workflow integration - claude foundations

"I use Claude" vs. "our workflow uses Claude"

The first is a personal productivity habit;

the second is a repeatable process where Claude performs specific steps every time, run by a team.

Your team has a pile of unstructured material — a 40-page RFP, an email thread of half-formed internal answers, meeting notes — and has to act on it. What do you do with Claude (4 things), and what should come back?

Requirements analysis. Hand Claude the raw inputs and ask for a *structured analysis*, not a narrative summary:

every distinct requirement, each with a short label,

the exact source section it came from,

whether existing material already answers it,

and a flag on anything ambiguous.

Ask for it as a table.

RFP example: "From the attached RFP and the email thread, extract every distinct requirement the client is asking us to address. For each, give a short label, the exact RFP section it comes from, whether our thread already has an answer, and flag anything ambiguous that needs clarification. Return it as a table."

A stakeholder tells you "we need better reporting". Why can't you act on that yet, and what do you do with Claude?

Translate the business need into task definitions. "We need better reporting" is not actionable. Claude converts it into specifics: what report, for whom, how often, from what data, in what format. Each specific becomes a requirement you can build against and check completion on.

Claude has extracted a requirements table from a long client document. What are the 2 questions in the second pass? [if not already asked]

Ask Claude to challenge its own list:
AI (mnemonic)
- which requirements are Ambiguous
- which are Implied only indirectly by the source.

This is for Pressure-testing. Extraction is only the first pass.

What makes a requirement ambiguous?

Unquantified words that two builders would resolve differently — "a large number of concurrent users", "fast", "user-friendly". A requirement naming exact formats, systems, or thresholds ("export to PDF and CSV", "SSO through the existing identity provider", "sent within five minutes") is specific even if it is demanding.

name the two types of work Claude can do

Synthesis, where Claude excels (gathering considerations, structuring options, laying out trade-offs),

and calculation,

You are planning next quarter's headcount and you have four quarters of ticket data. How do you get the figures the plan rests on?

Code execution, not prose.

When a plan depends on numbers, compute them: upload the dataset, use code execution for the trend and throughput figures, then have Claude synthesize the recommendation from the verified figures.

What is AI good at vs what are humans good at?

Not every step benefits equally.

Synthesis-heavy steps — many considerations to weigh and structure — are where Claude adds most.

Judgment steps, where a person weighs risk appetite, budget, or political reality - for humans

The iteration loop (solution design) - what are the 4 steps?

Ideate → prototype → gather feedback → refine, repeating until the solution holds.

Claude is a design collaborator, not a vending machine. Run the loop inside a Project so context, constraints, and prior decisions stay stable; otherwise each iteration restarts and you get a pile of one-off drafts instead of a solution. The worked example builds a dashboard artifact in three cycles with no code written by the team — each refinement described, not coded.

When should you escalate to developers?

The moment people rely on the solution as infrastructure — a dashboard artifact three departments now depend on daily, or anything with uptime, security, or integration requirements — it is no longer a prompt-and-iterate exercise and belongs with Developer or Architect expertise. Dependency is the signal, not size or complexity.

You have mapped a workflow step by step and classified each step. What do you build in next?

Once the workflow is mapped, embed the right feature at each AI step: a Skill for repeatable procedure steps, code execution for data steps.

You are describing your team's AI-assisted workflow to a manager, or a client. What is the approach?

State the limits as clearly as the value, and name the human gate.

- Accurate: "Claude drafts the redline; a lawyer approves every change."

- Overstated: "Claude handles contract review."

Overstating is how you lose stakeholder trust on the first visible miss.

You have to explain the same AI-assisted workflow to several different stakeholders. What changes between the explanations, and what must not?

calibrate to the audience: what changes is the level of detail each one needs in order to trust the workflow. What never changes is the workflow itself and the human review gate, which every version names.

- A technical stakeholder wants the feature detail and the known failure modes.

- An executive wants the outcome, the oversight in place, and the risk posture.

- An assurance audience wants the human sign-off stated plainly.

When describing an AI-assisted workflow to stakeholders, which three common phrases overstate what Claude actually does, and what is the one fix?

- "Fully automated"

- "Claude handles X"

- "Basically as good as a person at Y"

Fix: say what the tool does, then name the human checkpoint. Explicit gates make stakeholders trust the workflow more, not less.