How Anthropic Uses Claude Skills Internally

YouTube Summary · 7 Jul 2026

VIDEO

⚡ TL;DR

EXECUTIVE SUMMARY

Anthropic's internal playbook for Claude Skills distills into five lessons: categorize skills into four types (utility, verification, data enrichment, orchestration), power them with scripts/templates/setup-prompts, treat verification as the highest-leverage category, maintain a running gotchas list as your personal moat, and write trigger descriptions that describe WHEN to fire — not what the skill is. The meta-lesson: skills are living folders that compound over time, not static markdown files.

Source: YouTube — How Anthropic Uses Claude Skills · Duration: 16:16 · Channel: The creator does not self-identify in the transcript

🎯 Key Highlights

4 Skill Types — Anthropic's 9 technical categories collapse into 4 universal types: utility (small reusable tasks), verification (check final output), data enrichment (pull external data), orchestration (chain skills together). Best skills fit cleanly into ONE type.
Skills are folders, not files — The biggest misconception. A skill is a folder containing scripts, assets, templates, data — not just a markdown file. Claude can discover and manipulate everything inside.
Scripts partition deterministic vs non-deterministic — Push repeatable logic into scripts (deterministic); let AI handle only the non-deterministic parts. Fewer tokens burned, more repeatable output.
Verification = highest leverage — Anthropic says verification skills had the most measurable impact on output quality internally. Two types: verify for correctness (facts/numbers) and verify for quality (meets the bar). Worth an engineer spending a week on.
Gotchas are your moat — A running list of things Claude actually got wrong. Only add gotchas you've personally seen — never speculative ones. Skills start small and compound: structure = skill, leverage = verifier, moat = gotchas.
Description = trigger condition, not summary — Claude scans skill descriptions to decide when to fire. A good description names who it serves and when it should fire, using the actual words a user would type.
Orchestration skills call other skills by name — Complex workflows should be broken into sub-utility skills first, then chained. Updating a utility skill auto-updates all orchestration skills that call it.
Manager-as-reviewer pattern — Anthropic's Amol Agrawal built a skill that simulates his manager's feedback using her public writing + Slack data. Every week it tells him what she would have said — pre-vetting his work before she sees it.

⏱ Topics by Timeline

00:00 Intro — Anthropic released their internal playbook on how they use Claude Skills across engineering, marketing, and legal teams

00:22 Lesson 1: Understand the Categories — 9 technical categories bucketed into 4 skill types (utility, verification, data enrichment, orchestration)

02:23 The one trap to avoid — skills that straddle multiple categories confuse the agent; best skills fit cleanly into one

03:00 Lesson 2: Use the Power Components — skills are folders, not just markdown files

03:32 Power component 1: Scripts — partition deterministic (scripts) vs non-deterministic (AI) workflows to reduce token burn

04:28 Power component 2: Assets/Templates — upload past reports as templates so Claude fills in instead of improvising

05:00 Power component 3: Setup Prompts — config.json for first-run values, ask-user-question tool for structured input, arguments field for input reminders

06:55 Lesson 3: Focus on the Verifier — Anthropic says verification has the most measurable impact on output quality

07:20 Two verification types: correctness (did Claude get facts right?) vs quality (does output meet the bar?)

08:00 Built-in verifiers: /verify (run app, confirm change works) and /run (launch app for Claude to see its own work)

08:33 Skill-driven verification — tweak existing skills to have pass/fail or grade output instead of building verifiers from scratch

09:16 Manager-as-reviewer case study — Amol Agrawal's skill simulates his manager's weekly feedback using her public writing + internal data

10:31 Channel promotion — 50K subscribers, Claude Max giveaway, anti-slop subscribe agreement

11:22 Lesson 4: Write the Gotchas — running list of issues Claude hits; only include things you've actually seen go wrong

12:36 The framework: skill = structure, verifier = leverage, gotchas = personal moat. Skills compound over months — don't aim for perfection on day one

13:30 Lesson 5: Tune the Trigger — description field describes WHEN to fire, not what the skill is. Name who it serves and when it should fire

14:16 Anthropic's front-end design skill as example — 'Use this skill when the user asked to build web components, pages, or applications'

15:00 Orchestration skills — reference other skills by name; break complex workflows into reusable sub-utility skills first

15:46 Recap of all 5 lessons and outro

🧠 Hermes Integration

💡 Audit & Categorize Our Skills

We have 80+ skills across apple, creative, devops, github, mlops, productivity, research, etc. Run a categorization audit mapping each skill to one of the 4 types (utility, verification, data enrichment, orchestration). Flag skills that straddle multiple categories — they're confusing the agent. Split them into focused sub-skills.

✅ Actionable: Run a skills_list + categorization pass this week

💡 Add Verification Components to Existing Skills

Anthropic says verification is the highest-leverage skill type. Our skills like youtube-summary-zeus, html-dark-template, and github-code-review already have implicit verification steps (verify file with ls -la, curl check, link verification). Make these EXPLICIT — add pass/fail or grade outputs. E.g., after rendering HTML, the skill should output 'PASS: 15 links verified' or 'FAIL: 3 broken links found'.

✅ Actionable: Patch 3-5 high-usage skills with explicit verification outputs

💡 Gotchas Section as Standard Practice

Many of our skills already have 'Pitfalls' sections (youtube-summary-zeus has 6, html-dark-template has 19). Rename/standardize these as 'Gotchas' and enforce the rule: only include failures you've actually encountered. The pitfall lists are already valuable — they're our moat. Going forward, every new skill should start with 1 gotcha minimum and grow from real usage.

✅ Actionable: Already partially done — standardize naming and enforce on new skills

💡 Trigger Description Optimization

Audit our skill descriptions. Many are written as summaries ('YouTube transcripts to summaries, threads, blogs') rather than trigger conditions ('Use when the user shares a YouTube URL or asks to summarize a video'). Rewrite key skill descriptions to name WHO they serve and WHEN they should fire, using the actual words M~ would type.

✅ Actionable: Audit and rewrite 10 highest-traffic skill descriptions

💡 Manager-as-Reviewer Pattern for M~

The Amol Agrawal pattern is powerful — encode a trusted advisor's judgment as a verification skill. For M~, this could be a 'Devil's Advocate Reviewer' skill that simulates critical feedback from a specific advisor or investor before shipping deliverables. We already have the opc-war-room and opportunity-scorecard skills — extend this pattern to a personal advisor simulation using GBrain-stored knowledge of their preferences and past feedback.

⚖️ Explore: Prototype as a GBrain-anchored verification skill

💡 Scripts for Deterministic Workflows

Several of our skills already embed scripts (fetch_transcript.py, html-template-render.py, generate-exam-html.py). The lesson: push MORE deterministic logic into scripts. For skills that currently use inline LLM reasoning for repeatable tasks (e.g., JSON parsing, file naming, slug generation), extract those into scripts. This reduces token burn and makes output repeatable.

✅ Actionable: Review skills with high token usage for script extraction opportunities

😈 Critical Thinking & Devil's Advocate

Anthropic's Playbook ≠ Universal Truth

This video presents Anthropic's internal practices as universal best practices. But Anthropic is a well-funded AI lab with engineers who can spend a week on a single verification skill. Most solo builders and small teams don't have that luxury. The 'worth spending a week on a verifier' claim is aspirational, not practical for most users.

Categories May Over-Constraint

The 4-type categorization (utility, verification, data enrichment, orchestration) is clean but artificial. Real-world skills often blur lines — a data enrichment skill that also verifies the enriched data is arguably both. Forcing every skill into one bucket might cause over-splitting, creating micro-skills that are too small to be useful on their own.

Gotchas as Moat is Overstated

The claim that gotchas are your 'personal moat' is a strong statement. A gotcha list is valuable, but it's not a competitive moat — it's just good documentation. Anyone can read your skill, see your gotchas, and replicate the behavior. The real moat is the underlying data, integrations, and workflows — not a list of things that went wrong.

Manager-Clone is Risky

The 'simulate your manager's feedback' pattern is clever but risky. You're training a model on someone's public writing and Slack messages to approximate their judgment. This can produce confidently wrong feedback that sounds like your manager but isn't. It may also create an echo chamber where you optimize for the AI-clone's approval instead of the real person's. The real manager's feedback includes context the AI doesn't have — body language, strategic pivots, private knowledge.

Trigger Description ≠ Full Solution

The video frames the description field as the key to automatic skill triggering. But in practice, Claude's skill selection depends on the entire context window — the user's message, conversation history, and competing skill descriptions. A perfect description doesn't guarantee correct triggering, especially when multiple skills could apply. The video oversimplifies a complex matching problem.

Self-Promotion Conflicts

The video creator runs buildpartner.ai (a SaaS product for AI-simulated expert advice) and positions it as a natural extension of the manager-as-reviewer pattern. This is a content marketing play disguised as a lesson. The 'internal focus group' and 'expert advice' patterns are framed to funnel viewers toward his product. Take the Anthropic-sourced insights; apply more skepticism to the creator's own recommendations.

What's Missing

The video doesn't address: (1) skill versioning and migration — what happens when a skill changes and breaks dependent workflows? (2) skill discovery at scale — with 50+ skills, does Claude's triggering degrade? (3) cost analysis — how many tokens do scripts vs LLM reasoning actually save? (4) failure modes of orchestration skills — chaining skills can create cascading failures. (5) security implications of scripts running arbitrary code from skill folders.