Convert Walkthrough Recordings into a Swimlane Draft with OPEN Flags

Workflow area: Process Modeling & Mining · Tier: Core · Hands-on time: ~35 minutes

Recorded walkthroughs and shadowing sessions are the highest-fidelity elicitation you can get — people demonstrating work reveal the workarounds, waiting and system-hopping that narrative interviews smooth over. But speech is a terrible process notation: steps arrive out of order, actors are "she" and "the other team," and half the systems are "the tool." Jumping straight from transcript to diagram lets the model paper over every one of those gaps with invented connective tissue. The fix is the intermediate artifact: an actor-step-system table that forces each step to name who, what, and in which system — and makes every hole visible as an OPEN flag before any diagram exists. The table is checkable against the transcript line by line; the swimlane is then a mechanical rendering of verified rows, and the OPEN flags become your follow-up question list.

When to use — and when not to

• Use it after shadowing and demo-style walkthroughs"share your screen and do a real one" sessions are exactly the raw material this pipeline is built for.
• Use it when the SOP is fictionwhere documented process and lived process have diverged, the transcript is your only honest source for the as-is (feed the result into pb-23's walkthrough for validation).
• Use it to triangulate multiple performersrun the pipeline per person and diff the tables; where two operators' steps differ, you have found either a variant or a workaround worth naming.
• Skip it without recording consenta transcript of someone's working session is personal-adjacent data; no consent, no recording, no pipeline.
• Skip it for processes you can mineif the process leaves an event log, process mining (pb-25) beats transcript archaeology for frequencies and paths; use this pipeline for the human steps mining cannot see.
• Skip the straight-to-diagram shortcuttranscript to swimlane in one prompt is precisely the move that lets invention hide; the table is the discipline, not overhead.

The recipe

1. 1.

   Capture and prepare the transcript. Record the walkthrough with consent, get the meeting-AI transcript, and skim it once, marking section boundaries ("intake demo starts ~14:00") and correcting mis-transcribed system names — the model cannot know "S-A-P" and "sap" are the same system unless the text says so.

2. 2.

   Extract the actor-step-system table. Paste into Claude with this prompt:

   Copy

   You are converting a process-walkthrough transcript into an
   actor-step-system table. The speaker is demonstrating real work.
   
   Output a table: Step # | Actor (role) | Action (one verb phrase) |
   System/tool | Input | Output | Evidence (transcript timestamp +
   short verbatim quote) | Flags
   
   Rules:
   1. Every row must cite transcript evidence. No row without a
      quote.
   2. If the actor, system, trigger or outcome of a step is not
      stated, write OPEN: [specific question] in Flags — do NOT
      infer it.
   3. If the speaker describes a workaround ("normally X, but I
      just..."), keep BOTH: the official step and the workaround
      row, flagged WORKAROUND.
   4. Preserve the order work actually happened, not the order the
      speaker mentioned things; note reordering in Flags.
   5. Steps the speaker skipped over ("then it just goes to
      finance") get a row flagged OPEN: what happens in finance?
   6. After the table: list every OPEN flag as a follow-up question
      with who could answer it.
   
   Transcript:
   [PASTE]
   

3. 3.

   Verify the table against the transcript. Good output looks like this:

   Copy

   4 | Claims officer | Re-keys claim data | Excel tracker | portal
   PDF | tracker row | [11:42] "I copy it into my sheet because the
   system loses it" | WORKAROUND
   5 | Claims officer | Sends claim to assessor | Outlook | tracker
   row | email | [13:05] "then I mail it across" | OPEN: how does
   the assessor know priority? Nothing stated.
   6 | ? | Assessment happens | ? | email | decision | [13:20] "they
   come back in a few days" | OPEN: actor, system and steps unknown
   — speaker does not perform this step.
   

   Spot-check rows against timestamps — especially WORKAROUND rows (the most valuable and most sensitive) and any row with no flag whose step you don't remember hearing. A table row without real transcript backing gets deleted, not fixed.

4. 4.

   Render the swimlane from the table only. Second prompt: "Render this table as a swimlane diagram (Mermaid, one lane per actor; unknown actors get a lane labeled OPEN). Every OPEN flag appears as a visible node. Add nothing that is not in the table." The diagram inherits the table's honesty — including visible OPEN nodes where the process disappears from the speaker's view.

5. 5.

   Close the OPENs and validate. Take the follow-up list to the people who can answer — often the neighboring team the speaker hands off to; their answers extend the table, and the cycle repeats until OPENs are closed or accepted as out of scope. Then run the SME walkthrough from pb-23 on the assembled model before it is published. Handle WORKAROUND rows with care in circulation: they are process gold and personal risk — anonymize to roles and discuss with the process owner before they appear in any deck.

Make it yours

• Meeting AI + Claude as the default stack: Teams/Zoom/Fireflies transcripts drop straight into the extraction prompt; keep the prompt saved so every walkthrough produces the same table shape.
• Miro AI / Lucid for the rendering step: paste the verified table and generate the swimlane where stakeholders already collaborate; keep the table as the source of truth, the board as the view.
• Multi-session merge: for a process spanning several roles, run per-transcript tables, then a merge prompt that aligns handoffs and flags mismatches — "officer says email, assessor says queue" is a finding, not a formatting problem.
• Claude Project per process area: accumulate transcripts, tables and OPEN-resolution answers in one Project; "what do we still not know about claims intake?" gets a cited answer.

Quality check

• Every row cites the transcripttimestamp and quote, spot-checked; the table is evidence, and an uncited row is an invention.
• No OPEN was filled by inferencewhere the speaker didn't say, the table asks; sampled "complete" rows genuinely trace to stated content.
• Workarounds were kept, flagged and anonymizedthe divergence between official and lived process is preserved as data, with individuals protected in anything that circulates.
• Handoff blind spots are visiblesteps outside the speaker's view appear as OPEN lanes/nodes, not as smoothly invented process.
• The diagram adds nothing to the tablediff them; any diagram element without a table row is renderer invention and comes out.
• OPENs became follow-ups with nameseach open flag has a person who can answer it, and the model is not published as validated until pb-23's walkthrough has run.