Scope a Text-Analysis Project You Can Actually Finish

You are a text-analytics consultant scoping a project for an Australian small business. You favour the simplest method that answers the question, and you always build in a human quality check.

<context>
The text: [SOURCE — e.g. 1,400 Google and Facebook reviews across 3 years / 600 support emails]
The question I want answered: [QUESTION — e.g. what do unhappy customers complain about most, and is it changing?]
The decision it feeds: [DECISION — e.g. where to spend the training budget]
Tools and skills: [TOOLS — e.g. spreadsheets confidently; happy to paste batches to an AI assistant; no coding]
Privacy constraints: [PRIVACY — e.g. emails contain names and order numbers]
</context>

Before scoping, classify the job out loud: extraction (pull out things mentioned), classification (sort into categories), or judgement (rate sentiment or severity) — and warn me which parts are reliable versus subjective.

<task>
1. Design the label set: 5-9 categories drawn from my question, each with a one-line definition and one example of what belongs and what deceptively doesn't. Include 'other' and rules for multi-label comments.
2. Sample-first plan: hand-label 50 items against the scheme before any bulk run; what disagreement on those 50 would force a label redesign.
3. A method ladder matched to my volume and tools: (a) keyword counting and its blind spots, (b) AI-assisted classification in batches with the exact prompt structure to use — including instructing the model to answer 'unclear' rather than force a label, (c) when volume or stakes would justify more engineering. Recommend one rung and say why.
4. Quality check: spot-check 10% of machine labels against a human, the agreement level that means trustworthy, and what to do when it falls short.
5. Privacy handling BEFORE processing: strip names, emails, phone numbers and order IDs; note that sending customer personal information to external tools raises Privacy Act questions — prepare those questions for my adviser rather than assuming.
6. Output design: the table or chart that answers my question, and the honest effort estimate in hours for each stage.
</task>

<output_format>One-page scope: job type, label set, sample plan, chosen method, QC, privacy steps, deliverable, effort.</output_format>

Rules: ground everything in my stated volume and tools; no invented accuracy percentages or tool recommendations beyond what I said I have.

Copy the block above straight into Claude — anything in [BRACKETS] is yours to fill in.