[research@ai4se] : ~ $
cd ../
[methodology] | | 11 min

# Run AI4SE as a Business: Double Diamond, Specs, then Skills

Treat AI4SE as a business and the deliverable cannot stop at slides. Consulting engagements can run Spec-Driven: use Design Thinking / Double Diamond to move through pains, opportunities, and solutions; clarify IT4IT-related specifications; then agentify and skillify—atomic Skills first, Goal-driven Agents later, plus Commands, Hooks, and MCP. The hard part is defining the problem and converging on a solution.

[pilot-transformation][sdd][design-thinking][ai4se-framework]

Many peers who do not really do consulting — engineers by trade, or trainers by trade — treat some market “authoritative answer” as treasure: a hot open-source stack, a popular course, the Agent lineup from a flagship case study. A tiny team can sometimes copy-paste and still limp along. In a mid-sized organization, that cargo-cult almost always fails.

The reason is simple: every organization has its own logic, role boundaries, and collaboration inertia. Someone else’s answer solved someone else’s problem. AI4SE transformation has to rediscover a path that fits this organization — you cannot land with the answer already chosen, and you cannot treat “which Agent / which curriculum we adopted” as the deliverable. What clients buy is a verifiable change in how work gets done: which value stream runs cleaner, which human–AI handoff can be accepted, which assets are reusable next week. The three workshop failure modes on this site usually share the same root: after the excitement, what remains is someone else’s tool memory — not an operable path of your own.

Thesis: Operate a Spec Pipeline

Running AI4SE as a business means you need a value chain, specifications, and shippable assets — not a pile of demo scripts.

The Open Group positions IT4IT as managing the business of IT — a reference architecture and value streams for “IT as a business.” Consulting is the same shape: clarify the Specifications that describe R&D and delivery work, then talk automation and Agents. Otherwise an Agent is only executing the goal it invented.

One-line formula:

Double Diamond cadence × IT4IT Spec × assetization (Skill / Agent / …) = operable AI4SE consulting

That is Spec-Driven Development lifted from “write a Spec before code” to the consulting layer: the engagement itself can be Spec-Driven. Specs are the source of truth for ways of working and value streams; agentic / Skill packaging is the second half of delivery.

For the org-level pilot frame, see AI4SE Pilot Transformation for Mid-to-Large R&D Organizations. This piece is about how consulting operates that chain.

Three Frameworks, Three Layers

Do not mash Double Diamond, IT4IT, and SDD into one buzzword. They sit on different layers:

FrameworkOwnsConsulting artifact
Double DiamondDiverge / converge cadencePain inventory → problem definition; opportunities → solution selection
IT4ITReference architecture and value streams for the “IT business”End-to-end capability and data-object map (Spec skeleton)
SDDSpec as source of truth and gateSpecs that can be reviewed, accepted, and handed to Agents

The Design Council’s Double Diamond insists: understand the problem before answering it — diamond one is problem space, diamond two is solution space. Our piece on Design Thinking: cadence, templates, AI brainstorm already covers the three-layer stack; here we only borrow the cadence discipline.

On the SDD side, Thoughtworks / Fowler discussions often place practice on a spectrum: spec-first → spec-anchored → spec-as-source. Most consulting work is fine at spec-first / spec-anchored — Specs must be reviewable, editable, and usable as pilot gates. You do not need to fetishize spec-as-source (“humans never touch generated artifacts”).

Pipeline 1: Pain → Opportunity → Solution

Do not skip the order. Jumping to tool choice before problem definition is an empty second diamond.

PhaseWhat you doTypical output
DiscoverInterviews, as-is value streams, friction (where IT4IT streams break)Raw pains and evidence
DefineFirst hard converge: a debatable problem statementWho, what pain, what evidence, what success, what we will not solve
DevelopParallel opportunities: which handoffs to accelerate, what can become a SkillOpportunity wall / elevator-pitch set
DeliverSecond hard converge: a pilotable solution slice1–2 bounded pilot slices + first Spec

In practice, the workshop Pair → Group → Plenary funnel hardens both convergences: decisions must persist, not die on sticky notes. Elevator pitches and journey maps supply exit criteria — without shape, multi-round brainstorming never agrees on “done clarifying.”

Cadence rule: no vendor selection while diverging; no scope growth while converging.

Pipeline 2: Clarify the IT4IT Spec (Consulting SDD)

In code-side SDD, Specs are the truth of function and design. In consulting, Specs are the truth of ways of working and value streams.

A one-page consulting Spec should at least answer:

SlotQuestion
Value-stream segmentWhich delivery slice? What are the upstream/downstream handoffs?
Roles and dutiesWhat definition rights and vetoes stay with humans?
Inputs / outputs / gatesWhat must be true before work proceeds?
Human–AI boundaryWhat can be script / Skill / Agent; what must stay human judgment
Acceptance evidenceWhat counts as pilot success — reproducible and measurable

“Change goes through the Spec” still applies: if pilot scope moves, edit the Spec before Skills or Agents, or demos will drive drift. A tool list without a Spec falls into the trap described in Humans define, machines execute — machines scale a process that was never made explicit.

Pipeline 3: Duty Transfer and Asset Evolution

Agentic / Skill packaging is transferring already bounded human duties to AI. The usual carriers are Skills and Agents; in coding-agent ecosystems you also accumulate Commands, Hooks, and MCP as connectivity assets.

StageCarrierWhen
EarlySkill (atomic capability pack)Procedure is describable, trigger surface is narrow, verification is writable
ConvergeAgent (Goal-driven)Multi-step loops need planning autonomy, with humans still owning acceptance
CompanionCommand / HookDeterministic gates and repeatable actions — things the model must not merely “remember”
ConnectMCPReach external systems (tools and data); not the methodology itself

They compose; they are not mutually exclusive choices:

  • Skill answers “how we do this the right way” — see Nine types of Claude Code Skills
  • MCP answers “can we reach that system” — see MCP as a standard interface
  • Hook answers “this must happen; the model cannot skip it”
  • Agent answers “under a clear Goal, who runs the closed loop in an isolated context”

Start with Skills to keep blast radius small: atomic capabilities are reviewable, replaceable, and fail narrowly. Once Goals and acceptance are clear, converge Goal-driven into Agents — do not stack “universal collaborators” first; enlarge autonomy only after duties are sharp.

Where It Breaks: Two Convergences

The whole engagement diverges and converges, but most failures land on Define the problem and Deliver the solution.

SignalUsually means
Still comparing models / plugins, no “what we will not solve”Define incomplete
Opportunity wall is full; no pilot slice with Spec and asset boundariesDeliver incomplete
Skills proliferate without Goals or acceptanceAssetization spinning; never converging to Agents or operable capability
Treating MCP or an Agent product as the solution itselfSpec skipped; second diamond captured by tools

Pin the cadence: no deep tool selection before the first diamond converges; on the second diamond, leave with 1–2 slices and park the rest in the backlog. Funnels and ORID-style process discipline exist so consensus lands in one Spec — not in meeting atmosphere.

Anti-Patterns and a Minimal Week

Anti-patterns

  • Treating slides as the deliverable — no Spec, no assets, no acceptance evidence after the pilot
  • Buying Agents / platforms first, hunting for pain later
  • Skills only grow; never Goal-driven converge
  • Encoding deterministic gates as “please remember to lint” Skills instead of Hooks
  • Counting MCP connections or Agent headcount as success

Minimal actions this week

  1. Pick one friction point on a real value stream
  2. Write a one-page problem definition (including what you will not solve)
  3. Add a half-page consulting Spec (human–AI boundary + acceptance evidence)
  4. Run one Skill-level pilot (reproducible and reviewable)
  5. Decide next week whether to Goal-ify into an Agent — on evidence, not vibes

One-line close: Running AI4SE as a business means operating a Spec pipeline — Double Diamond to find the right problem, IT4IT Specs as truth, then transfer human duties onto Skills, Agents, and companion assets. Divergence is easy; the two convergences are the work.

References and further reading