Source: raw/How_Anthropic_Engineers_ACTUALLY_Prompt_Claude_Code.md — YouTube tutorial qOvc9IUKEIc. Creator unidentified in transcript (refers to their own “channel” + “anti-slop agreement” + a prior video on Boris Cherny’s Claude usage as a follow-up). Quotes Anthropic engineers Eric and Barry directly from the AI Engineering Code Summit + cites the Anthropic engineering blog.

Four rules distilled from Anthropic-internal prompting practice, sourced from named Anthropic engineers’ AI Code Summit talks. Thesis: most users still prompt Claude Code like they prompt chat — sentence-shaped intent, one-shot. Anthropic’s own engineers prompt skills, treat skills as 3-layer composites (description + instructions + tools), keep skills small and composable, and update the skill every session to compound. Reframes the prompt-engineering mental model around skills as the unit of leverage, not prompts.

Key Takeaways

Rule 1 — Prompt skills, not Claude

  • Stop writing one-off prompts for repetitive tasks.
  • Anthropic’s framing (Eric): “Skills are organized collections of files that package composable procedural knowledge for agents. In other words, they’re folders.” Barry’s description: “procedural knowledge for agents.”
  • The mental shift: instead of typing a long prompt “respond to this email in my voice / tone / style”, you write a /draft-email skill once, then call it with /draft-email <email>. Same task. Skill gets sharper each time it’s used.
  • Three-layer architecture Anthropic adapted from their engineering presentation:
LayerWhat it isWho uses it
1AI model (Claude Opus 4.7, etc.)Anthropic builds it
2AI agents + raw promptsWhere most users live today
3Skills (the application layer)Where Anthropic engineers spend their time

“Anthropic is building the phone itself. You have to create the apps. That’s the layer to control.”

Rule 2 — Skills are more than prompts (the 3-layer skill model)

A skill is not just a prompt that lives in a folder. Inside every skill, three load-bearing layers:

LayerPurposeRisk if neglected
DescriptionWhat Claude reads to decide whether to invoke the skill. Acts as the “label on the folder.”Vague description → Claude can’t recognize when to use it. Specific description → auto-invokes correctly without explicit /slash call.
InstructionsThe step-by-step playbook Claude follows once it grabs the skill.This is where most users stop.
ToolsScripts, API calls, reference files the skill has access to.Where most leverage lives — and where most users skip.

Eric (Anthropic) on the symptom of users stopping at the instructions layer:

“The funniest things I see is that people will put a lot of effort into creating these really beautiful, detailed prompts. And then the tools that they make to give the model are sort of these incredibly bare-bones — no documentation, like the parameters are named A and B, and it’s kind of like, oh, an engineer wouldn’t be able to work with this as a function.”

Anthropic engineers do the opposite — they focus on the tools layer. Example from the video: the creator built a check-domain skill where instead of having Claude reason about domain availability in tokens, the skill calls a programmatic domain-check tool. Result: 10 subagents can iterate through 10,000+ domains in parallel to find the right one — a task that wasn’t possible with prompt-only skills.

Rule 3 — Build composable skills, not custom skills

Pulled directly from Anthropic’s engineering blog: skills should be composable, portable, efficient, and powerful.

  • Composable = multiple skills can work together with Claude automatically coordinating which to use.
  • Anti-pattern: one massive /content-creation skill that generates ideas, writes scripts, drafts social posts, all in one file. Unmanageable.
  • Pattern: split into focused units — /youtube-idea-research, /youtube-script-writer, /linkedin-post. Each can call the others. Claude does the chaining.

Three concrete benefits of composability:

  1. Issues are easy to spot. When a focused skill breaks, you know exactly where to look. With a mega-skill, you don’t know which part broke.
  2. Improvements compound. Update /youtube-idea-research once → every workflow that calls it gets upgraded. Mega-skill = overlapping functionality + you fix in one place, still broken in another.
  3. Reuse instead of rebuild. A /check-domain skill plugs into any workflow. Not rebuilding the wheel each time.

Technical patterns from Anthropic engineers:

  • Save scripts inside skills (tools-layer pattern). Barry at AI Engineering Code Summit: “We kept seeing Claude write the same Python script over and over again to apply styling to slides. So we just asked Claude to save it inside of the skill as a tool for his future self. Now we can just run the script — that makes everything a lot more consistent.” Token math: code is deterministic (same input → same output → reusable), AI is non-deterministic (interprets, guesses, costs tokens per run). General rule: if you can use code instead of AI, you should.
  • Two invocation control flags most users don’t know about:
FlagWhat it doesUse case
user_invocable: falseHides the skill from the /slash menu. Only agents can invoke.AI-internal sub-skills you don’t want polluting your menu.
disable_model_invocation: trueOnly the user can run it. The model can’t invoke it autonomously.High-risk skills — deploy to production, send a message, drop a database.

Rule 4 — Update your skill every session (the compounding loop)

Where Anthropic engineers pull ahead is treating skills as living artifacts.

  • A one-off prompt evaporates the moment you close the chat. A skill persists. Every use is a chance to sharpen it.
  • Anthropic engineering team quote: “Our goal is that Claude on day 30 of working with you is going to be a lot better than Claude on day one. Every time Claude learns something about how you work — your voice, your process, your edge cases — you write it down in the skill. Next session starts smarter than the last.”

The one question Anthropic engineers ask after every skill run:

“Is this a one-time fix or should this be in the skill forever? If forever, update the skill.”

Add the rule, the example, or the edge case. Most users skip this entirely — run skill → get output → continue. Anthropic engineers run skill → get output → update the skill → compound.

One-line meta-prompt for the update loop (paraphrased from the transcript):

“Review the back and forth I just had after using this skill. Can we enhance the skill so this is handled automatically or we don’t make the same mistake again?”

The four rules zoomed out

  1. Use skills, not prompts.
  2. Build tools, not just skills with prompts. (The 3-layer model — description + instructions + tools.)
  3. Build skills that are composable, not custom. (Small focused vs giant mega.)
  4. Update your skills every time you use them. (Compounding loop.)

The transcript closes with: “Using Claude like an engineer doesn’t have to be complicated.” Followup video promised on Boris Cherny’s specific skill usage.

Open Questions

  • Creator identity — not named in transcript. Likely a separate creator from the Agentic Academy (skill-systems author) given different framing + reference to “my channel” + “anti-slop agreement” branding.
  • AI Engineering Code Summit primary source — the talks by Eric and Barry are the canonical sources for rules 1-3. Recording / publication URL not in transcript. Worth tracking down.
  • Anthropic engineering blog post — referenced as the source for “composable, portable, efficient, powerful” framing. Specific URL not given.
  • Boris Cherny skill-usage video — promised as a follow-up by this same creator.
  • user_invocable / disable_model_invocation flag documentation — referenced as if shipped in the official skill frontmatter spec. Verify in Anthropic skill docs.

Try It

  1. Audit one repetitive prompt. Pick a task you’ve prompted Claude 3+ times this month. Convert it into a skill (/your-skill-name) with a clear description, instructions, and at least one tool reference. Use skill-creator to scaffold it.
  2. Add the tools layer. Wherever the skill currently asks Claude to “do X” via prompt, ask whether X could be a script. If yes — write the script once, save it inside the skill folder, reference it from the instructions.
  3. Audit one mega-skill. Find your largest SKILL.md (probably one you wrote that does end-to-end content/research/posting). Decompose into 2-3 focused units. Have a tiny orchestrator skill chain them. Same logic — better leverage.
  4. Set user_invocable: false on any skill that’s only meant for the agent to call autonomously (e.g., internal sub-skills). Cleans up the /slash menu.
  5. Set disable_model_invocation: true on any skill that touches production, sends messages, or runs destructive ops. Forces a human confirmation step.
  6. Add the compounding loop. End every skill-run session by asking the meta-prompt: “Review what we just did. Can we enhance the skill so this is handled automatically next time?” Update if yes. Skip if it was truly one-off.
  7. Track day-30 improvement. Pick one skill. Note its SKILL.md size + behavior on day 1. After 30 days of updates, compare. If it hasn’t gotten more specific to your voice/workflow, the compounding loop isn’t running.