspect once; repeat no earlier than 60 minutes after each completed check. At a due check compare the actual progress and source requirements with the contract, send at most one useful correction and make at most one justified clarification. Keep independent work moving, and recover failed critical paths through a distinct permitted route. When all presently required outcomes are proven, publish the pink summary with optional next-direction prompts, archive the accepted executor and delete its schedule with receipts. Keep future outcomes and exact blockers explicit.
Read-only audit of all five weekly root chats with confirmed CCC loads. Call IDs reconcile every load’s source line. Quotes and metadata checked against the frozen UTC window. Only one complete lifecycle is demonstrated, and that result was subsequently rejected by the user. The current skill read for this audit is the pre-enhancement version; parent owns canonical edits. No cost/token-savings measurement or live provider recheck.
Screened from this week’s observed skills and current source. These additional skills were not five-example audited. The table is a source-backed recommendation, not a completed migration.
Seven transfer priorities: lll, lll-unhinged, ccc, sales-prep-kit, the offer-in-email-kit + contract-to-docuseal-kit bundle, get-leads, and the customized ads family. This is a judgment about useful capabilities and current sources, not a measured success ranking.
| Skills | Why carry them over | Claude Code adaptation |
|---|---|---|
ccc | Scheduled autonomous follow-through: first checkpoint after 20 minutes, then hourly; verify original requirements before acceptance and cleanup. | Newly enhanced canonical source. Bundle the LLL shared references; supply a real Claude Code checkpoint scheduler, native status/return and archive adapter. No Claude runtime execution has been tested. |
sales-prep-kit, sales-kit-runtime | URL-to-sales-kit workflow with measured research, two designs, stable review links and one Figma handoff. Runtime already defines Windows/macOS path and publishing adapters. | Keep package scripts/templates and runtime setup. Map Codex browser/Figma UI integration to available Claude tools, parameterize receiving operator and model routing, and preserve source engine dependencies. sales-kit-runtime is internal support, not a fourth operator command. |
offer-in-email-kit, contract-to-docuseal-kit | One commercial case ledger continues from exact transcript through offer, review, final PDF, DocuSeal and unsent mailbox draft. Kits already separate local CLI from provider writes. | Keep both folders together plus sales-kit-runtime, find-transcript, look-for-access, review host and PDF dependencies. Replace Codex previews with available browser/file opening. Use verified Missive API and DocuSeal region; keep exact source/hash and signer-role checks. |
get-leads | Agency-specific reviewed sourcing, enrichment, personalization, suppression, human row edits and run/spend ledger. High business value despite substantial runtime coupling. | Adapt orchestration and model IDs, preserve store/ledger/schema and resume contracts. Verify project-local runtime and devbox dispatch dependencies before transfer. Refresh dated cohort rules from current task. Do not promise a self-contained command by copying only SKILL.md. |
cold-email-sequence-review | Checks saved sequence, all personalization/random branches and actual lead previews separately from campaign sending. | Business contract is portable. Retarget authenticated provider reads/writes and preview browser; retain exact campaign/variant IDs and saved-version readback. |
campaign-clone-sample-review | Narrow paused campaign cloning and lead sample readback preserves follow-ups, settings, source records and review state. | Use current task sample size/parallelism rather than historical ten-worker example. Requires outreach provider plus get-leads run ledger and review host. Replace devbox/native worker launch with available execution route. |
notion-client-lifecycle | Useful shared CLI for exact client resolution, idempotent Activity logging and verified CRM snapshots. | Keep Python library and V1 property contract, configure exact current databases/accounts, and serialize writes per client. Use same canonical timestamp conversion; port has no reason to change that provider behavior. |
szamlazz-invoice | Real invoice and storno API scripts, redacted dry run, PDF receipt and filing convention form a distinct valuable business command. | Parameterize credential location, current seller tax treatment, number series and Drive folder/connector. Replace unmeasured 96% certainty language with exact legal-recipient evidence. Preserve issue/void evidence and task-specific binding-commitment authorization. |
docuseal-cli | Provider-independent CLI protocol and concrete PDF/DOCX/HTML field references support signing-package work. | Install the actual CLI separately, verify region/account and current flags. Keep reference pack; do not equate template creation or signing URL with invitation or signature. |
ads, ads-google, ads-budget, ads-attribution, ads-optimize, ads-creative, ads-landing | Evidence-first Ads audit family already derives from claude-ads while retaining agency account, measurement, budget and business-result rules. | Keep this customized canonical family, shared references, finding schema and scoring registry rather than reinstalling upstream over it. Wire current read-only Google Ads API/export. These skills do not supply a live mutation adapter. |
google-ads-lessons-for-agents, meta-ads-lessons-for-agents | Observed historical lessons encode account mistakes and current skill maps not supplied by generic ads tooling. | Bring source references as private supporting context and consult only relevant lessons. Refresh dated account facts from provider; do not replay past task authorization. |
meta-ads-operator | Deterministic Meta API wrapper supplies pagination, preflight, paused writes and receipts; a different function from the read-only ads audit family. | Preserve script and its mode contract, parameterize dotenv/account configuration, and verify current Graph API/account route. Existing validation is part of source behavior; no new security layer is being proposed. |
tracking-verification | Separates interaction, tag firing, network receipt, analytics, destination record and attributed outcome. | Retain controlled test marker and evidence ledger. Supply available browser/network inspection and correct Tag Assistant/GA4/provider identities; do not collapse stages into a passed tag screenshot. |
kpc, keywords-performance-control-setup, keywords-performance-control-refresh, keywords-performance-control-synthesis, keywords-performance-control-advice | Frozen Ads snapshot, deterministic SQLite dashboard and separately grounded model advice are niche and valuable beyond generic audit prompts. | Transfer five-skill family with Python scripts/config/schema/templates as one installed unit. Replace only Sol advice routing and Codex browser opening. Keep read-only account checks, snapshot lineage and validated advice JSON. Dashboard generation is not Ads mutation. |
riport-ppc-seo-cro-30-days | Client reporting combines actual paid-media/SEO windows and landing-page evidence with qualified-outcome limits. | Keep source references and report templates, replace named worker/model routes, resolve client API exports and OpenSEO connection, and preserve paid-data source windows. |
riport-4-decks, riport-4-decks/subskills/next-period-strategy | Useful report-pack rendering and next-period strategy workflow builds on source evidence and exact media manifests. | Keep parent pack and subskill together; adapt named Astra reviews and audio/session integration. Python renderer, templates and hash manifest remain package dependencies. Use only when four decks are the requested form. |
scrape, image-catalog | Structured site crawl plus evidence-linked raster catalog has actual code, manifests, image counts, hashes and useful downstream sales/design inputs. | Keep scrape.v2 and catalog schemas and scripts. Parameterize fixed Agty/Desktop paths and publishing project through sales-kit-runtime; map Luna vision role to available image-capable reviewer. Preserve failed rows and source coverage limits. |
scrape-seo-volumes, seo-keywords-search-volumes | Produces measured scrape/keyword/CPC source bundles used by sales preparation. | Transfer producer scripts and source contracts together, parameterize local paths and actual Ads/DataForSEO access. Preserve unknown values and national/local distinctions, not model estimates. |
lll | Review-gated goal contract and verification board are a distinct observed workflow. | Preserve Prepare/Launch semantics when invoked under that explicit mode. Bundle the currently external board template. Replace pinned Codex session/model/todo/wait APIs with supported Claude execution and durable file state. Never describe Codex pinning as a generic Claude command. |
lll-unhinged | Autonomous goal inference, observable criteria, evidence board and bounded recovery are distinct from review-gated lll. | Keep autonomous mode and proof criteria separate from lll. Bundle external template, replace Sol/Astra/Luna routes and Codex task-thread waiting/recovery with actual target capabilities. Do not claim model names, screenshot agent or retries survive a verbatim copy. |
launch-agent, to-my-parent, dont-supervise | Bounded ownership, self-contained goal, exact parent/report return and passive completion define useful execution primitives. | Consolidate transport adapter for Claude native agents/background tasks. Keep business authority and report ownership. Replace Codex thread IDs/host/pinning tools, native wait calls and app messages with verified target return/wait semantics. |
am-i-blocked | Structured goal/authority/workaround check avoids escalating fixable tool problems. | Mostly prose portable. Preserve affected-action scope and factual blocker evidence, use available target tools, and avoid turning the questionnaire into a compulsory approval gate. |
document-changelog | Small two-file CHANGELOG/PROJECT_STATE contract makes future work resume with current facts. | Retain project-local files and recorded real state, replace product-specific wording only. No extra infrastructure. |
archive-skill | Checksum-verifiable package archive before replacing a personal skill supplies practical recovery. | Keep archive script/tests and canonical root parameter. Map target discovery path without copying an entire Codex runtime. |
writing-for-agents | Clear goal/source/constraint writing improves reusable instructions across harnesses. | Portable prose. Replace Codex-specific examples only where they name actual target tooling, keep references/examples. |
domain-modeling | Active domain vocabulary and proportionate ADRs are useful coding discipline despite only one root use. | Keep CONTEXT/ADR formats and repository conventions, no model or connector adapter required. |
explain-lite | Compact end-state/route Mermaid decision map is useful in planning and human handoff. | Preserve focused diagrams and template, replace in-app open action, use review host for deliverable HTML. |
handoff-to-dani | Verified branch/resources and Windows/WSL Claude Code handoff already target the intended receiving environment. | Keep validation/push helpers, configure actual recipient and project branch; replace only launcher/browser/app integration and do not send unless recipient/channel/purpose authorized. |
find-transcript | Exact source retrieval rule matters to offers/contracts and currently routes to Wispr Flow. | Map local/export lookup and available authenticated tools. Keep Wispr-only current source rule and full-call requirement. Do not restore outdated Fireflies dependencies found in old notes. |
cloudflare-staging | Concrete static deployment helper yields verified publicly reachable client staging URLs. | Keep deploy script and its existing expected environment, parameterize account/project/paths and CLI availability. Separate client staging from shared review publication. |
impeccable, clone, prototype, make-design-md, marketing-ui, gsap, copywriter | Observed design/build skills provide distinct parts of design production: critique, reference recreation, question-driven prototype, source design spec, marketing surface, animation and copy. | Retain distinct entrypoints rather than one enormous prompt. Read their source references on demand, replace browser/model integrations, and configure image/video generation only when actually available. Preserve source design tokens and question-specific prototype mode. |
daily-digest, plan-tomorrow-and-reflect, planning-session-done, daily-audio-guides | Personal planning family uses exact windows, live evidence, private plan state and intentional session closure. | Keep runtime and source-lane contracts, adapt native voice/Calendar/computer-history integrations to available Claude tools. Unavailable personal sources remain explicit gaps. Keep audio optional according to user instruction. |
tts, audio-to-html | Gemini speech CLI and synchronized HTML player are useful media capabilities independent of Codex conversation UI. | Keep TTS runner and HTML runtime assets. Parameterize current provider model/voice and publishing; explicit task delivery controls override normal Telegram/public-library defaults. Remove proprietary live voice-agent destination when unavailable rather than fake a working link. |
look-for-access | Observed access recovery is widely reused and mostly harness independent. | Keep dotenv loader and exact-account harmless read contract. Replace Codex in-app/CLI browser selection and connector discovery with available Claude routes. Private access records remain local. |
extract-gist-of-chat-jsons | Source-traceable user/assistant extraction is valuable for audit, instruction provenance and handoff. | Current parser accepts Codex session schema. Add/test a Claude transcript parser against real target records before claiming target support; keep exact file/line lineage and exclusions. |
setup-openseo, seo-project-setup, seo-audit, seo-report, keyword-research, keyword-clustering, local-seo, seo-coach, seo-tactics | Observed OpenSEO workflow family already names a Claude Code plugin route, shared project context and provider tools. | Reinstall/connect official OpenSEO target package first and compare canonical customizations before replacing files. Retain agency report/claim rules where needed. Do not port obsolete Codex template-literal HTML restrictions as generic Claude limitations. |
comment-html-review-host, comment-html | Stable document identity, comments and inline edits make human review usable across revisions. | Keep shared comment runtime and publisher as one implementation with a stable-host mode, plus scripts/templates/worker and configured existing Cloudflare access. Use sales-kit-runtime dedicated project on a new machine to avoid replacing a partial shared mirror. Replace Codex browser open/readback with actual target tools. |
luna-visual-qa-section-screehoshots | Section and multi-section screenshots with fixes and fresh postdeployment inspection are a portable visual verification contract. | Keep screenshot coverage and actual raster inspection. Map Luna named roles and Codex browser/context tooling to available image-capable agents/browser tools; preserve explicit user QA scope. No claim that a browser or model integration is installed by copying this manifest. |
Other useful candidates include KPC snapshots and dashboards, PPC/SEO reports, measured keyword research, image catalog, transcript retrieval, Notion lifecycle, DocuSeal CLI and invoice scripts. The local transfer inventory also dispositioned 51 consolidation/subordinate sources and 23 unavailable, conditional or lower-priority names.
Known package gaps: LLL’s board template lives outside its skill folder. Sales runtime onboarding references missing ZIP-level installer resources. The Missive reply drafter needs its owning app/config and drafting rules recovered. These are implementation dependencies, not working ports.
Keep exact authority and completion boundaries, one canonical report, stable review URLs, preservation of unrelated work, real deployed UI checks, and honest failed-check disclosures. The five reporting examples cover the child branch, so they do not justify changing the human audio policy.
The window is September 23 at 08:31:56 through September 30 at 08:31:56, Europe/Budapest. Local active and archived Codex files were covered. Remote-only chats were not. Literal-path detection can miss injected or variable-only loads. The selected examples are not a random sample.
All 149 observed names were screened for Claude Code transfer: 75 candidate sources in 38 groups, 51 consolidated/subordinate sources, and 23 unavailable, conditional or lower-priority names. No Claude Code runtime test was performed. No percentage token savings, cost savings or frontier-model performance gain was measured. A repeated read may be justified after context loss or an instruction change. Current skill revisions already address parts of the historical friction.
For the three enhanced skills, compare the next real run; for the remaining suggestions, compare after implementation: instruction reads, tool-output size and successful completion. Keep any extra instructions that protect the actual outcome.
The observation-led advice is consistent with official OpenAI guidance on narrow triggers and loading references only when needed. Model names and tool availability need a Claude Code adapter. Do not assume a named Codex reviewer exists there. The private appendix also links the visible weekly dialogue transcripts. Personal skills and isolated skill execution are documented in the official Claude Code skill documentation.