Your Guide

The Claude Opus Workflow Interrogator

4 drop-in Claude Opus Skills that map workflow decay, audit handoff rot, and force a binary deploy/don't-deploy call before AI ever touches the process.

The complete guide is below, free. When you're ready to run it with your leadership team:

Book 30 Minutes →

No deck. No pitch. Just a conversation.

The 4 Claude Opus Skills that replace a $1M Big Consulting workflow audit — drop in your SOPs, system architecture, and process docs and run the intelligence pass yourself.

Word Count: ~4,200 words | Reading Time: ~21 minutes | Replaces: A $400K–$1.5M Big Consulting workflow discovery engagement

A COO at a Fortune 50 manufacturing firm sat across the boardroom table from me last month. He had just received the final deliverable from a massive enterprise workflow audit. Twelve weeks. Eight Big Consulting associates. A 200-page slide deck mapping their operations, beautifully bound, filled with heat maps and maturity quadrants. He pushed the heavy document across the table.

“I still have no idea what we should actually go build first.”

Listen to this. This happens every single day inside the enterprise. Operators are paying millions of dollars for workflow intelligence that produces absolutely zero deployable clarity. They are told they need to map their business before they can apply AI, which is true. But the artifact they buy to solve that problem is fundamentally BROKEN.

This is why 95 percent of enterprise AI pilots fail. They fail for one very specific, entirely preventable reason: operators automate chaos. They take a fundamentally broken, undocumented, rotting workflow and they layer an expensive agentic AI on top of it. You cannot automate chaos. The technology will just execute the broken process a thousand times faster, creating a blast radius that will take your data teams months to clean up.

Workflow intelligence has to come BEFORE AI. You have to expose the decay before you build the tool.

Here is the secret Big Consulting does not want you to know: the entire workflow audit playbook is completely commoditized. You do not need a massive team of associates interviewing your staff for months. You just need the right intelligence pass.

I mapped the entire forensic discovery process to Claude Opus. This guide gives you the 4 drop-in Skills to run the interrogator yourself. Drop in your SOPs, your system architecture, and your process docs. Run the passes. Expose the decay.

I am giving you the recipe. The recipe is free. But the RESTAURANT is what matters. AI handles the analysis and the audit. But judgment, change management, and actual production deployment still require senior human operators.

By the end of this guide, you will have a binary, operator-grade verdict on exactly which workflows are ready for AI, and which ones will set your capital on fire.

Let’s get to work.

Why Big Consulting’s Workflow Audits Fail

The standard $400K–$1.5M workflow discovery engagement runs roughly the same shape no matter who you buy it from.

A partner sells the engagement on the premise of identifying your highest-ROI AI automation targets. A few weeks later, six associates land in your Microsoft Teams channels. They spend eight weeks interviewing your business unit leaders. They take copious notes. They build a massive bottleneck report. You get a workflow heat map spanning dozens of processes, color-coded to show where the “friction” lives.

You cannot make a single production decision from this artifact.

There are three structural reasons this playbook fails, and none of them have to do with the intelligence of the associates. It is a failure of the mechanism.

Associates cannot map institutional knowledge. Big Consulting associates write down what people SAY happens. They do not map what actually happens. When an associate asks a manager how an invoice is approved, the manager recites the official SOP. The associate documents the happy path. What the associate does not capture is that the official system has been broken for three years, and the actual approval happens when an analyst sends a direct message to a VP, who replies with a thumbs-up emoji. The real workflow lives in the shadows. Associates miss the shadow workflow entirely. When you deploy an AI agent against the official SOP, the agent breaks on day one because it does not know about the shadow process.

The handoff problem. The team that writes the workflow audit is a strategy team. They build decks. They do not build software. Six months later, a completely different team of engineers is hired to layer AI on top of the workflow. The engineering team was not in the room for the discovery interviews. The strategy deck lacks the technical depth the engineers need to design the agentic system. The audit goes stale before the first line of code is written. The handoff between the people who map the business and the people who build the tool destroys the continuity required to ship working AI.

Mismatch with operator reality. Operators do not want a 200-page heat map. Heat maps do not ship. An operator needs a binary call. Can we deploy an autonomous agent against this supply-chain workflow next quarter? Yes or no. If no, what specifically is broken, and in what order do we fix it? Big Consulting structurally refuses to answer that question with a binary yes or no, because a binary call carries accountability. A heat map is safe. A heat map says “everything is sort of a priority.”

The 4 drop-in Skills below eliminate these failures. They do not produce heat maps. They produce binary, operator-grade deployment calls.

The 4-Skill Interrogator At A Glance

The Claude Opus Workflow Interrogator is built on four progressive intelligence passes. Each Skill requires the output of the previous Skill. You cannot skip a step.

Skill 1: Workflow Cartography. You drop in your SOPs, training docs, and process maps. Claude outputs a current-state map that exposes the decision points, exception paths, and named human owners of institutional knowledge.

Skill 2: Handoff Forensics. You drop in the mapped workflow from Skill 1. Claude forensically hunts for operational ROT — broken handoffs, queue time, undocumented approvals, and single points of failure.

Skill 3: Architecture Auditor. You drop in your system architecture diagrams, integration lists, and data dictionary. Claude exposes the structural decay beneath the workflow, mapping the blast radius of what happens if an AI agent makes a mistake.

Skill 4: Automation Readiness Verdict. You drop in the outputs of the first three Skills. Claude synthesizes the intelligence and forces a binary GREEN or RED call per workflow. You get your realistic 12-month deployment surface.

Plan on 2 to 4 working days per business unit to run this interrogator. You need two senior operators who deeply understand the business, and you need Claude.

The output is not a strategy deck. The output is a binary GREEN/RED call per candidate workflow and a sequenced fix list for the workflows that fail the gates.

Each Skill below ships as a copy-paste drop-in prompt. Copy the prompt, paste it into Claude, drop in your inputs, and run it. NEVER accept an output that deviates from the strict formatting demanded in the prompt.

Skill 1 — Workflow Cartography

Every workflow audit has to start with mapping the current state. But where Big Consulting maps the happy path, you have to map the DEEP reality.

If you do not map the reality of how the work actually gets done, you cannot audit the handoffs in Skill 2. This is the prerequisite. You drop in your standard operating procedures, your training manuals, your outdated Visio diagrams, and any process documentation you have.

Skill 1 forces Claude to map the workflow while explicitly hunting for the three things every consulting audit routinely misses.

Decision points. Every single place in a workflow where a human chooses path A versus path B. If an analyst looks at a customer email and routes it to Tier 2 support instead of Tier 1, that is a decision point. Why did they make that choice? Was it the tone of the email? Was it the account size? If the decision logic is not explicitly mapped, an AI agent cannot replicate it. Big Consulting writes “Review Email.” You need to map the exact cognitive boundary of the decision.

Exception paths. What happens when the happy path breaks. The vendor portal goes down. The customer inputs a blank field. The API times out. Enterprise workflows spend HALF THE TIME in exception states. If your process map assumes everything works perfectly, you are building a toy, not an enterprise system. You must map the exact sequence of events that occurs when the primary route fails.

Institutional knowledge owners. The named human beings who hold undocumented context in their heads. In every enterprise workflow, there is a person who knows why a specific workaround exists. There is a person who remembers why the legacy database requires a manual refresh on Fridays. If you do not attach a named human owner to these nodes of the workflow, that knowledge disappears the moment you try to automate the process.

Run the Skill. Force the cartography.

Copy-paste prompt — Skill 1: Workflow Cartography

You are an enterprise workflow cartographer mapping current-state operations for AI deployment readiness.

I will paste below: raw standard operating procedures (SOPs), training documents, and process maps for a candidate enterprise workflow.

For every discrete step in the workflow, score it on these three axes from 1 (worst) to 5 (best):
1. Decision Point Clarity — 1 means "magic happens here," 5 means the exact logical conditions for choosing path A vs path B are strictly defined.
2. Exception Path Mapping — 1 means only the happy path is documented, 5 means every failure state has a documented recovery route.
3. Knowledge Owner Identification — 1 means anonymous "team" ownership, 5 means a specific named human role is explicitly tied to the step's execution.

For each axis, give the score AND a one-sentence justification grounded in the input I gave you. Reject any score that does not include the justification.

Decision Rule: If any step scores a 1 or 2 on any axis, flag that specific step as an AUTOMATION BLOCKER. If the inputs provided are too vague to even identify the discrete steps, refuse to output the table and state: "Insufficient documentation to map current state."

Output as a markdown table: Workflow Step | Decision Point Score | Exception Path Score | Knowledge Owner Score | Justification | Blocker Status.

End with: "Top 3 undocumented decay points currently blocking the cartography of this workflow:" — list them.

Inputs:
[paste your SOPs, training docs, and process maps here]

The decision artifact from this Skill is a structured workflow map with named owners, explicit decision gates, and mapped exception paths. If the Skill flags automation blockers, you stop. You go find the named human who does the work, you interview them, and you update the input.

Skill 2 — Handoff Forensics

Once you have the mapped workflow from Skill 1, you have to look at the seams.

Work does not break in the middle of a step. Work breaks in the handoff between steps. Handoff decay is the number one killer of agentic AI deployments. If you automate a process that contains a broken handoff, the AI agent simply inherits the broken handoff and executes the failure at scale.

Skill 2 takes the map you just built and forensically interrogates it for five specific markers of operational ROT.

Queue time. Work waiting for a human. Where does the workflow sit idle? Does a completely automated data extraction process dump a file into an inbox where it sits for 48 hours waiting for a human to open it? If you layer AI over the extraction but ignore the queue time, your total cycle time does not improve.

Role ambiguity. Multiple people think they own the handoff, or nobody does. The marketing team thinks compliance owns the review. Compliance thinks legal owns the review. Legal thinks they only review it if marketing flags it. The handoff drops. An agentic system requires absolute, deterministic role clarity.

Undocumented approvals. The gates that exist in practice but not on paper. The shadow IT of the enterprise. If the official SOP says a purchase order goes from generation to fulfillment, but reality requires an email to the regional director for a “quick sanity check,” you have an undocumented approval. Agents cannot navigate shadow approvals.

Single-point-of-failure people. The workflow halts completely if one named person takes a vacation. If Dave is the only person who knows how to reconcile the legacy inventory codes, Dave is a single point of failure. You cannot scale an AI deployment that depends on Dave being awake.

Silent dependencies. Upstream system A breaks. Downstream workflow B halts. Nobody in workflow B knows why, because the dependency is silent. The teams do not talk. AI deployed in workflow B will hallucinate or fail endlessly because it is starved of upstream context it does not know it needs.

Run the handoff forensics.

Copy-paste prompt — Skill 2: Handoff Forensics

You are an enterprise workflow forensic auditor hunting for operational decay and broken handoffs.

I will paste below: the completed Workflow Cartography map from Skill 1, including the steps, decision points, and exception paths.

For every handoff (the transition between one step and the next), score it on these five axes from 1 (severe decay) to 5 (frictionless):
1. Queue Time — 1 means work sits idle for >24 hours, 5 means immediate machine-speed transition.
2. Role Ambiguity — 1 means ownership is disputed or unknown, 5 means deterministic, single-threaded ownership.
3. Undocumented Approvals — 1 means known shadow approvals exist, 5 means all governance is explicitly mapped.
4. Human Bottlenecks — 1 means a single named individual is required to proceed, 5 means load-balanced or automated transition.
5. Silent Dependencies — 1 means the handoff relies on upstream unmonitored systems, 5 means all dependencies are visible and alerted.

For each axis, give the score AND a one-sentence justification grounded in the input I gave you. Reject any score that does not include the justification.

Decision Rule: Any handoff that scores a 1 or 2 on ANY axis is flagged as RED DECAY. A score of 3 or 4 is YELLOW DECAY. A 5 is GREEN. If the cartography map lacks sufficient detail to score the handoffs, refuse to output the table and state: "Handoff visibility too low to audit."

Output as a markdown table: Handoff Point (Step A to B) | Queue Time | Role Ambiguity | Approvals | Bottlenecks | Dependencies | Worst Marker | Status (RED/YELLOW/GREEN).

End with: "Ranked remediation list: The top 3 handoffs that must be fixed before AI can be deployed."

Inputs:
[paste your Skill 1 workflow map here]

The decision artifact is a ranked decay list. It tells you exactly what to fix first, second, and never. You cannot proceed to architecture until you know exactly where the operational process is bleeding.

Skill 3 — Architecture Auditor

Handoffs reveal process decay. Architecture reveals system decay. You must expose both before you can make a deployment call.

Skill 3 moves from the human process to the technical foundation. You drop in your system architecture diagrams, your integration list, and your data dictionary. If you are about to give an AI agent read/write access to your enterprise systems, you have to know what happens if the agent makes a mistake.

The Architecture Auditor surfaces four deeply hidden structural failures.

Data lineage gaps. A value appears in a downstream dashboard, but the upstream source is completely unknown. A financial metric is populated, but nobody can point to the raw table it was calculated from. If you point an AI agent at a system with broken lineage, you cannot debug the agent’s output. When it hallucinates, you will not know if the model failed or if the underlying data was garbage.

Duplicate systems of record. The official source of truth, and the shadow source of truth. The enterprise CRM says one thing. The sales operations Excel sheet says another. Both are treated as truth by different teams. If you do not resolve duplicate systems of record, your AI will confidently give the wrong answer to the right question.

Integration debt. Point-to-point glue that nobody can describe. Scripts written by an engineer who left four years ago that move data from a legacy mainframe to a modern data warehouse. It works, but nobody knows how. Integration debt shatters when new AI workloads increase the call volume or change the data payload.

AI blast radius. For any candidate workflow, what other systems share data and would propagate a bad AI output? If an agentic system incorrectly updates a customer tier in the CRM, does that automatically trigger a billing change in the ERP? Does it trigger a compliance alert? You must map the blast radius of an AI failure before you give the agent the keys.

Drop in your technical truth and run the auditor.

Copy-paste prompt — Skill 3: Architecture Auditor

You are an enterprise system architecture auditor evaluating structural readiness and blast radius for AI deployment.

I will paste below: system architecture diagrams, an integration list, and a data dictionary.
Fallback rule: If no data dictionary is available, list the top 10 tables with their owning team and last-updated date instead. Process the audit using the fallback if invoked.

For every core system involved in the candidate workflow, score it on these four axes from 1 (critical risk) to 5 (production solid):
1. Data Lineage — 1 means origin data is untraceable, 5 means perfect provenance from origin to downstream consumer.
2. System of Record Clarity — 1 means conflicting shadow databases exist, 5 means a single, globally enforced source of truth.
3. Integration Health — 1 means undocumented legacy point-to-point scripts, 5 means modern, documented, monitored APIs.
4. Blast Radius Containment — 1 means a bad data write instantly corrupts downstream critical systems, 5 means strict validation gates contain errors locally.

For each axis, give the score AND a one-sentence justification grounded in the input I gave you. Reject any score that does not include the justification.

Decision Rule: If any system scores a 1 on Data Lineage or Blast Radius Containment, flag the system as a SEVERE STRUCTURAL BLOCKER.

Output as a markdown table: System Name | Lineage Score | Record Clarity Score | Integration Score | Blast Radius Score | Justification | Blocker Status.

End with: "System-level decay map: The specific blast-radius propagation paths if an AI agent commits a critical write error in this architecture."

Inputs:
[paste your architecture diagrams, integration list, and data dictionary (or fallback list) here]

The decision artifact is a system-level decay map. You now know exactly what systems are too fragile to support autonomous agents. Both the human process and the technical systems are now on the table. It is time to make the call.

Skill 4 — Automation Readiness Verdict

This is the move that justifies the entire interrogator. This is where you replace the $1M Big Consulting heat map with a binary operator decision.

Skill 4 synthesizes the outputs from the Cartography, the Forensics, and the Architecture passes. It forces every candidate workflow through four non-negotiable readiness gates.

Cartography gate. Is the workflow fully mapped? Does it include explicit decision logic, documented exception paths, and named human knowledge owners?

Handoff gate. Are there zero RED decay markers from the Skill 2 forensics? You cannot deploy AI across a broken handoff, an undocumented approval, or a silent dependency.

Architecture gate. Are there zero SEVERE STRUCTURAL BLOCKERS from the Skill 3 auditor? The lineage must be traceable and the blast radius must be contained.

Operator gate. Does a named human owner exist for the post-AI operating model? Who monitors the agent? Who handles the exceptions the agent kicks out? If there is no named human owner for the new reality, the workflow fails the gate.

The verdict output is completely BINARY. It is GREEN, meaning you can deploy AI against this workflow right now. Or it is RED, meaning you must fix the decay first.

There is no YELLOW. There is no “in progress.” There is no “maybe.” If you allow a yellow tier, you are flinching. You are building a heat map. Force the call.

When you run this Skill across your list of candidate workflows, your realistic 12-month deployment surface will emerge. You will realize that out of 30 theoretical AI use cases, only three actually pass all four gates. That is the truth Big Consulting hides from you.

Copy-paste prompt — Skill 4: Automation Readiness Verdict

You are an enterprise AI deployment authority making final, binary production calls.

I will paste below:
1. The Cartography map (Skill 1 output)
2. The Handoff Forensics scores (Skill 2 output)
3. The Architecture Auditor scores (Skill 3 output)
4. The proposed human owner for the post-AI operating model.

Evaluate the candidate workflow against these four strict gates:
- Cartography Gate: Pass only if decision points and exceptions are fully mapped with >3 scores.
- Handoff Gate: Pass only if there are ZERO "RED DECAY" markers in the forensics.
- Architecture Gate: Pass only if there are ZERO "SEVERE STRUCTURAL BLOCKERS" in the architecture.
- Operator Gate: Pass only if a specific, named human owner is provided for post-AI exception handling.

For each gate, state PASS or FAIL with a one-sentence justification.

Decision Rule: You must assign a final, binary verdict for the workflow.
GREEN = Deploy AI now. (Requires PASS on all 4 gates).
RED = Fix first. (Any FAIL on any gate).
You may NOT use YELLOW. You may NOT use "in progress." You must force the binary call.

Output as a markdown table: Workflow Name | Cartography Gate | Handoff Gate | Architecture Gate | Operator Gate | Final Verdict (GREEN/RED) | Single Largest Gap.

End with: "Workflows that pass all 4 gates: [Count]. Workflows blocked: [Count]. Quarter to ship deployable workflows: [Estimate based on green status]."

Inputs:
[paste Skill 1, 2, and 3 outputs, plus the proposed post-AI human owner here]

The decision artifact is a one-page table. Workflow. Verdict. Single Largest Gap. Quarter to Ship.

This is the document you take to the board. This is the document you give to engineering.

How To Run The 4 Skills In Sequence

You have the tools. Now you need the operating cadence. Do not stretch this out over eight weeks. Compress the timeline and force the momentum.

You need two senior operators in the room. Not juniors. Not project managers. You need people who actually understand how the business makes money and where the operational bodies are buried. One operator drives Claude and executes the prompts. The second operator challenges the outputs. When Claude says a handoff is clean, the second operator must verify if that is actually true on the floor.

Day 1–2: Cartography. Take one business unit. Gather the SOPs. Run Skill 1. When the automation blockers trigger, get on the phone with the floor managers. Fill the documentation gaps immediately. Map the reality.

Day 2–3: Forensics. Take the mapped workflows and run Skill 2. This is where the arguments will happen. A VP will realize their flagship process is completely dependent on undocumented shadow approvals. Let the forensics expose the rot. Do not soften the scores.

Day 3–4: Architecture. Run Skill 3 across the technical stack that underpins the mapped workflows. Pull in a senior data engineer to validate the integration lists and the blast radius estimates.

Day 4: The Verdict. Run Skill 4. Synthesize the intelligence. Force the GREEN/RED calls.

What do you do with the RED workflows? You do not buy AI tools for them. You build a sequenced remediation roadmap. You hand the exact failure markers (the broken handoffs, the lineage gaps) to your operations and data teams. You fix the process as discrete operational bets. Once the process is fixed, it becomes GREEN.

What do you do with the GREEN workflows? You deploy. You now have a mathematically proven, structurally sound foundation. You can layer agentic AI onto these workflows without setting $1M on fire. The agents will actually work, because the process they are executing is clean, documented, and governed.

What Comes Next

You now have the playbook. The mechanism is real. The prompts are operator-grade. You can run this workflow interrogator yourself.

If you are going to do this internally, here is an honest estimate of what it takes. Plan on 80 to 160 hours of senior operator time per business unit. You will need 2 to 4 weeks of calendar time to track down the shadow documentation and interview the knowledge owners. You need two senior leaders dedicated to the outcome, and you need to budget for Claude API spend.

But no drop-in Skill can replace the hardest parts of enterprise transformation.

Claude cannot handle the politics when a business unit head learns their flagship, multi-million-dollar workflow is rated RED and fundamentally broken. Claude cannot negotiate with the CFO when your realistic 12-month deployment surface is three workflows, not the 30 they were promised in the vendor’s strategy deck.

A prompt cannot sequence the talent hiring and technology vendor decisions against a rigid board calendar. And an AI cannot keep the verdict honest when a highly sponsored, politically protected workflow fails the readiness gates but the executive sponsor demands it ships anyway.

Those are the parts where the 95 percent AI-pilot-failure rate lives. It does not live in the audit. It lives in the execution. It lives in the courage required to look at a broken process and REFUSE to automate it.

You have to follow the only sequence that works.

Understand the work. Build the target operating model (inclusive of AI). Build the tool.

Get The Next Brief

One operator. Every other Wednesday.

Real business leaders. Real AI deployments. Real numbers — including the parts that didn’t work.

Subscribe at thecraftofai.com.

If you want to talk about running this interrogator on a real enterprise, book a session.


Run the prompts. Then, if you want a second set of eyes:

I read every result a reader sends me. If your program looks stalled and you want 30 minutes on it, grab time on my calendar. No deck. No pitch. Just a conversation.

Book 30 Minutes →

Get the next Brief

One operator. Every other Wednesday.

Plus the AI Glossary and the Failure Museum.
Real names. Real numbers. Honest analysis.