curate-transcript

You are curating a Defra LAP stakeholder interview transcript. Your job is to remove off-topic passages. All kept text is preserved mechanically — you never write or rewrite any transcript content.

Input

The transcript file path is: $ARGUMENTS

Steps

1. Derive the output path by taking the filename, replacing the .txt extension with _curated.txt, and placing it under output/transcripts/. For example:

   • transcripts/interview-1.txt → output/transcripts/interview-1_curated.txt
   • transcripts/deep-dive.txt → output/transcripts/deep-dive_curated.txt

2. Ensure the output directory exists by running mkdir -p output/transcripts.

3. Copy the original file to the output path using cp. This creates an exact mechanical copy that preserves all text verbatim.

4. Read the output file using the Read tool.

5. Identify passages to remove. Classify content into two categories:

   Keep — content directly about the application or its domain:

   • Application walkthrough content: screen descriptions, user flows, functionality explanations, how the system behaves
   • Domain knowledge: business rules, terminology, processes, regulations that the application implements
   • Technical details about the application: architecture, data flows, infrastructure it runs on
   • Integrations and dependencies: third-party systems, external services, APIs, data exchanges, upstream/downstream applications the system connects to

   Remove — everything else, including but not limited to:

   • The incumbent modernisation team's agenda, plans, opinions, or approach discussions
   • Stakeholder feature requests or wishlists for new functionality
   • Project management discussions: team structure, suppliers, contracts, timelines, staffing, resourcing, delivery plans
   • Meeting logistics and scheduling: arranging follow-ups, calendar availability, session housekeeping
   • Personal context about individuals: career history, background, role introductions beyond job title
   • Technical tangents unrelated to the application itself: DPIA processes, data protection policy, procurement procedures, programme governance
   • Social pleasantries, small talk, greetings, sign-offs, thanks, and chitchat
   • Meta-discussion about the interview process itself
   • Tangential conversation not related to the current application or its domain

   When in doubt, remove — but never remove content that describes how the application behaves, what it connects to, or what domain rules it implements. These transcripts are typically recordings of sessions where third-party vendors explored legacy applications as part of a migration engagement. We are not interested in the vendors' migration plans, modernisation approach, or opinions — only the as-is state of the application they were examining. The test is: does this passage describe what the application does today, or what someone planned to do with it? Keep the former, remove the latter.

6. Remove each identified passage using the Edit tool on the output file. For each passage:

   • Set old_string to the exact text of the passage to remove (copy it precisely, including whitespace, timestamps and line breaks)
   • Set new_string to an empty string ""
   • If removing a passage leaves adjacent blank lines, make a follow-up Edit to collapse them to a single blank line

7. Return a confirmation message containing:

   • The output file path
   • A brief summary of what categories of content were removed, with approximate counts (e.g. "Removed: 3 passages discussing the incumbent team's migration timeline, 2 feature requests for a new reporting dashboard")
   • If nothing was removed, state that the transcript contained no off-topic material