Skip to main content

DOSSIER I · The Agentic Organization (Entry 01 of 05)

An AI Adoption Strategy That Works Without a Terminal

Build an AI adoption strategy around distribution, not hype: a 127-skill catalog, a web storefront, SSO MCP connectors, and role playbooks teams run weekly.

9 min read

By Dakota Smith

Share article

XLinkedIn
On this page6 parts

Architectural scope

Anonymized enterprise consultancy rollout pairing a 127-skill Git catalog, a 49-component web storefront, and dual terminal/browser MCP surfaces.

Cadence
9 min read
Structure
6 sections · 1 code specimen

Buying enterprise seat licenses and pinning a PromptThe brief. Closer to writing instructions for a new contractor than to programming.The text you give the model — question, instructions, context. Its quality shapes the quality of the output.Full definition cheat sheet in Slack is not an AI adoption strategy. Early rollouts fail the same way across almost every mid-market and enterprise organization. Ten percent of software engineers automate half their week in the terminal, while engagement leads, analysts, strategists, and pre-sales directors bounce off an empty browser chat box after two tries.

Designing and shipping an anonymized multi-practice consultancy rollout taught me that an AI adoption strategy is a distribution architecture problem, not a model-capability problem. Frontier models already clear the reasoning bar for discovery synthesis, statement-of-work review, legacy platform scoping, and regression triage. Adoption stalls because internal capabilities hide inside developer-only CLIs, tribal SharePoint folders, and blank-page chat prompts.01 · Catalog FootprintThe production catalog shipped 127 role-scoped skills across 8 capability groups and 30+ consulting roles from one Git spine.

Here is the distribution architecture that got technical and non-technical practices working from one governed source — a 127-skill Git catalog, a 49-component internal web storefront, a zero-terminal claude.ai MCPUSB-C for AI — one standard plug so any agent can connect to any data source or tool.The Model Context Protocol: an open standard (Anthropic, 2024) for connecting AI systems to external tools and data without building a custom integration for each one.Full definition path, and a 45-repository engineering HarnessThe cockpit around the engine. The model is the engine; the harness is everything that makes it useful and safe to fly.The runtime around the model — the loop, tool access, memory, prompts, and guardrails. The model reasons; the harness does everything else.Full definition — along with where this model breaks down.

Why Seat Licenses Are Not an AI Adoption Strategy

Leadership teams usually misdiagnose low AI usage as a training gap and schedule more lunch-and-learns. Training alone fails because rational employees are responding to three structural friction walls.

First is the terminal wall. Builders instinctively package extensions as CLI commands, Node scripts, and repository dotfile trees. That fits developers living inside Git and locks out engagement directors, business analysts, UX researchers, and account leads who run the commercial half of delivery from a browser.

Second is the blank-page and context-scavenger tax. A solution lead with twelve minutes between client calls will not write a system prompt, hunt down SharePoint staffing curves, and paste SOW templates into chat. Their old spreadsheet wins every single time because AI setup costs more effort than the manual habit.

Third is unguarded policy ambiguity. Telling staff to "experiment responsibly" without protocol-level guardrails produces two bad behaviors at once: careful employees avoid AI completely to protect client confidentiality, while rushed contributors paste sensitive discovery notes into unmanaged personal tools.

One Git Spine, Two Consumption Surfaces, and a Web Storefront

The architectural fix was separating where institutional knowledge is maintained from how each role invokes it. Every skill, rubric, hook, and MCP connector lives in one versioned Git catalog, discovered through a browsable internal web hub and projected onto two distinct daily surfaces.

PLATE 02DUAL-SURFACE DISTRIBUTION
REV 2026.11 // ONE SOURCE, TWO HABITS

One Governed Git Spine Projected Into Terminal and Browser Surfaces

Architecture diagram showing a central Git-backed catalog of 127 skills, playbooks, and MCP servers plus a 49-component Starlight storefront branching into two consumption surfaces: a Claude Code CLI marketplace for engineers and a browser-based claude.ai OIDC SSO read-only MCP path for non-technical roles.

SHA-PINNED · CI-VALIDATED · NAMED PRACTICE OWNERS

CENTRAL GIT CATALOG + WEB STOREFRONT

127 Role-Scoped Skills · 8 Capability Groups · Astro/Starlight Discovery Hub

SURFACE A · TERMINAL & IDE

Claude Code Plugin Marketplace

Software Engineers, Solution Architects, QA Automation

  • Single-command CLI install & byte-locked repo rules
  • 8 hard pre-commit/push hooks & sanitized audit logs
  • Local AST, serialization & test-runner tool execution

Human-is-engineer-of-record owns every push and PR

SURFACE B · ZERO-TERMINAL BROWSER

Hosted OIDC MCP + Deliverable Playbooks

Engagement Leads, Business Analysts, UX, Pre-Sales

  • Browser sign-in on claude.ai via enterprise OIDC SSO
  • Live retrieval over 617 DXP records & 81 delivery benchmarks
  • Calendar-named workflows (SOW audit, RFP, backlog, QA matrix)

Read-only protocol surface — zero live system mutation

Shared invariant: no prompt wikis, no copy-pasted API keys, and no divergent skill forks across practices.

PLATE 02: Capabilities are authored, versioned, and CI-checked once in Git, then projected into a CLI marketplace for builders and a zero-terminal SSO web storefront + MCP path for delivery, strategy, and pre-sales roles.

To eliminate repository browsing for non-developers, we paired the Git catalog with a 49-component Astro and Starlight internal documentation storefront. Any practitioner could filter all 127 skills by role, practice, or client lifecycle phase, inspect required inputs and data-handling rules on one screen, and launch the workflow in under thirty seconds.

Surface A met engineers directly inside their repositories. Technical pods consumed the catalog through an internal Claude Code plugin marketplace built on the same registry pattern as my CMS analysis marketplace and ObjectCore. On one global Sitecore XM 10.4 and Azure DevOps engagement spanning 45 repositories and 30+ brand sites, the team harness packaged 24 specialist agents and 8 hard hooks. Its day-loop commands covered /fde morning briefs over a local read-only ADO mirror, /triage, /implement, /smoke, /handoff, and /timesheet. Three-layer zero-secret packaging kept versioned agents in package/, PR-governed repo blast-radius maps in config/, and secret names — never raw values — in machine-local env.json.02 · Engineer of RecordEight hard hooks blocked pushes to main, redacted credentials from audit logs, and kept a named human as engineer of record on every PR.

Surface B gave non-technical roles a zero-terminal browser path. Delivery managers, strategists, and pre-sales leads never installed Node, Git, or a CLI. They signed into claude.ai via corporate OIDC single sign-on, attached pre-approved remote MCP connectors, and invoked structured Claude Code Skills. Behind those connectors sat cited, read-only knowledge engines. Practitioners queried a 617-record DXP operations ontology across 9 version tracks, 81 agency staffing benchmarks, and an 8-role extractor compiling SharePoint history into Open Knowledge Format (OKF) bundles.

connectors/consulting-knowledge-mcp.json
{
  "connector": "practice-intelligence-mcp",
  "auth": "oidc-entra-sso",
  "mode": "read-only",
  "citations": "mandatory-source-url",
  "resources": [
    "dxp-kb://operations/178-canonical-ops",
    "agency-benchmarks://staffing-curves-and-sow-baselines",
    "account-okf://approved-discovery-dossiers"
  ]
}

Making non-technical MCP connectors strictly read-only solved hesitation overnight. An account lead could pull version-pinned upgrade effort factors and house SOW exclusions in seconds, knowing an errant prompt could never mutate a live Azure DevOps board, SharePoint library, or client environment.

Deliverable-Bound Role Playbooks Over Prompt Wikis

Prompt libraries rot because nobody reads a 200-row wiki table before a client workshop. Instead of generic prompt tips, the 127-skill catalog organized around eight lifecycle capability groups mapped to calendar deliverables across 30+ roles.

Capability GroupTarget RolesRepresentative Catalog SkillsGrounded Context + Human Gate?
Scope & Discover (16 skills).Business Analysts, UX, Strategy.discovery-synthesizer, stakeholder-map, opportunity-qualifier, sharepoint-doc-intake.Grounded in OKF dossiers; analyst verifies every requirement traces to a source.
Propose & Win (15 skills).Engagement Directors, Account Leads.sow-generator, rfp-response, estimation-helper, pricing-calculator, partner-brief.Grounded in 81 agency benchmarks; practice director signs off on margin and exclusions.
Design Solutions (15 skills).Solution Architects, Data, UX.architecture-proposal, data-model, design-handoff, adr-writer, prompt-engineer.Grounded in 617 vendor-cited DXP records; principal architect validates blast radius.
Develop, QA & Ship (4 groups).Full-Stack, DXP, QA & Cloud./implement pipeline, /smoke matrix, OWASP/CVE audit, cloud FinOps & incident runbooks.Grounded in repo AST and ADO mirror; human engineer of record owns every PR.

Every playbook on the Starlight storefront answered four operational questions on one screen. It specified when to run the skill, which connector to attach, which client data and PII classes stayed off-limits, and who owned sign-off. Enforcing "no source, no record" inside our underlying platform knowledge base meant upgrade assessments cited exact vendor documentation URLs rather than plausible-sounding model guesses.

Measuring Compounding Adoption Without Vanity Metrics

Seat-activation charts and raw message counts measure curiosity, not operational value. An all-hands demo always produces a one-week traffic spike followed by quiet abandonment if workflows fail under deadline pressure.

Four operational signals separated real adoption from novelty.03 · Dependency SignalYou know an internal catalog is real infrastructure the first morning a PM files a bug ticket against an estimation skill's licensing table.

  1. Four-week cohort retention per deliverable skill. Tracking whether the same analyst or engagement lead ran sow-generator or discovery-synthesizer across four consecutive engagements proved the workflow beat their spreadsheet habit.
  2. Pull requests and bug tickets filed against skills. When non-technical directors started opening issues because a pricing skill missed a new cloud licensing tier, the catalog crossed from experiment to shared operating manual.
  3. Cycle-time compression on bounded consulting artifacts. We measured elapsed time from workshop transcripts to reviewed backlogs, and cut 8-section repo handoff packets to under three minutes, replacing subjective "hours saved" surveys.
  4. Zero-secret audit cleanliness. Checking that team harnesses ran across dozens of client repositories without a single credential leak or unauthorized direct push to main.

Once pre-sales, strategy, and engineering pods depend on the same 127-skill catalog, informal review stops scaling. Keeping unvetted scripts and over-scoped MCP tools out of that shared spine requires an AI governance framework compiled into CI.

Failure Modes and When Not to Build This

Dual-surface distribution solves the terminal wall, yet it introduces real infrastructure responsibilities that rollout slide decks gloss over.

OIDC TokenA chunk of text — roughly three-quarters of a word. The unit the model reads and bills in.The fragments a model breaks text into. Around 100 tokens ≈ 75 English words. Limits and pricing are counted in tokens, not words or characters.Full definition expiry breaks non-technical trust immediately. Developers shrug when a CLI session needs re-authentication; an account executive ten minutes before an RFP readout will not debug an OAuth callback error inside claude.ai. Remote MCP connectors require production uptime monitoring, clean token refresh flows, and clear fallback instructions on the storefront.

Monolithic MCP servers poison tool selection. Exposing forty lookup tools on one universal MCP endpoint bloats context tokens and causes the model to pick the wrong retrieval tool mid-conversation. Partitioning connectors by domain — pre-sales benchmarks, DXP version ontology, and tenant-scoped OKF account memory — kept tool schemas compact and permissions tight.

Unowned skills decay every quarter. Vendor licensing tiers, accessibility standards, and house SOW terms shift constantly. Every skill frontmatter block must name an accountable practice owner and review date, or teams silently fork local copies again.

When NOT to use this approach: Do not build a 127-skill catalog, a custom Starlight storefront, and SSO-brokered MCP servers for a fifteen-person startup or an all-engineering product squad already working inside Git. Below roughly thirty mixed-discipline staff, hosting and curating two consumption surfaces costs more than it returns — start with a shared Git folder of power-user Claude Skills and a tight AGENTS.md file instead.

Conclusion

Getting a whole company to use AI has almost nothing to do with teaching prompt tricks. Adoption compounds when workflow setup drops below the cost of the old habit. Maintain one SHA-pinned Git catalog, publish a plain-English web storefront, give engineers a hook-enforced CLI harness, and wire non-technical teams to read-only SSO MCP connectors.

Key Takeaways:

  • Solve distribution before persuasion. Meet developers inside repo harnesses and non-technical roles inside claude.ai + a browsable web storefront rather than forcing everyone into a terminal.
  • Ground skills in cited institutional records. Pair role workflows with source-enforced ontologies (617 platform records, 81 delivery benchmarks, OKF dossiers) so outputs reflect firm standards instead of generic prose.
  • Enforce safety at the protocol and hook layer. Keep browser MCP connectors strictly read-only and wire hard Git hooks so a human remains engineer of record on every client change.
  • Organize around calendar deliverables. Package skills by real lifecycle outputs — discovery synthesis, SOW risk audits, C4 proposals, smoke matrices — with named practice owners.
  • Right-size the distribution stack. Reserve hosted MCP brokers and custom portals for multi-practice organizations; keep small technical teams on plain repo-level skills.

Next, see how we kept 127+ skills, MCP connectors, and executive AI bets auditable in An AI Governance Framework in Code, Not Slide Decks. Explore our AI Security pillar for the complete permission architecture.

DOSSIER I · ENTRY 01 OF 05

The Agentic Organization

How an engineering practice moves from idle seat licenses to dual-surface distribution, CI-enforced governance, and accountable agent fleets.

Security & Governance Pillar
  1. 01An AI Adoption Strategy That Works Without a TerminalReading now
  2. 02An AI Governance Framework in Code, Not Slide Decks10 min read
  3. 03Agentic Engineering: From Solo Copilots to Governed Fleets9 min read
  4. 04AI Isn't Just for Coders: How Claude Skills Bridge the Gap7 min read
  5. 05Claude Skills for Power Users: Scripts, Teams, and Architecture9 min read

Further analysis

Selected by shared architectural signals and recent publication history.

  1. 01AI tools9 min read

    Claude Skills for Power Users: Scripts, Teams, and Architecture

    Explore Claude Skills' three-tier progressive disclosure architecture and learn to build advanced workflows using executable scripts and team deployment.

  2. 02AI tools7 min read

    AI Isn't Just for Coders: How Claude Skills Bridge the Gap

    Discover why half of all workers avoid AI — and how Claude Skills turn any workflow into a repeatable, quality-controlled playbook. No coding required.

  3. 03AI10 min read

    Building a Deep Research Agent That Verifies Its Sources

    Learn how I built a deep research agent that fans out across sources, then refutes its own findings with three skeptics before reporting a single cited claim.

Discussion

Loading comments...