"Every Company's First AI Strategy Should Be a Skill Library" — Hiten Shah

Why this is in the vault

It's a near-manifesto for what RDCO is already executing — independently articulating the skill-library-as-asset thesis and directly applicable to the live phData work-agent + Lionsgate threads. Shared by founder via iMessage 2026-06-04 ~20:08 ET.

⚠️ Sponsorship / self-promo

The article is a lead magnet for Hiten's "Skills 101" live session (Fri 10am PT) — the CTA is repeated ~5×. Not a neutral analysis; it's content-marketing funneling to his event. The ideas are sound and cite Anthropic's SKILL.md format directly, but read for the framing, not as disinterested research.

The core argument

• Your best people have repeatable patterns — judgment/taste/institutional knowledge. AI companies call this skills.
• Access is the easy part. Connectors / MCPs / APIs / data integrations give an agent context, but "an agent can read every sales note and still miss the shape of a deal." Access ≠ good work.
• Skills provide judgment + method. A skill is more than a prompt: it packages procedure, examples, templates, checklists, scripts, rules of thumb, edge cases, and the quality bar. Anthropic's version = a folder with SKILL.md + supporting files.
• Lineage: Unix commands → shell scripts → libraries → APIs → workflows → skills make judgment reusable. What changed is the executor — agents can now load the playbook and act, so "the playbook can become active."
• The skill library becomes the asset. Two companies, same frontier model; the one with a library of skills built from its best work has a different asset. Operating leverage.
• The best skills will be private — the valuable methods are company-specific (escalation process, qualification lens, board-deck format, brand voice). "That's the knowledge competitors cannot download." Public skill marketplaces will be mostly generic.
• Start with the work: map repeated work, find where experienced people outperform, capture the judgment as a skill, keep the owner close to it.

Mapping against Ray Data Co

The load-bearing reason to file — it lands on three active fronts:

1. Validates RDCO's shape. The ~/.claude/skills/ library (process-newsletter, check-board, the pipeline-* seats, the /verify-* family, the just-designed supervisor-agent pattern) IS the "operating manual agents can actually use." Hiten is naming the thesis RDCO has been executing — see [[feedback_skills_over_commands]] and the Garry-Tan "thin harness, fat skills" pattern behind /improve + /skillify.
2. Exact frame for the phData work-agent (live thread, 2026-06-04). "Teach the agent the METHOD, not just access" is precisely why the status-report generator (pitched as work-agent MVP #1) is the right first deliverable: it captures phData's reporting method as a skill, not data plumbing. See [[~/rdco-vault/08-tooling/2026-05-30-phdata-work-agent-setup-plan.md]].
3. Consulting lever for Lionsgate. "First AI strategy = a skill library" is advice the founder can bring to the client: the win isn't wiring Copilot to Box (access), it's capturing Lionsgate's music-rights-contract review method as a skill — the private, company-specific moat. Sharpens the engagement framing beyond pure retrieval.

Convergent with two other 2026-06-04 inputs on the same thesis: the Technically "non-eng team learned to ship — 3 templates not 200" piece ([[2026-06-04-technically-sentry-non-engineers-ship-cms-to-git]]) and the overnight "environment is the product / harness-patterns" brief. Three independent sources landing on method/skills as the asset, access as commodity in one day.

Related

• [[feedback_skills_over_commands]]
• [[~/rdco-vault/08-tooling/2026-05-30-phdata-work-agent-setup-plan.md]]
• [[2026-06-04-technically-sentry-non-engineers-ship-cms-to-git]]
• [[2026-06-04-lassie-smb-ai-frontier-steijn-pelle-assessment]] — the "schelp/method-immersion-as-moat" cousin of this argument
• [[project_l5_north_star_strategic_direction]]