On June 5, 2026, serial founder Hiten Shah posted a long X thread arguing that every company's first AI strategy should be a skill library. The point was practical: top performers already use repeatable patterns, but most companies keep those patterns scattered across Slack, docs, calls, templates, onboarding notes, and the heads of experienced people.
In Shah's framing, AI agents do not become useful just because they can access your CRM, Slack, Google Drive, GitHub, or data warehouse. Access tells the agent where information lives. Skills tell the agent how the company works.
This article turns that thesis into an implementation guide: what a company skill library is, how it differs from prompts and connectors, which workflows to capture first, how to write useful skills, and where private skill libraries can become either durable operating leverage or another messy knowledge base.
TL;DR
Question
Practical answer
What happened?
Hiten Shah argued on X that companies should start their AI strategy by building a reusable skill library from top performers' methods.
Core thesis
A company's AI advantage comes less from choosing the same frontier model as everyone else and more from teaching agents the company's specific way of working.
A prompt is a one-off instruction; a skill is a reusable method with procedures, examples, checklists, templates, references, scripts, and quality bars.
Skill vs connector
Connectors expose data and systems; skills encode judgment and process. Real agents need both.
Best first skills
Sales call prep, support escalations, PRD review, incident postmortems, renewal risk, board decks, launch briefs, contract review, and forecast narratives.
Main risk
Encoding stale, biased, or unsafe process as "best practice" and giving it too much authority without owners, tests, and review.
The real insight: your company already has skills
Most companies talk about "experience," "judgment," "taste," or "institutional knowledge" as if those qualities are too fuzzy to package. Shah's thread makes a sharper claim: much of that work is already procedural.
A strong salesperson does not simply "have good instincts." They inspect account history, identify the real buyer, look for unstated objections, check promises made in prior conversations, and shape open questions around renewal risk.
A good support lead does not just read a ticket. They look at tone, customer history, account value, product pain, timing, prior escalations, and whether a small issue is about to become an executive problem.
A finance leader does not just read a dashboard. They separate meaningful movement from noise, decide which metric changes need a story, and know what will create board-level concern.
Those are methods. They may not be written as methods yet, but they are methods. A skill library is the process of turning that invisible operating knowledge into reusable agent-readable infrastructure.
What an AI skill library is
An AI skill library is a collection of versioned workflow packages that teach agents how to perform repeatable work in a specific context.
In the Claude ecosystem, Anthropic's official docs describe skills as directories centered on a SKILL.md file with YAML frontmatter and Markdown instructions. The Claude Code skills docs say to create a skill when you keep pasting the same checklist, instructions, or multi-step procedure into chat. Anthropic's public Agent Skills repository shows the common shape: self-contained folders with SKILL.md, optional templates, examples, references, and scripts.
For a company, that means a skill might contain:
Skill component
What it captures
Instructions
The steps the agent should follow every time
Examples
Good briefs, bad briefs, real anonymized cases, before/after outputs
Internal policy, product docs, tone guide, pricing rules, architecture notes
Scripts
Safe data pulls, validation helpers, formatting tools, report generators
Rules of thumb
The judgment experienced operators apply before making a call
The important word is library. A single skill is useful. A maintained collection of skills becomes an operating manual that agents can actually load and apply.
Skills, prompts, connectors, MCP, and plugins
Shah's thread separates a common source of confusion: companies often start AI strategy by connecting data, then wonder why output still feels generic.
Access is necessary. It is not sufficient.
Layer
Primary job
Example
Failure mode
Prompt
Ask for one result now
"Write a renewal brief for this account"
Re-explained every time; quality varies
Connector
Give the agent access to systems
Salesforce, Slack, Google Drive, GitHub, warehouse
Agent can read data but lacks method
MCP/API tool
Let the agent call live tools or resources
Query accounts, open tickets, fetch docs, create issues
Tool access without judgment creates noisy automation
Skill
Teach the agent a reusable way of working
"How our team prepares renewal calls"
Stale or vague skills become ritualized bad process
Plugin/workflow
Bundle instructions, tools, actions, and distribution
A sales-prep plugin with CRM access and a skill
Higher operational and security surface
Connectors answer: what can the agent see or do?
Skills answer: how should the agent think and work?
Plugins and richer agent systems combine both. They expose tools, load procedures, and execute workflows. But if the method is missing, the agent is still mostly improvising with access.
What changed, why it changed, what is noise, what needs action, what questions the board will ask
People ops
Performance review calibration
Evidence quality, bias checks, level expectations, examples, growth plan clarity
Legal
Contract fallback review
Standard terms, redlines, deal size thresholds, escalation points, non-negotiables
The best first skill is not always the flashiest workflow. It is the workflow where the same mistakes keep getting corrected by the same senior person.
A practical audit for finding company skills
Use this as a two-hour workshop with a functional team.
1. List repeated work
Write down tasks that happen every week or month: renewal prep, launch review, customer escalation, incident postmortem, PRD review, release notes, pipeline review, board update, hiring scorecard, competitive analysis, security review.
2. Identify uneven quality
Circle workflows where output quality varies heavily by person. Skill libraries are most valuable where judgment is uneven, not where the task is already mechanical.
3. Find the top performer
Pick the person whose work others keep copying, forwarding, or asking to review. Interview them while they do the work live. Do not ask only for principles; watch the sequence.
4. Extract the method
Capture:
What they inspect first
Which signals they ignore
Which edge cases change the recommendation
Which examples shaped their approach
Which mistakes they are trying to prevent
What a "good" output looks like
What they would reject and why
5. Turn it into a skill
Write a short SKILL.md with trigger conditions, a numbered procedure, required inputs, output format, examples, and failure checks. Add supporting templates and references only when needed.
6. Test against real work
Run the skill against three to five historical tasks. Compare the output to what the top performer actually did. Revise until the skill helps a capable operator, not just a demo prompt.
What a company skill should look like
Here is a lightweight shape for a private company skill. This is not the only valid format, but it is a good starting point for Claude-style and SKILL.md-style agents.
md
---
name: renewal-call-prep
description: Use when preparing for an important customer renewal, expansion conversation, or executive account call.
---## Goal
Prepare a concise renewal call brief that surfaces risk, context, open questions, and recommended next moves.
## Required inputs- Account name and renewal date
- CRM opportunity history
- Last three customer conversations
- Open support issues
- Usage and billing changes
- Prior commitments made by sales, support, or product
## Process1. Identify the economic buyer, champion, blockers, and missing stakeholders.
2. Summarize the current account state in five bullets or fewer.
3. Look for unresolved promises or follow-ups from prior conversations.
4. Flag renewal risks using product pain, usage changes, support tone, and executive involvement.
5. Draft open questions that reveal risk without sounding defensive.
6. Recommend the next best action and owner.
## Output format- Account snapshot
- Renewal risk
- Hidden objections
- Promises to verify
- Questions to ask
- Recommended next action
- Confidence and missing data
## Quality bar
Do not produce generic discovery questions. Tie every risk and question to a specific account signal.
Notice the skill is not a motivational essay. It is a working procedure with trigger conditions, required context, steps, output format, and a quality bar.
What belongs in the skill versus the data layer
Teams often overload skills with data that should live elsewhere. Keep the distinction clean.
Put in the skill
Keep in systems/connectors
How to inspect account history
Live CRM records
Which signals matter
Ticket data
Output format and review checklist
Call transcripts
Escalation thresholds
Current customer contract
Examples of good reasoning
Warehouse metrics
Brand voice and legal posture
Product analytics
Skills should explain the method. Connectors, MCP servers, APIs, and internal tools should provide fresh data.
This is why a serious skill library usually grows alongside better access control, retrieval, and MCP infrastructure. The skill tells the agent what to do; the tool layer lets it do the work against current information.
Versioning: a skill library is not a wiki
A skill library should be managed more like product code than like a static documentation folder.
At minimum, each important skill needs:
Owner: the person or team accountable for correctness
Version history: what changed and why
Review cadence: monthly or quarterly for fast-moving workflows
Test cases: real historical tasks or synthetic fixtures
Permission model: what data and tools the skill may use
Deprecation path: how stale skills are retired
If nobody owns a skill, it becomes organizational folklore with better formatting.
The security and quality problem
Private skills can improve work, but they also create a new control surface. A skill can tell an agent what to trust, which tools to use, how to interpret private data, and when to escalate. That means it deserves security review.
The agent follows old rules after the business changes
Owners, review cadence, changelog
Encoded bias
One top performer's preferences become policy
Cross-functional review and outcome testing
Overbroad permissions
A skill can call tools or access data it does not need
Least privilege and scoped tool access
Prompt injection
External text manipulates the agent through connected data
Treat untrusted content as data, not instructions
Leaked private method
Sensitive internal procedures are copied into public repos or vendor tools
Classification, repo boundaries, vendor review
No measurement
A skill feels useful but does not improve outcomes
Before/after quality checks and operational metrics
The easiest trap is to confuse "top performer wrote it" with "this should be automated." Skills should be extracted from great work, then tested against reality.
A 30-day rollout plan
Week 1: Inventory and selection
Pick one department and list repeated workflows. Score each workflow by frequency, business impact, quality variance, and availability of a strong owner.
Choose two skills. One should be low-risk and high-frequency. The other should be judgment-heavy enough to prove whether the library matters.
Week 2: Shadow and draft
Watch the top performer do the work on real examples. Capture their sequence, not just their advice.
Draft each skill with a trigger description, required inputs, process, output format, examples, and anti-patterns. Keep the main skill concise. Put bulky examples and policy docs in supporting files.
Week 3: Test and revise
Run each skill against past cases. Compare the output against real accepted work. Ask the owner to mark what was useful, wrong, generic, missing, or unsafe.
Revise the skill until it reliably raises the floor for a competent teammate.
Week 4: Ship with governance
Publish the skill in a private repo or managed agent environment. Assign ownership, document tool permissions, define review cadence, and set one outcome metric.
Do not announce "the AI strategy." Announce the first two repeatable methods the company has made agent-usable.
Metrics that tell you whether skills work
Skill libraries should be measured by work quality, not install count.
Useful metrics include:
Time to prepare a sales or support brief
Number of manager rewrites required
Escalation accuracy and false positives
Review pass rate on PRDs, launch briefs, or incident reports
Reduction in repeated explanation from senior employees
Consistency of output format across teams
Improvement in task outcome, not just faster artifact generation
Number of stale or unused skills removed
A good skill library should make work more consistent without hiding accountability. The agent can draft, inspect, and recommend. Humans still own decisions.
How this connects to the broader skills ecosystem
Shah's thread lands because the tooling layer is finally catching up to the management idea.
Anthropic's skills docs describe personal, project, enterprise, and plugin skill locations for Claude Code. The public Anthropic skills repo shows SKILL.md packages with optional examples, templates, references, and scripts. Google has published vendor skills for Google Cloud and Gemini workflows. Public registries now index community skills, while companies are starting to build private libraries for internal operating methods.
Hiten Shah's skill-library thesis is not that every company needs more prompts. It is that every company already has repeatable ways of working, and the next AI advantage comes from making those methods visible, reusable, versioned, and agent-loadable.
The first-order AI strategy is not "connect everything and hope the model figures it out." It is: choose important repeated work, study the people who do it best, extract the method, encode it as a skill, test it on real cases, and keep improving it.
Models will keep changing. Your company's operating method is the thing worth compounding.
Status note: this article is based on Hiten Shah's June 5, 2026 X thread as provided in the prompt, plus public agent-skills documentation from Anthropic and explainx.ai's existing skills coverage. Verify live social-post details and event links directly on X or hiten.com before citing the thread in formal materials.
