Source: raw/How_to_build_Claude_Skills_Better_than_99%_of_People.md, raw/How_to_Use_Claude_Skills_2.0_Better_than_99%_of_People.md, raw/How_to_Actually_Build_Claude_Skills_like_a_Pro.md, raw/Anthropic_Just_Revealed_8_New_Rules_for_Building_Great_Claude_Skills.md, raw/Stop_Writing_Claude_Skills._Reverse_Engineer_Them.md; Anthropic claims checked against ai-research/anthropic-skill-authoring-best-practices-2026-10-27.md and ai-research/anthropic-best-practices-claude-code-2026-05-17.md

Creator: Ben AI · Platform: YouTube · five videos:

Ben runs an AI agency and a paid “AI accelerator” community, and across eight months of videos he has refined one method for building skills that automate real business work (his LinkedIn writer, newsletter, YouTube title ideation and email-triage skills are the running examples). This article puts the five videos together in the order you would use them: decide what to build, pick a build method, structure the files, test, then improve through use. Most of his structural rules trace to Anthropic’s own docs; the per-rule trace is in Skill Authoring Best Practices § Second relay. The parts that are his own are the build-method ranking, the reverse-engineering prompt, the “multiple options at every checkpoint” rule, and his eval prompts for subjective knowledge work.

Key Takeaways

  • Build “work skills”, not just download “general skills.” General skills (humanizer, fact-checker, find-skills, skill builders) help with any task and are worth downloading. Work skills encode your own repeat workflows and have to be built by you. Work skills come in two kinds: autonomous (email triage, YouTube research, run by a scheduled task at 8 a.m.) and assistant (LinkedIn writer, where you steer at each step).
  • Build a skill when a task repeats, when you keep giving the same correction, when you keep pasting the same context, or when you want to hand a process to someone else.
  • Scope to the smallest valuable part. Do not put research, post-writing, newsletter and image generation in one skill. Ben split his YouTube workflow into ideation, title, packaging and outline skills that he runs one after another in the same chat. His reason is context rot: the further a chat goes, the less Claude follows the instructions it read at the start. Anthropic’s guide has the same advice (“Iterate on a single task before expanding”).
  • Do the task with Claude first, then convert it. Of his build methods, one-shot prompting (“I’ve yet to see one good skill coming out of that method”) ranks last. For work you already do, he works through the task in a chat, corrects Claude as it goes, then says “turn this into a skill.” He explains it with Polanyi’s paradox (“we know more than we can tell”): the edge cases come out when you do the task, not when you describe it. Anthropic’s best-practices page lists the same first step: “Complete a task without a Skill.”
  • Or start from finished outputs. For tasks with a clear deliverable, give Claude your best and worst examples, have it name the patterns that separate them, then have it build the skill from those patterns. Ben rebuilt his newsletter skill from his 6 best and 6 worst issues this way and says the result beat the version he wrote step by step.
  • SKILL.md holds only the trigger, the steps and the routing to reference files. Examples, voice guides, ICP notes and prompts go in reference files, and each step names the files to read right before it runs.
  • Route connector calls and research through subagents. MCP output and web research fill the context window before the step that matters (interpreting the data). His title skill runs its VidIQ research in eight subagents and keeps only their summaries.
  • Give 3 to 10 options at every human-in-the-loop step. Picking from five angles, outlines or hooks is faster than explaining what is wrong with one. Ben says this is not in Anthropic’s guide but copies how Anthropic builds its own skills.
  • Turn off auto-triggering for skills you always call by name. disable-model-invocation: true in the frontmatter.
  • Add a self-improvement rule. On a correction or an output you like, the skill asks whether to make it permanent, then updates itself or saves the output as an example.
  • Test function first, taste second. Functional checks (right order, right files loaded, connector called correctly) are measurable. Copy quality is not, so it improves through use and saved examples. Optimize one variable per eval round.

Pick a build method

MethodWhen Ben uses itHis verdict
One-shot prompt (“build me a skill that…”)Never by choiceProduces skills that need endless fixing
Plan first (plan mode or his process interviewer skill, 10 to 15+ questions, 30 to 45 minutes)New processes, fuzzy workflows you have never written downBetter than one-shot; outputs a skill brief (PRD) to build from
Do the task once with Claude, then “turn this into a skill”Any knowledge work you already doHis default
Record a skill (Cowork, narrated screen recording)Click-heavy tasks with judgment calls, e.g. inbox triageSame principle with less typing; see [[claude-ai/record-a-skill-cowork
Reverse-engineer from outputsTasks with a clear finished artifact (newsletter, proposal, reel)Faster and better than describing the process (Oct 2026)
Adapt someone else’s skillMarketplace or team skillsStill needs adapting to your connectors and workflow

The process interviewer is Ben’s adaptation of Matt Pocock’s grill-me skill (Matt Pocock’s skills), extended with Anthropic’s skill-building guidance and set to answer questions from your second brain before asking you. The general pattern is in Context-Extraction Skills.

Reverse-engineering a skill from examples (three steps)

  1. Gather examples. Best and worst performers if you have data (Ben: 6 and 6 newsletters ranked by open and click rate). Without data, use a creator in your niche or anything you like (he gave three reels in a style he wanted).

  2. Have Claude find the winning patterns. His prompt, close to verbatim:

    Compare them. Find the structure, hook, CTAs, tone, length and subject-line
    patterns, and say why each one likely worked or didn't. Tag every finding as
    seen, inferred, or unknown.

    The seen / inferred / unknown tag is the important line: it stops Claude inventing reasons and shows you which findings are guesses. In his data, open rates were similar between best and worst issues while click-through differed, so body and CTA mattered more than subject line.

  3. Review the patterns, then “now build the skill,” then test it on a real input. His rebuilt newsletter skill made issues shorter, moved the CTA earlier and gated some content behind the link.

This does not fit tasks without a clear output (email triage); use do-it-once or record-a-skill there. Anthropic’s best-practices page supports the principle: “Examples convey the desired style and level of detail to Claude more clearly than descriptions alone.” The “reverse engineer anything” repo that prompted the video is covered in Reverse-Engineering Software With Coding Agents; Ben’s “200,000 stars in one week” figure is unverified.

The skill prompt (February 2026 version)

When he does write a skill from a prompt, Ben fills these sections. He says it is “not hard science”:

  1. Name and trigger: “should be triggered any time a user mentions they want to generate an infographic.” This becomes the description.
  2. Goal: one or two lines.
  3. Connectors, APIs, MCPs, including the exact page or table to use (his newsletter skill names the Apify YouTube-transcript MCP).
  4. Reference files. His reusable copywriting set: what we do, ICP, voice personality, a channel strategy doc (newsletter, YouTube, LinkedIn), a writing framework, and output examples. He says this set came from testing in the PromptMetheus IDE and that examples affect quality most. Build missing ones with Claude by having it interview you (his YouTube strategy doc came from plan-mode questions).
  5. Process, step by step. For each step: what to do, where a human checks in and in what form (Claude’s question boxes offer checkboxes, single select or open field), which reference file to read, and what the step outputs.
  6. Rules: predict what could go wrong. Restate that reference files are mandatory at their steps (he finds Claude skips them otherwise) and restate the multiple-options rule.
  7. Progressive updates: “every time I define a clear thing not to do any more, update the rules section,” and save an approved final output as a good example.

Structural rules, with Ben’s prompts

Ben gives a ready prompt for each rule (he distributes the full set as a free download). Use them on a new skill or to retrofit an old one:

RulePrompt to Claude
Lean SKILL.md”Keep the SKILL.md steps short and direct; write each step as a plain instruction.” Move anything else to reference files.
Per-step progressive disclosure”Use the skill creator skill to update this skill so it uses progressive disclosure for reference files. Map each reference file to the specific step where it’s actually needed.”
Subagents for connectors and research”Use the skill creator skill to update this skill so it keeps connectors and research out of the main context; have the skill route the connector call or research to a subagent.”
Human-triggered only”Make this a human-triggered skill only. I don’t want Claude to auto-invoke it. Set disable-model-invocation to true in the frontmatter.”
Self-improvement”Add a self-improvement rule: any time I correct your output during the process or get an output I explicitly like, ask me whether that should become a permanent update to the skill.”
Options at checkpoints”Use the skill creator skill to update this skill so every human-in-the-loop step gives me multiple options to choose from.”

Notes from the videos:

  • Ask whether you need the live call at all. If the data does not need to be current, store it as a reference file instead of fetching it with a connector each run.
  • Per-step loading paid off in his LinkedIn skill. Reading the hook-template file right before the hook step “really improved the outputs” compared with loading it at the start. His run reads ICP and what we do first, three more files at step two, one more later.
  • About 90% of skills online skip even the basics (lean SKILL.md, references split out), by Ben’s estimate. His old win-loss analysis skill is his own example of a bloated one.
  • Anthropic’s own skill-creator does not apply per-step disclosure reliably, in his experience, which is why he built his own skill builder (sold in his community). It applies all six rules in one pass: in his demo it added subagents, a self-improvement rule, disabled auto-triggering and split SKILL.md from references.

Test

Ben calls the March 2026 skill-creator update “Skills 2.0” (Anthropic did not use that name; the wiki records it as the skill-creator evals overhaul in Building Skills guide). After building, Claude offers to run test cases, spins up subagents in parallel and produces a report with each run’s steps, outputs, pass/fail grades and a feedback box per variation.

  • The default test only proves the skill runs. With no criteria given, Claude invents its own (his: a PS with a soft pitch, a sign-off, a word-count range, a .docx produced).
  • To optimize, state one goal, the criteria, and the test design. His example: optimize for matching Ben’s voice; criteria = closeness to his reference examples, no em dashes, length, personal stories from his background file; design = one YouTube video, five variations. Result: style match failed 2 of 5, personal stories 1 of 5, which told him what to fix. “Optimize one thing at a time.”
  • Functional eval for subjective skills (July 2026). “Use the skill creator skill to run an eval using three subagents. Test whether the skill functions the way it’s supposed to according to the goals and steps laid out in the SKILL.md. Does it execute the process in the right order? Does it load the reference files when it should? Does it call the connector correctly?” Add one real test case, or Claude will make one up. His LinkedIn writer passed 100% with minor suggestions.
  • A/B tests only once a skill works. Speed: Claude made a leaner variant; Ben reports 93,000 tokens and 204 seconds for the original versus 77,000 tokens and 160 seconds for the lean one, which also failed transcript extraction and came out too short. Claude blamed the transcript tool, a rerun passed, and he kept the lean version. Context: run the skill with all eight reference files versus without the voice-personality file to see whether that file helps. Anthropic suggests rerunning skill-versus-no-skill comparisons when a new model ships.
  • Hard, measurable criteria let you run an auto-improvement loop such as Karpathy’s AutoResearch; Ben leaves that out because most knowledge work lacks such criteria.

Improve through use

  • Expect 2 to 10 iterations. Ben compares a skill to onboarding an intern; his best skills were revised 5 to 10 times and his infographic skill about five.
  • When it skips or breaks a step, ask Claude why first. The cause is often an unclear line in the skill or a reference file.
  • Put each fix in the right place: process problems go in SKILL.md; extra information becomes a reference file; “never do this” becomes a rule (his: no black infographic backgrounds because shadows disappear); if Claude struggles with a tool, guide it through the action once by hand, then have it write an MCP-instructions reference file.
  • Separate builder and tester chats. Refine in one chat, test in a fresh one, report mistakes back to the builder. This is Anthropic’s Claude A / Claude B method.
  • Save good outputs as examples. For copy skills Ben wants 3 to 5 examples minimum and has seen up to 20 help.
  • Share a finished skill as a zip (Settings → Capabilities → upload) or through GitHub; bundle several into a plugin, and a company’s plugins into a marketplace (Plugins and Marketplaces).

Try It

  1. List your repeat tasks and mark each “clear finished output” or “judgment process.” The first kind suits reverse-engineering; the second suits do-it-once or record-a-skill.
  2. Pick one and cut it to the smallest step that would still save you time on its own.
  3. Build it with the method from the table. For reverse-engineering, paste the seen / inferred / unknown prompt with your best and worst examples.
  4. Apply the six structural prompts above, one at a time, and read the diff after each.
  5. Run the functional eval prompt with three subagents and one real test case before relying on the skill.
  6. Use it for two weeks with the self-improvement rule on. Each time it asks “make this permanent?”, answer deliberately.
  7. When a new model ships, A/B the skill against a lighter version of itself and against no skill.

Implementation

Tool/Service: Claude Cowork or Claude Code skills; Anthropic’s built-in skill-creator for evals and A/B tests.

Setup: Skills live as a folder with SKILL.md plus reference files (and optional scripts, e.g. his infographic skill calls the Nano Banana API from a script, run in the Code tab). Ben’s process interviewer, skill builder and prompt pack come from links in his video descriptions; the skill builder sits in his paid community.

Cost: Not given beyond token counts. Evals and A/B tests run several full skill executions in parallel, so they cost several times one run.

Integration notes: A human-in-the-loop step works best with Claude’s question boxes (checkbox for “generate all variants”, single select for an angle). An approve-and-save box at the end of a run builds the example library automatically.

Open Questions

  • Does disable-model-invocation: true also keep the description out of every chat’s context, as Ben says? The wiki’s copy of Anthropic’s Claude Code best practices only documents the field for “workflows with side effects that should be triggered manually.” Check the current Claude Code skills docs.
  • “Turn into a skill” button in every chat. Ben describes one in the October 2026 UI; not confirmed in first-party material in the wiki.
  • Boris Cherny tweet on background subagents by default (July 2026 video): not located; confirm against the Claude Code changelog.
  • The token and time figures (93K/204s vs 77K/160s) and the “90% of skills” estimate are single demonstrations and opinions, not benchmarks.