Jump to section tap to open
The setup · Claude Code half

Vlad's actual Claude Code, snapshotted.

Chapter 5 explains skills as a concept. Chapter 11 walks through building one. This page is the receipts — the 54 skills, 0 custom agents, and 10 plugins running live in Vlad's ~/.claude/ right now. Auto-generated from the actual filesystem; updates as the stack evolves.

The point isn't that you should install all of these. The point is that this is what an operator's library looks like at hour two thousand. Steal anything. Skip anything. Build your own. The auto-extracted list is just names — scroll past it to the six swarm compositions first to see how the pieces actually chain together.

And it's a garden, not a museum: the Self-Audit pruned it on 2026-06-09 — fifteen dead skills archived on telemetry evidence, every trigger description re-trimmed, one plugin benched. The 2026-10-24 refresh counts 36 more skills gone, 24 new, and an empty custom-agents folder — Claude Code reads that folder recursively, so even an _archive/ inside it kept loading, and the agents moved out. A library that only ever grows is a library that's rotting.

54 skills · 0 custom agents · 10 plugins
Why skills
Ch 5 — Recipes the chef reads
How to build one
Ch 11 — Build a skill end-to-end
Sister page
Cowork setup (sanitized) →
Compositions · how the pieces chain

Six swarm compositions Vlad actually runs

The names in the list below are inert until you wire them into a Wave-1 / Wave-2 / synthesis pattern. These are the six chains that earn their rent — each one a real workflow with real receipts, not a marketing diagram. Read Ch 6 for the swarm philosophy and Ch 16 for the hooks + subagents wiring.

Strategic-plan swarm

Greenfield venture — go from one paragraph of intent to a 25-document strategic plan before lunch.

Trigger · manual · Vlad says "draft me plans for [new venture]" → /swarm-strategic-plan fires
  1. Wave 1 — foundation
    4 agents in parallel: VISION, BUSINESS-MODEL, MONETIZATION, MARKET. Read/Write/WebSearch each. Each writes one master doc.
  2. Wave 2 — ops
    4 agents: ROADMAP, GTM, OPERATIONS, RISK. Read the Wave-1 outputs as context, write their own.
  3. Wave 3 — brand + voice
    4 agents: BRAND-IDENTITY, MESSAGING, CONTENT-PILLARS, VOICE-GUIDE.
  4. Wave 4 — revenue
    4 agents: PRICING, REVENUE-MODEL, SALES-PLAYBOOK, RETENTION.
  5. Wave 5 — tech + personas
    4 agents: ARCHITECTURE, STACK, 4 buyer personas. Final synthesis writes the BRIEF.md tying all 20+ docs together.

Receipt · Fired this for the 10k MRR ideas portfolio scaffolding — 25 docs in ~14 min wall-clock, ~$8 of tokens. Saved the week of "where do I even start" that kills every greenfield idea.

Don't use this for · Don't use this for a feature inside an existing product — the 5-wave depth is overkill and the docs won't fit the repo's existing voice.
→ Ch 6 — The swarm

6-agent codebase audit

Pre-PR or pre-investor-demo sweep on a repo nobody has audited end-to-end in months.

Trigger · manual · /audit on the repo root
  1. Wave 1 — 6 parallel reviewers
    security · types · performance · dependencies · test coverage · dead code. Each gets Read + Grep + Glob (read-only). Each returns findings as file:line with severity.
  2. Wave 2 — synthesis (mine)
    I dedupe overlapping findings, prioritize by blast radius × fix cost, write the fix order. Subagents do not edit — the main session applies fixes in one pass.

Receipt · A B2B SaaS pre-launch audit: 10+ real bugs surfaced per pass — 3 NestJS injection footguns, a BSON cast that nuked auth on a specific email shape, two dead workspace routes. Wall-clock ~9 min vs. half a day of manual grep.

Don't use this for · Don't run this on a 200-line repo — the per-agent overhead dwarfs the work. Use bughunter solo under ~500 LOC.
→ Ch 16 — Hooks + subagents

Editorial swarm

Naming the chapter, the product, the tier — anywhere a single answer would mask the decision.

Trigger · manual · "give me 3 editorial takes on this"
  1. Wave 1 — 3 framings in parallel
    Agent A: topical (lead with the noun). Agent B: action (lead with the verb). Agent C: question (lead with the reader's objection). Same brief, same data, three philosophies.
  2. Wave 2 — pick or hybridize (mine)
    The output is a decision surface, not a single answer. I usually steal the verb from B and the noun phrase from A.

Receipt · Retitled all 36 chapters of this book for Edition 4 — 3 candidates per chapter, kept ~60% topical-primary, ~30% action-primary, ~10% question. Topical won on SEO + scannability.

Don't use this for · Don't use this for technical accuracy questions — three opinions on a fact is just three guesses. Reserve for style, voice, framing.
→ Ch 6 — The swarm

Design-variant swarm

"Redesign this page" — when one aesthetic answer hides the better one.

Trigger · manual · "give me 3 design variants of this section"
  1. Wave 1 — 3 designers in parallel
    Same data inlined into each prompt (never read from disk — variance kills comparison). Agent A: Bold. Agent B: Editorial. Agent C: Dashboard. Each gets Read + Write + a self-contained design brief from the 11-brief reference library.
  2. Wave 2 — comparison (mine)
    View the three variants side-by-side, pick one, or fuse (Bold typography + Editorial spacing is a common steal).

Receipt · A creative-agency proposal page: spawned Bold/Editorial/Brutalist variants in one wave. Editorial won the agency vote, Bold won the client click-through, shipped the hybrid. Each variant cost ~$0.40 in tokens.

Don't use this for · Don't use this when you already know the constraint (e.g. "must match Belkins brand"). Variant swarms are for exploring — not for executing.
→ Ch 6 — The swarm

Pre-flight external-deps swarm

Before writing a single line of code that depends on 5+ external APIs — verify each is still alive in 2026.

Trigger · manual · /preflight-external-deps at the start of any integration-heavy project
  1. Wave 1 — N agents, one per dependency
    One agent per named API/data source. WebSearch + WebFetch. Each verifies: still active? rate limits changed? auth flow same? pricing tier still affordable? 2026 deprecation notes? Returns a one-page viability card.
  2. Wave 2 — go/no-go matrix (mine)
    Tabulate viability. Anything yellow gets a fallback plan documented before the first commit. Anything red kills the dep or the project.

Receipt · 10k MRR ideas portfolio pre-flight: 6 show-stoppers caught in one ~15-min pass — a deprecated TikTok endpoint, a Stripe Tax country gap, an Apify actor that 403'd on the listed plan. Saved 2-3 weeks of scaffolding against a dead dep.

Don't use this for · Don't use this for a project with 1-2 well-known deps you've used this quarter. The swarm overhead beats a single WebSearch only when N ≥ 4.
→ Ch 16 — Hooks + subagents

Multi-AI weekend-build swarm

Saturday-ship project where you want to know which AI provider's vibe wins on the same brief.

Trigger · manual · Saturday morning, one PRD, three providers
  1. Wave 1 — same brief, 3 providers
    Claude (architecture-first), GPT (feature-completeness-first), Gemini (UI-polish-first). Each gets the same one-page PRD, the same target stack, the same 4-hour budget. Each ships to its own branch.
  2. Wave 2 — adversarial review (mine)
    Compare the three branches: which one merges cleanest? which has tests? which UI ships without 5 more hours? Cherry-pick across — Claude's server logic + Gemini's components is a frequent merge.

Receipt · Weekend-builds page synthesis: three parallel passes (wide ideation, PRD shaping, adversarial rubric) — the rubric killed five "obvious" picks and pushed one boring one to the top. Shipped in one Saturday.

Don't use this for · Don't use this for production code on the critical path — three half-finished branches is worse than one finished branch. Reserve for greenfield Saturday builds where exploration is the point.
→ Ch 6 — The swarm

Six compositions, not sixty. The other patterns Vlad runs (memory-consolidation triplets, adversarial proposer/critic/arbiter trios, map-reduce email classifiers) are variants of these — once you can see the Wave-N shape, you stop counting agents and start counting outcomes.

Chains · skills in sequence

Eight skill chains Vlad runs

A skill on its own is a recipe card. A chain is a service ticket — three recipe cards pinned in the right order, fired by one phrase, finished before the operator's coffee cools. These eight scenarios show what actually happens on a Tuesday morning when the library is doing the work and the operator is just reading Slack. Real portfolios, real numbers, real failure surfaces.

Tuesday 06:35 ET

The deal-pipeline tuesday

Skills that fire (in order)
  1. deal-watcher — twice-daily HubSpot sweep that flags stalls (4+ days quiet) and surges (2+ stage jumps in 48h) per-deal, not as a digest
  2. friday-wrapup — last Friday's aggregator memo still pinned in #pipeline-watch, carries the "deals to revisit Tue" list
  3. belkins-sales-intelligence — joins HubSpot stage data with Gong transcript signals + calendar moves

The scene · 06:35. Coffee on, kid still asleep. Slack opens to two deal-watcher pings from the 04:00 scheduled tick — Acme Corp $42k quiet 5 days in proposal, Northwind $18k advanced two stages overnight (good news, but a last-mile slip risk). Below those, friday-wrapup from Apr 26 still pinned: "Tuesday revisit — three deals from last week's calls had no follow-up touch." belkins-sales-intelligence had already cross-referenced the Gong transcripts: two of the three had verbal commitments that never made it to HubSpot notes. Vlad reads for 7 minutes, fires 4 outbound DMs to Kate and the AEs at 06:48. Acme replies at 09:12, deal closed Apr 30.

Receipt · $42k deal recovered from the silent-dead window. 7 minutes of operator attention vs. the old 45-minute "read the dashboard, decide" Tuesday ritual.

Where it's brittle · deal-watcher posts per-deal pings; on a 12-flag morning the channel becomes a wall and the operator skims. If flag count > 6, the chain needs a fallback digest mode — currently absent.
Wednesday 14:00 ET

The newsletter ship loop

Skills that fire (in order)
  1. vlads-newsletter — drafts in Vlad's voice from vault sources (Vlad-Brain/Drafts/, recent retros, last week's calls)
  2. voice-calibration — pulls 3 sample paragraphs from the published archive, scores the draft against them on cadence + sentence-length variance
  3. rigor-enforcer-gate — kills "Key takeaways" lists, numberless claims ("many users…"), and the four banned superlatives

The scene · 14:00, draft prompt opens with "newsletter on why CFOs are mispricing AI tokens." vlads-newsletter pulls three angle candidates from the vault, picks the one with the strongest opening hook (the Folderly $9k/mo bill conversation from Apr 11), drafts 1,150 words in 8 minutes. voice-calibration runs at 14:12 — flags two paragraphs as "too LinkedIn thinkfluencer", suggests cuts. rigor-enforcer-gate runs at 14:25 — catches one "studies show" claim with no link, one "Key Takeaways" closer, three em-dash overdoses (rare for the gate to fire on em-dashes since Vlad likes them, but the threshold is 11 per piece). Vlad hand-edits for 22 minutes. Published draft in Substack at 14:47.

Receipt · 47-minute cycle from idea → published draft. Old cycle was 4 hours including the "wait, this sounds AI-written" rewrite pass. 1,150 words shipped at vladsnewsletter.com.

Where it's brittle · voice-calibration scores on syntax, not argument shape. A draft can pass voice and still be a topical retread of last month's piece — the chain needs a "have I shipped this argument before" check that doesn't exist yet.
Tuesday, 30 min before a recurring mentee call

The mentee tuesday prep

Skills that fire (in order)
  1. mentoring-lifecycle — pre-session mode: loads Session History + Action Tracker + Patterns file + last Session Prep for the mentee
  2. friday-wrapup — pulls any cross-portfolio context the mentee should know (e.g., a warm intro the mentee flagged in a prior session)
  3. vault-note — writes the 1-page prep into the mentoring folder as a dated Session Prep note with proper frontmatter + wikilinks

The scene · T-30. mentoring-lifecycle fires in pre-session mode, loads four files from the vault: the session-history table, the action tracker (a few commitments overdue on Vlad's side), the patterns file (the recurring behavioural pattern for this mentee), and the last session prep (a blocker discovered last time). friday-wrapup adds a cross-portfolio status item the mentee had flagged. vault-note writes the prep ~8 min in. Vlad reads for 6 minutes and walks into the call knowing the one load-bearing question.

Receipt · 1-page prep doc written in 8 minutes vs. the old 35-minute scramble through vault files. The retainer stays defensible because the session opens on the mentee's actual question, not the last mentee's context bleeding through.

Where it's brittle · mentoring-lifecycle assumes the vault paths are stable. When Cowork's bindfs mirror got re-synced on May 9 the chain returned an empty Patterns.md for 18 minutes — verify-file-durability now runs as a pre-check, but the chain doesn't fail loudly when a source file is silently empty.
Friday 16:45 ET

The CFO defense friday

Skills that fire (in order)
  1. billing-alert-template — already flagged the Anthropic spend spike Tuesday morning, $9,240 month-to-date vs. $5,800 budget
  2. cfo-case — defense generator, builds the "what this spend bought" memo from git activity + shipped artifacts + revenue-attached outcomes
  3. model-comparison — pulls Claude Opus 4.7 vs. GPT-5 cost-per-task on the actual prompts Vlad ran that month

The scene · 16:45. Slack DM from the CFO: "Vlad, the Anthropic bill is up 59% this month, can you justify before the board on Monday?" billing-alert-template had already pinged Tuesday — Vlad had ignored it because the spike was a swarm-strategic-plan run for a new portfolio venture, deliberate. cfo-case fires at 16:48: pulls git activity across four active repos, maps Claude usage to specific outcomes (a portfolio upsell recovered Tuesday, the 36-chapter book shipped, the payment links live). model-comparison runs at 16:55: same prompts on a competing model would have cost ~$14k with worse output on the long-context chapter drafts.

Receipt · 4-paragraph response to the CFO at 17:08, board memo attached. $9,240 bill defended with ~$74k of attached deal-and-shipped-product outcomes. 23 minutes total vs. the old "spend Saturday reconstructing what the money bought" pattern.

Where it's brittle · cfo-case maps spend to git activity, but a third of Vlad's Claude usage is conversational (mentoring prep, strategic thinking) with no git trail. The defense undercounts the actual value — currently the chain reports the floor, not the truth.
Saturday 10:00 ET, day before any public artifact ships

The pre-launch sweep

Skills that fire (in order)
  1. preflight-external-deps — spawns parallel research agents to verify every external API/data source is still alive and rate-limit-shaped as documented
  2. cross-trio-audit — reads landing + day-1 fulfillment page + welcome email side-by-side, catches tier/price/refund/cadence drift
  3. rigor-enforcer-gate — voice check + claim check on every customer-facing string before money moves

The scene · Saturday 10:00, the day before the 10k MRR Ideas demand-test decision (May 18 noon). preflight-external-deps spawns 4 parallel agents at 10:02 — one per external dep: Formspark (lead capture), Stripe Payment Links (18 of them), Resend (welcome email infra), Mailgun (Belkins sender for follow-ups). Wave returns at 10:14: Formspark's free-tier monthly cap was 250 submissions, the hub had logged 312 — silent throttle starting in 6 days. cross-trio-audit runs at 10:18 across the 10 subdomains: catches that 2 of them had a stale $19/mo tier on the landing while the welcome email quoted $29/mo. rigor-enforcer-gate runs at 10:32 on every customer-facing string — flags 3 "we'll get back to you soon" hedges that should name the actual SLA.

Receipt · 3 launch-killing drifts caught before the May 18 decision. Form provider swapped to a $9/mo tier by 11:00. Saved roughly 2-3 weeks of "why aren't leads coming in" debugging on a portfolio of 10 day-1 pages.

Where it's brittle · preflight-external-deps only checks deps that the operator named in the brief. If a dep is implicit (Vercel's Edge Config quietly underpins the redirect map) the chain misses it. Implicit-dep discovery is still a manual eye.
Wednesday 11:15 ET, week 2 of SDR onboarding

The sales-team adoption loop

Skills that fire (in order)
  1. mentoring-lifecycle — SDR-onboarding variant, loads the new hire's first-week artifacts (sent emails, recorded discoveries, HubSpot notes)
  2. email-deliverability-eval — gate before any SDR's first owned campaign — runs SPF/DKIM/DMARC + Postmaster reputation check
  3. customer-call-truth-extractor — Belkins SDR's first prompt — strips marketing hedge language out of discovery call summaries

The scene · 11:15. Belkins onboarded 3 new SDRs on May 1. mentoring-lifecycle (SDR variant) loads their week-1 artifacts at 11:17 — reads 47 sent emails, 6 Gong-recorded discoveries, HubSpot deal notes. One SDR (Jamie) had been writing notes like "buyer seems excited, likely close Q3" — empty signal. customer-call-truth-extractor flagged 9 instances of marketing-hedge language across her notes, suggested rewrites ("buyer named budget $40k, named decision date May 30, named blocker = legal review"). email-deliverability-eval ran on jamie's new sender domain before her first owned send Tuesday — caught a missing DKIM selector, fixed in 4 minutes before the 8AM send window.

Receipt · 3 SDRs in week 2, customer-call-truth-extractor stuck (used 38 times across the team that week), email-deliverability-eval caught the DKIM miss that would have torched a 2,400-message send. Week-2 patterns memo written for the next cohort.

Where it's brittle · the SDR-onboarding variant of mentoring-lifecycle exists as a mode toggle, not a separate skill — and the trigger phrases between mentee-prep and SDR-onboarding overlap. Wrong mode fires roughly 1 in 7 invocations; the operator catches it because the prep doc lands in the wrong vault folder.
Saturday 09:00 → 15:00 ET

The saturday product build

Skills that fire (in order)
  1. gstack-office-hours — builder mode, six forcing questions on whether the idea is worth the Saturday
  2. swarm-strategic-plan — 5-wave × 4-agent scaffold if the idea survives office hours
  3. agent-wave-verify — between-wave audit, catches silent agent failures + scope drift before wave N+1 fires
  4. pitch-html — single-file interactive artifact, email-attachable, no build step

The scene · Saturday 09:00. Idea: a "deal-cause-of-death autopsy" tool for outbound agencies — given a lost deal, generate the autopsy. gstack-office-hours runs at 09:05: six questions kill the "everyone wants this" assumption, narrows wedge to "Belkins-internal first, 30 lost deals/month, 15-min autopsy each." Survives. swarm-strategic-plan fires at 09:25, wave 1 (BRIEF + MARKET + PROBLEM + PERSONA) returns at 09:48. agent-wave-verify at 09:50 — wave 1 clean. Waves 2-5 fire across the morning with audits in between. agent-wave-verify catches wave 3 at 12:14: the monetization agent silently wrote to the wrong folder, scope-drift on positioning. Re-run wave 3 with corrected scope at 12:20. pitch-html runs at 14:30 with the BRIEF + MONETIZATION + REVENUE-MODEL + BRAND-IDENTITY → single .html file at 14:51. Sent to two Belkins ops leads at 15:00.

Receipt · 6 hours from idea → 20 strategic docs + 1 shippable HTML pitch. Wave-3 silent failure caught by audit before it cascaded into 4 downstream wave-4 docs writing against bad positioning. Roughly 3 hours of rework avoided.

Where it's brittle · swarm-strategic-plan assumes 4-agent parallelism stays under the empirical ceiling. Saturday morning's API load is low, but if a Saturday build coincides with a Belkins prod incident also burning agent budget, wave-2 agents start returning 429 — the chain doesn't backoff gracefully.
Monday 08:30 ET

The cross-portfolio hygiene check

Skills that fire (in order)
  1. friday-wrapup — Friday's aggregated state across Belkins / Folderly / LinguaLive
  2. cross-portfolio-check — runs context-leak detection: did a Belkins-flavored prompt accidentally reach a Folderly customer-facing draft, and vice versa
  3. claude-md-portfolio — verifies the portfolio CLAUDE.md instructions haven't drifted from the actual project state

The scene · Monday 08:30. friday-wrapup memo from Apr 26 still pinned, summarizes the 5 portfolio cos at end-of-week. cross-portfolio-check fires at 08:32, reads through draft Slack messages, draft emails, draft newsletter pieces from the weekend — flags one: Vlad's weekend draft for a Folderly partner outreach quoted a Belkins-internal closed-won number that's confidential. claude-md-portfolio at 08:45 — checks the per-project CLAUDE.md files against current state: catches that LinguaLive's CLAUDE.md still references the old payment provider (swapped to Stripe Connect on May 4), a B2B SaaS project's still says "v10.46.12" when prod is on 10.47.8.

Receipt · 1 confidential-number leak caught before send — would have been a Monday-morning lawyer email otherwise. 2 stale CLAUDE.md drifts fixed in 11 minutes. Cross-portfolio hygiene done before 09:00, instead of being the thing Vlad notices Wednesday when something has already shipped.

Where it's brittle · cross-portfolio-check matches on entity names (Belkins, Folderly, etc.) but doesn't catch paraphrased leaks — "an agency client of mine closed $X" reads as generic, fails the regex, ships. The chain catches the obvious leaks; the lawyer-grade ones still need a human eye.
Deep-dives · the spawn templates

Five swarm deep-dives

These five patterns are the ones I keep reaching for when the work doesn't fit in one context window. Each one has a real receipt — a moment it fired, what it produced, and what it would have cost to do the slow way. The spawn templates are really usable: paste, swap the bracketed names, send.

One rule under all of them: 3-4 agents per wave is the empirical sweet spot. 5+ invites filesystem contention. More agents in one wave doesn't mean more output — it means two agents writing to the same file, one agent silently winning, and you discovering it three hours later in a diff. Respect the ceiling.

Pattern 1: the strategic-plan swarm

When to use · you have a greenfield venture — a name, a hunch, maybe a sentence — and you need 25 documents (vision, market, monetization, ops, brand, tech, legal, persona pool) before you'll know whether it's worth a weekend.

Wave shape
  1. Wave 1 — Foundation (4 agents, parallel)
    Agent 1 (vision-architect): writes VISION.md — what this is, what it isn't, the one-sentence positioning. Agent 2 (business-modeler): writes BUSINESS-MODEL.md — value chain, unit economics frame, who pays whom. Agent 3 (monetization-strategist): writes MONETIZATION.md — pricing tiers, take rate, expansion vectors. Agent 4 (market-analyst): writes MARKET.md — TAM/SAM/SOM, 3 incumbents, the wedge.
  2. Wave 2 — Operations (4 agents)
    ops-playbook + sourcing-strategy + house-rules + legal-frame.
  3. Wave 3 — Brand + Access (4 agents)
    brand-identity + marketing-channels + access-model + atmosphere-design.
  4. Wave 4 — Revenue surfaces (4 agents)
    revenue-streams + sponsorship-deck + streaming-rights + merch-line.
  5. Wave 5 — Build + community (4 agents)
    tech-stack + persona-pool + community-loops + investor-deck.
  6. Synthesis (1 agent, runs last)
    reads all 20 docs, produces TOP-INSIGHTS.md (5 cross-cutting themes the individual specialists couldn't see) and a brainstorming-personas.md (8 voices to pressure-test the plan in subsequent sessions).

each wave waits for the prior to finish — wave 2's ops doc references wave 1's monetization assumptions, so foundation has to set before operations can build. within a wave, 4 agents max. that's the contention ceiling. if you try to do 6 ops specialists in wave 2, two of them will both try to edit ROADMAP.md and one will lose silently.

Spawn template (paste into Claude Code)
/swarm-strategic-plan [VENTURE-NAME]

Context: [one paragraph — what the venture is, who it serves, what makes it
different]. Stack assumption: [Next.js + Postgres / Notion-native / physical
space / etc.]. Scope: 25 documents, 5 waves of 4 specialists + synthesis.
Output folder: notes/strategic-plans/[venture-name]/.

Constraints: real numbers per claim, lowercase tolerant, em-dashes welcome,
no "amazing/powerful". Each doc 800-1500 words. Cross-link by relative path.

Receipt · ran this on 2026-04-22 for a hypothetical members-only ops salon — 20 docs in 11 minutes wall-clock, $6.20 total. synthesis caught that the brand doc's "exclusive" framing contradicted the monetization doc's volume assumption — would have been a month of confused positioning if i'd written them sequentially.

Anti-pattern (when NOT to use) · the venture has fewer than three real questions in it. if you can write the plan in one tab in 90 minutes, 20 specialists will produce 19 docs of filler around the one doc you needed. swarm cost > thinking-clearly cost.
→ Chapter 6 — The Swarm → /showcase swarm receipts gallery

Pattern 2: the 6-agent codebase audit

When to use · you have a diff (or a whole repo) and you need to know what's wrong with it across six dimensions before merging — and a single reviewer agent will read 9,000 lines and miss half.

Wave shape
  1. Wave 1 — Domain reviewers (6 agents, parallel, read-only)
    Agent 1 (security-reviewer): tools Read/Grep/Glob. Hunts secrets, injection sinks, missing auth checks, unvalidated input. Agent 2 (types-reviewer): hunts any, untyped returns, missing zod, prisma type drift. Agent 3 (performance-reviewer): hunts N+1 queries, sync work in hot paths, unbounded loops, missing indexes. Agent 4 (deps-reviewer): hunts deprecated packages, vulnerable versions, unused imports, version-skew between workspace packages. Agent 5 (test-gap-reviewer): hunts untested branches, missing edge cases, mocked-but-uncalled assertions. Agent 6 (dead-code-reviewer): hunts unreachable code, unused exports, orphan files, stale feature flags.
  2. Wave 2 — Synthesis (1 agent)
    reads all 6 reports, dedupes overlapping findings (a security issue is often also a types issue), ranks by severity × confidence, writes AUDIT.md with file:line citations and a top-10 fix list.

i know i broke my own rule with 6 in one wave. exception is legal here because each reviewer reads the diff and writes to its own report file (audit-security.md, audit-types.md, etc.) — no write contention. the synthesis agent in wave 2 reads all six and merges. if you tried 6 agents all editing the same audit file, you'd hit the contention ceiling instantly.

Spawn template (paste into Claude Code)
/audit [DIFF or PATH]

Spawn 6 read-only review subagents in parallel against [PATH]:
  - security-reviewer → notes/audit/security.md
  - types-reviewer → notes/audit/types.md
  - performance-reviewer → notes/audit/performance.md
  - deps-reviewer → notes/audit/deps.md
  - test-gap-reviewer → notes/audit/tests.md
  - dead-code-reviewer → notes/audit/dead-code.md

Each returns findings as: file:line | severity (P0/P1/P2) | description |
suggested fix. No edits. Read-only tools only (Read, Grep, Glob).

After all 6 return, spawn one synthesis agent to dedupe + rank + write
notes/audit/AUDIT.md with top-10 fix list.

Receipt · ran on a B2B SaaS pre-release branch 2026-04-09 — 6 reviewers returned 47 raw findings, synthesis dedupe collapsed to 12 unique issues, 4 were P0 (one missing auth check, two N+1s, one vulnerable dep). that single pass caught more than two human reviewers had in the prior week.

Anti-pattern (when NOT to use) · the diff is under 200 lines or single-file. overhead of spawning 6 agents costs more than reading it yourself. swarm pays off when the diff is across 10+ files or you're auditing a repo you didn't write.
→ Chapter 25 — Evals: Smoke, Regression, Golden → Chapter 16 — Hooks and Subagents

Pattern 3: the 3-agent perspectival editorial swarm

When to use · titling, voice calibration, slug-picking, headline copy. one agent will hand you a single answer. you don't want a single answer — you want a decision surface so you can pick by ear.

Wave shape
  1. Wave 1 — Three philosophical stances (3 agents, parallel)
    Agent A (topical-declarative): writes 10 candidates in topical-declarative form — "The N-Agent Codebase Audit," "Parallel Subagents and Fan-Out." Topic-led, no verb. Agent B (action-verb): writes 10 candidates in action-verb form — "Audit Your Repo with 6 Agents," "Spawn Six Reviewers, Merge One Report." Imperative, the reader is the subject. Agent C (question-form): writes 10 candidates in question form — "What If Six Reviewers Read Your Diff in Parallel?" "Why Does One Reviewer Always Miss the N+1?" Curiosity gap, no answer until the click.
  2. Wave 2 — Decision step (you, 90 seconds)
    read the 30 candidates side-by-side. pick by ear. there is no synthesis agent here — the value is the comparison, not the merge.

3 agents is the floor for this pattern, not the ceiling — 2 perspectives is a coin flip, 4+ stances dilute distinctiveness. 3 is the empirical sweet spot for editorial choice. (the contention ceiling doesn't bind here because each agent writes its own candidate list to its own file.)

Spawn template (paste into Claude Code)
Spawn 3 editorial subagents in parallel for [PIECE]:

Agent A — topical-declarative voice. Write 10 candidate titles in noun-led,
declarative form. No verbs in title. Examples: "The Swarm." "Hooks and
Subagents." Save to notes/titles/A-topical.md.

Agent B — action-verb voice. Write 10 candidate titles starting with an
imperative verb addressed to the reader. Examples: "Spawn 6 Reviewers."
"Stop Talking to Claude." Save to notes/titles/B-action.md.

Agent C — question-form voice. Write 10 candidates phrased as a question
that creates a curiosity gap. Examples: "What if six agents read your diff?"
Save to notes/titles/C-question.md.

Subject: [topic]. Audience: [who reads this]. Word budget: 4-9 words per
candidate. Return file paths only.

Receipt · used this on 2026-05-09 retitling all 36 chapters of this book. 3 agents × 10 candidates × 36 chapters = 1,080 options, generated in ~7 minutes. final picks were a mix: 22 topical, 9 action, 5 question. doing this with one agent would have produced 36 topical titles (the default voice for LLMs) and i'd never have seen the action-form variants that worked better for the operator chapters.

Anti-pattern (when NOT to use) · when the constraint is hard. if the SEO target keyword has to be in the title and in a specific position, you don't need 3 perspectives — you need one agent slotting the keyword into 5 templates. perspectival swarm is for subjective decisions where ear matters more than rules.
→ Chapter 34 — Persona Agents and the Four NEVERs → /weekend-builds for the 3-agent variant pattern applied to layout

Pattern 4: the pre-flight external-dependencies swarm

When to use · before you scaffold any project that depends on external sources or APIs. you've listed 5-10 third-party deps in your plan doc. before you write a line of code, you need to know which of them silently changed policy in 2026.

Wave shape
  1. Wave 1 — One agent per dependency (3-4 agents, parallel)
    Agent 1: audits dep A — current rate limits, deprecation notices, breaking changes since your last known version, 2026 policy shifts. Returns: status (green/yellow/red) + 1-paragraph what-changed. Agents 2-4: same shape against deps B/C/D.
  2. Wave 2 — Continue if >4 deps
    spawn next batch of 3-4. do not exceed 4 per wave — each agent runs WebFetch + reads provider docs + maybe hits a sandbox endpoint. file-system writes are cheap here, but rate-limit collisions on the provider side aren't (two agents WebFetching the same OpenAI docs page within 100ms gets one of them a 429).
  3. Wave 3 — Synthesis (1 agent)
    reads all dep reports, produces GO-NO-GO.md — list of green deps to scaffold against, yellow deps requiring fallback path, red deps to remove from the plan before they cost you a sprint.

remember: 3-4 agents per wave is the empirical sweet spot; 5+ invites filesystem contention AND rate-limit collisions on the upstream APIs you're researching. budget waves accordingly.

Spawn template (paste into Claude Code)
/preflight-external-deps [PROJECT-NAME]

Audit these external deps before scaffolding [PROJECT]:
  1. [dep-1] — current docs: [url]
  2. [dep-2] — current docs: [url]
  3. [dep-3] — current docs: [url]
  4. [dep-4] — current docs: [url]

Spawn 4 parallel research subagents (one per dep). Each agent:
  - WebFetch provider docs + changelog + status page
  - Verify: rate limits as of [today], deprecation notices, 2026 policy
    shifts, auth-flow changes since [last-known-version]
  - Return: status (green/yellow/red) + what changed + replacement
    candidate if red

Save each agent's report to notes/preflight/[dep-name].md. After all 4
return, spawn one synthesis agent to produce notes/preflight/GO-NO-GO.md
with traffic-light table.

If >4 deps: run in sequential waves of 4.

Receipt · ran on a content-aggregator side project 2026-03-28 — 8 deps audited in 2 waves. caught 6 show-stoppers in one session: Twitter API v2 free tier killed for write, Reddit API now $0.24/1K calls (was free), LinkedIn dropping non-partner access in Q3, two RSS providers consolidated, one OG-image service deprecated their v1 endpoint. would have been 2-3 weeks of scaffolding against dead pipes before discovery.

Anti-pattern (when NOT to use) · the project depends only on stable, first-party infrastructure (your own Postgres + your own Vercel + Anthropic API). preflight pays off when the dep is a third-party data feed, a partner API, or anything where the provider's business model could shift under you. don't audit your own database.
→ Chapter 31 — Six Stages from Idea to Deploy → /research-notes for the catalog of deps i've burned on

Pattern 5: the launch swarm

When to use · you're shipping a real launch — book, product, post, campaign — and you need a multi-channel sequence designed by specialists, not by you-at-2am the night before.

Wave shape
  1. Wave 1 — Three launch dimensions (3 agents, parallel)
    Agent 1 (distribution-sequencer, marketing-strategist subagent): designs the day-by-day plan across HN, X, LinkedIn, newsletter, subreddits, cold-side. Outputs notes/launch/distribution.md — 14-day plan with expected impressions per channel, swing variables, mitigation if HN flops. Agent 2 (viral-mechanics-auditor, general-purpose): reads the product itself, surfaces shareable artifacts (a tier list, a CFO defense, a side-by-side comparison) that have post-able life independent of the launch announcement. Outputs notes/launch/shareable-surfaces.md. Agent 3 (hook-and-messaging-auditor, general-purpose): pressure-tests the lead message across three audiences (operators, builders, executives). Identifies which framing converts which audience. Outputs notes/launch/hook-audit.md.
  2. Wave 2 — Synthesis (1 agent)
    reads all 3 reports, writes SYNTHESIS.md — single integrated launch doc with the day-by-day plan annotated by which shareable surface anchors each post and which hook lands per channel.

three is the floor here too. one launch strategist gives you a single plan with no audit on its assumptions; two gives you debate without arbitration; three (distribute / what-to-distribute / how-to-frame) covers the orthogonal dimensions and the synthesis step does the arbitration.

Spawn template (paste into Claude Code)
Spawn 3 launch-design subagents in parallel for [PROJECT]:

Agent 1 (marketing-strategist) — distribution sequence. Design day-by-day
launch plan across HN, X, LinkedIn, newsletter, subreddits, cold-side
amplifiers. Real impression estimates. Swing variables. Mitigation paths.
Save to notes/launch/distribution.md.

Agent 2 (general-purpose) — viral-mechanics audit. Read [PROJECT URL or
PATH]. Identify 3-5 shareable artifacts inside the product that have
independent post-able life. Each artifact: one-line description, target
channel, expected engagement floor. Save to notes/launch/shareable-surfaces.md.

Agent 3 (general-purpose) — hook + messaging audit. Pressure-test the lead
message against [3 audience segments]. For each segment: which framing
converts, which framing reads as funnel-bait, the one-line opener that lands.
Save to notes/launch/hook-audit.md.

After all 3 return, synthesis agent reads all three + writes
notes/launch/SYNTHESIS.md — integrated 14-day plan annotated by anchor
artifact + segment-specific hook per channel.

Receipt · this is exactly the swarm that produced the launch plan for the book you're reading. 3 agents fired 2026-05-10, returned in 4 minutes. synthesis caught that the X thread's launch-announcement framing would have suppressed reach — distribution agent had pencilled in "thread #1: I shipped a book" and the hook-audit agent flagged it as the lowest-engagement opener for the operator audience. final thread led with the LMArena disagreement instead. that single catch was probably 50K impressions of difference.

Anti-pattern (when NOT to use) · the launch is a Tuesday-quiet post to your own list. you don't need 3 strategists for a 400-word email. swarm pays off when the launch has 5+ surfaces and the cost of a wrong hook is > 30 minutes of swarm time.
→ Launch week — the distribution experiment
The meta-rule under all 5

If there's one line to walk away with: dispatch in one message, not five. The swarm only works when the parent session sends all wave-1 agents in a single tool batch — that's what makes them parallel. If you send them one at a time, you've just built a sequential chain in agent costumes and you'll wait 4× as long for the same answer.

Paste a spawn template, swap the brackets, send. If it stalls past 8 minutes, the wave is too big — split into two waves of 3-4 and run them sequentially. 3-4 is the empirical sweet spot. Respect the ceiling.

Daily ops

Standups, vault hygiene, the morning + Friday loops.

7
claude-md-hygieneops

Audit CLAUDE.md files for size bloat and the prepend-cascade accretion pattern (PRIOR ENTRY stacks, multi-date table-cell rows). Reports specific files, lines, and sections to compress. Run biweekly or when CLAUDE.md feels slow.

dreamops

Dreaming for CC — surface candidate memory-learnings from recent session transcripts (digest -> extract agents -> quote re-verification -> dated REVIEW file). PROPOSE-ONLY, never writes memory. Use on /dream or 'what did we learn lately'.

health-pulseops

Check CI, deploy, and cron health across all active projects. Updates ~/.claude/health/heartbeat-state.json. Alerts via Telegram on state transitions.

learnops

Write one lesson to this project's memory dir so future sessions load it — a correction Vlad made, a gotcha that burned time, or a confirmed non-obvious approach. Saves one lesson (new file or an update to the existing one) plus its MEMORY.md pointer; not the whole-session ceremony (that is /retro) and not the multi-session surfacer (that is /dream).

memory-hygieneops

Check MEMORY.md health — detect bloat, find stale memories, suggest consolidation. Run biweekly or after sprint bursts.

retroops

End-of-session retrospective — reviews what happened, extracts lessons, saves insights to memory for future sessions.

substack-pushops

Push a finished Vlad's Newsletter draft from the vault (Vlad-Brain/Newsletter/Drafts) into Substack as an UNPUBLISHED draft through Vlad's connected Chrome, title and body both verified after reload. Use on "push this week's Tuesday/Thursday edition", "push the newsletter draft to Substack", "create the Substack draft". Never publishes. Not for writing th…

Code review + QA

Audits, security passes, bug hunts, debugging.

15
antislopreview

De-slop, build or verify signed-in web app screens, iOS/Flutter/Expo screens, HTML email and generated documents against the repo's brand manifest, its slop checker and a two-theme render receipt. Invoked by design-taste Step 2 or as /antislop; never self-triggers on "design" asks. NOT for charts (dataviz), token extraction (design-system-extractor), func…

auditreview

Parallel multi-dimension codebase audit — finds bugs, security issues, and silent failures across the repo, with adversarial verification of every finding. Use before deploying, after big feature sprints, or when suspicious bugs keep recurring.

audit-contentreview

Verifies truthfulness, accuracy, and link integrity of content before publishing. Catches fabricated statistics, dead URLs, misattributed sources, and company claims that contradict the brand DNA.

audit-website-aeoreview

Audits a live website for AI-engine discoverability (AEO/GEO). Crawls the site, runs 16 deterministic checks plus a 6-dimension content evaluation, and produces a scored report (A-F) with prioritized fixes. Use to get a baseline before improve-aeo-geo, or to measure progress after changes.

credential-helper-lintreview

Lint a script before handing it to the owner to run with a real credential — catches secrets in argv, in usage lines, in output, in xtrace and in world-readable files. Use before handing any script that reads or sets a token, secret, password or API key to the owner, when writing a "credential helper", when asked to "set registry credentials" or to "pipe…

debug-swarmreview

Investigate a bug with multiple competing hypotheses

evalreview

Generate a synthetic test set for an AI feature, score its output two ways (LLM-judge + deterministic code-grade), and emit an honest HTML scorecard with a real pass rate. Use on '/eval', 'score this prompt', 'is this AI feature actually working', or before charging for model output. Wraps eval-harness/ (TS). Not for generic code review (that is /ship-fea…

fixing-accessibilityreview

Audit and fix HTML accessibility issues including ARIA labels, keyboard navigation, focus management, color contrast, and form errors. Use when adding interactive controls, forms, dialogs, or reviewing WCAG compliance.

gstack-carefulreview

Safety guardrails for destructive commands. Warns before rm -rf, DROP TABLE, force-push, git reset --hard, kubectl delete, and similar destructive operations. User can override each warning. Use when touching prod, debugging live systems, or working in a shared environment. Use when asked to "be careful", "safety mode", "prod mode", or "careful mode". (gs…

gstack-qareview

Systematically QA a web app and fix the bugs found: test, fix atomically, re-verify; tiers Quick/Standard/Exhaustive; before/after health scores. Use on 'qa', 'test this site', 'find bugs', 'test and fix', 'quality check'. (gstack)

negative-test-guardsreview

Prove a guard, watchdog, backup, tripwire, health check or CI gate can actually FAIL before trusting it green. Use whenever you write or review code whose job is to detect a problem — guards, asserts, alerts, monitors, backups, verifiers, CI checks. Also fires on "does this check actually work", "is this guard real", "why didn't this alert fire".

seo-opsreview

Structural checklist and checker script for SEO foundation engineering. Use when reviewing or accepting a site/page for SEO/GEO structural compliance (robots, sitemap, canonical, 301 canonicalization, JSON-LD, OG, hreflang, CWV, AI crawler access, rendering strategy), when running the machine checks to produce a report, or when judging whether a frontend…

ship-featurereview

Pre-ship gate — run a deep code review AND a multi-dimension audit on the current change in parallel, merge into one ship-readiness verdict, then fix blockers before commit. Use when Vlad says "is this ready to ship", "pre-ship check", "review + audit this change", or before committing a non-trivial feature.

ultrareviewreview

Deep code review of a PR, file, or module — architecture, correctness, performance, security, and test coverage in one pass

verify-live-surfacereview

Prove a deployed web surface actually serves what you just shipped — by content marker with controls, never by status code. Use after any deploy, or when checking whether a page, route, endpoint, bundle or auth gate is live and correct. Fires on "is it deployed", "did that go live", "check the page works", "verify the endpoint", "is the fix live on prod/t…

Build + ship

Scaffolding, deploys, env flips, framework-specific muscle.

18
bulk-rename-scopedbuild

Pre-flight + scoped bulk rename for identifier-like strings (emails, brands, domains, keys): greps the portfolio, breaks down matches by dir, flags DB-lookup contexts, BLOCKS for scope confirmation before writing. Use on 'replace X with Y across/everywhere'.

cutover-preflightbuild

Acceptance sequence for moving a running service to a new host on the SAME database — differential old-vs-new probe, the exact candidate digest exercised on staging with a real money path, scheduler tick reading, headless client smoke, and the evidence rows that go into the cutover packet. Fires on "cutover", "migration acceptance", "S1", "prove parity",…

deploy-to-boxbuild

Deploy a code/config fix to a git-tracked box reached over SSH (Raspberry Pi, self-hosted server) so it is LIVE, DURABLE across the next git pull, and visible in the user's warm browser. Use when editing files that run on such a box. Not for Vercel/Railway.

design-system-extractorbuild

Reverse-engineer a verified design system (DTCG-flavored tokens, css, playbooks, component inventory, kitchen-sink page) from an existing site/app/repo/screenshots. Use on 'extract/build the design system or tokens from <X>'. Not for from-scratch design or restyling an existing surface (→ frontend-design).

design-tastebuild

Front door to the taste-skill design roster: resolves brand first, then routes to the right taste-skill:* skill. Use for any build-or-change-a-web-surface ask and on '/design-taste'. NOT for schemas/APIs/architecture, charts (dataviz), token extraction (design-system-extractor), or functional QA (gstack-qa).

fixing-metadatabuild

Audit and fix HTML metadata including page titles, meta descriptions, canonical URLs, Open Graph tags, Twitter cards, favicons, JSON-LD structured data, and robots directives. Use when adding SEO metadata, fixing social share previews, reviewing Open Graph tags, setting up canonical URLs, or shipping new pages that need correct meta tags.

framerbuild

If Framer MCP tools (such as `open_project`) are available, use them instead and do not load this skill. Use when the user wants to design, edit, analyze, or publish a website or web page — creating layouts, editing sections, updating text or images, querying site analytics, managing CMS collections and content, syncing external data, creating or modifyin…

framer-code-componentsbuild

Framer code component implementation guidance, platform constraints, layout annotations, property controls, and authoring best practices. Use only after loading the framer skill and reading the Components row of the generated projects/<safeProjectId>/index.md task map first in the same task; never load this skill directly as the entry point.

improve-aeo-geobuild

Audits a website codebase and makes code changes so AI engines (ChatGPT, Claude, Perplexity, Google AI Overviews) can better discover, parse, quote, and cite the site. Covers structured data, content structure, technical signals, and freshness.

pitch-htmlbuild

Generate a single-file interactive HTML pitch artifact from a project's strategic docs. Use on 'beautiful artifact to send to friends', 'make me an HTML pitch', 'interactive slide deck / 1-pager'. Self-contained .html, brand-locked, no build step.

posthog-wizard-followupbuild

Run the PostHog wizard on a Nuxt or Vitest-tested codebase with baseline-test capture before/after — auto-applies the defensive trackEvent() wrapper if the wizard's `useNuxtApp()` injections break unit tests.

safe-git-cleanupbuild

Preflight before any destructive git housekeeping — worktree remove, branch delete, gc/prune, reflog expire, clean -fdx, reset --hard, stash drop. Fires on "clean up the repo", "remove the worktree", "tidy branches", "prune", or before deleting anything git-adjacent.

supabase-state-checkbuild

Report drift between your local Supabase migration files and the actual production DB state via Management API. Surfaces tables that exist but aren't in the tracker, migrations in the repo that aren't applied, and RLS gaps on user-data tables. Use before running `supabase db push` on a project where migrations are sometimes applied via dashboard.

use-railwaybuild

Operate Railway infrastructure: projects, services, databases, buckets, deploys, environments, domains, troubleshooting, docs. Use ONLY when Railway is explicitly named, the project is known to deploy on Railway (e.g. the folderly-dmarc worker), or a railway.toml/railway.app artifact is present. Do NOT fire for Vercel or generic deployment/build-failure/M…

vercel-env-verifybuild

Verify or flip an env var on Vercel and prove it reached the live bundle. Covers NEXT_PUBLIC_* (Next.js), VITE_* (Vite SPA) — both bake at build time, so changing them on Vercel does nothing until a redeploy — and plain production vars (rm + add + redeploy dance of CLI 50.x). Use after changing any Vercel env var, on 'did the env var reach prod', or to ro…

verify-file-durabilitybuild

Detect silent file reversion (bindfs mirrors, overlay mounts, sync agents, watchdog skills) by capturing a snapshot, waiting, then re-checking. Use after writing to any path where an upstream process might re-sync and overwrite the edit — Claude.ai Cowork /mnt/skills/, Codex worktrees, Dropbox/iCloud folders, network mounts.

visibility-flip-checkbuild

Pre-flight before flipping a GitHub repo public->private: checks Pages tier breakage, committed secrets, lost collab signals. Fire whenever Vlad says make/flip the repo private or runs gh repo edit --visibility private.

vlad-framerbuild

Work on Vlad Website Main or vladyslavpodoliako.com using its live Framer skills for design, copy, CMS, SEO, accessibility and release. Use for Vlad's personal website, not unrelated Framer projects.

Strategy + research

Office hours, competitor intel, ultra-plans, monetization.

6
ahrefs-budget-checkstrategy

Preflight check on Ahrefs subscription-info-limits-and-usage before running bulk keyword queries. Prevents 'API units limit reached' errors mid-batch. Estimates cost per operation (keywords-explorer-overview ~50 units minimum, serp-overview variable, site-explorer-organic-keywords 25+). Warns when <100 units remain before a bulk call that needs 500+.

humanizerstrategy

Rewrite AI-sounding text so it reads like the writer without changing what it says. Use when editing or reviewing prose for AI tells: not-X-but-Y contrasts, one-line closers, staged openers, forced triads, dashes everywhere, inflated claims, sales language, stock AI words, bold labels, or filler. Based on Wikipedia's "Signs of AI writing."

newsletter-draftstrategy

Draft a newsletter/longform post in the author's real voice: fetches 2-3 archive issues FIRST as voice calibration before writing a word. Use on 'draft me a newsletter/Substack/longform' or prose 'in my style' for a named publication.

researchstrategy

Launch a research swarm to explore code from multiple angles

swarm-strategic-planstrategy

Multi-wave swarm planning for greenfield ventures or deep planning on existing decisions. Use on /swarm-strategic-plan, 'mega ultrathink', 'full plan on X', 'design swarm for Y'. Modes: full-swarm / lite-mode / solo-deep-think.

ultraplanstrategy

Deep multi-step execution planning with goals, risks, implementation sequence, verification steps, and rollback — ported from claw-code

Portfolio-specific

Belkins/Folderly/LinguaLive — Vlad's actual book of business.

1
folderly-external-auditportfolio

External read-only email-deliverability audit of a domain list -> client-ready report + action register + HTML. Fire on 'Folderly audit', 'deliverability audit', or a domain/mailbox spreadsheet to audit. Not for Folderly app internals or generic Q&A.

Meta

Skills that build other skills, verify other agents, etc.

7
arsenalmeta

Four power questions on finished work — future-proofing, one industry-leading feature, unstated assumptions, what would have made the session smoother. Use on '/arsenal', 'arsenal check', 'power questions', or one of those questions asked after work completes. Report-only; a bare 'retro' means /retro, not this.

preflight-external-depsmeta

Audit external data sources/APIs for current-year viability BEFORE writing code — parallel research agents verify each dependency, catching deprecations, rate-limit and policy changes. Use when starting a project with listed feeds/APIs/integrations.

remove-ai-marksmeta

Strip AI provenance marks: invisible Unicode, statistical text watermarks (rewrite), and C2PA/EXIF/XMP metadata on PNG/JPEG/WebP/SVG/PDF/DOCX/HTML/MD. Use on 'remove watermarks / C2PA / AI metadata / invisible Unicode' or /remove-ai-marks (alias /remove-claude-marks).

teleportmeta

Jump to a file or symbol by name — searches filenames and content, returns up to 10 file matches and content matches with line numbers

verify-before-askingmeta

Run before asking Vlad for ANY credential, key, ID, URL, account detail, or fact a past session likely established. Exhausts the derivation paths first. Fires on the impulse "I need X from you", "please provide", "I don't have access to", "can you paste", or before writing a hand-off that asks for a value.

wfmeta

Front door to the compiled Workflow library. /wf lists workflows with token estimates; /wf <name> [scope] resolves the scriptPath, assembles args as real JSON, and launches (lite tier default). Use on '/wf' or 'run the <X> workflow'.

workflow-forgemeta

Author, lint, list, and audit the compiled Workflow library at ~/.claude/workflows/. Use when Vlad says "add a workflow", "new workflow for X", "lint the workflows", "what workflows do I have", "compile this skill into a workflow", or when a workflow script misbehaves. Encodes the verified DSL gotchas so new scripts don't repeat them.
How to read this list

Most operators discover skills by need. They explain a workflow to Claude three times, write a SKILL.md, drop it in ~/.claude/skills/, and never explain it again. That's how the list above grew. Every line in it started as a paragraph Vlad got tired of retyping.

The custom agents in ~/.claude/agents/ serve a different job — they're specialized subagents you spawn from a main session. code-reviewer reviews code. tdd-guide enforces test-first. pc-api-lead owns a B2B SaaS backend. Pick read-only tools for the agents that should never edit; broader tool surfaces for the ones that ship.

Plugins are bundles — skills + agents + slash commands + MCP servers, distributed as one. Browse the marketplace inside Claude Code with /plugins.

Stay close

The next edition lands when this list says it does.

No course. No paywall. Operator playbooks weekly. 10K+ subscribers.