kadns What Claude Did for Me
Turns a period of Claude usage into a personal impact report: what the person worked on, the job each piece of work was doing, estimated time saved, and what that time is worth. Output is a self-contained HTML file they can open, adjust, share, or print to PDF.
The person should never need admin access, a Console login, or an export file to get a useful report. Those are optional upgrades, not prerequisites.
Pick a mode
Try these in order. Do not ask the person to choose, just start with harvest and fall back silently if it is unavailable.
- Harvest (best): read the person's own recent Claude conversations to see what they actually worked on.
- Interview (always available): a short guided conversation, no data needed.
- Import (optional): a usage CSV from Claude Console, OpenAI, Azure, or Copilot, for anyone who has one.
Mode 1: Harvest
Use recent_chats to pull the person's conversations across the reporting
window, and conversation_search if you need to fill in specific topics. Ask
first which period to cover, defaulting to the last 30 days.
Paginate with recent_chats using the before parameter to walk back through
the window, up to about 5 calls. If the window is not fully covered, say so
plainly rather than presenting a partial picture as complete.
If these tools are unavailable or return nothing, do not announce a failure. Just move to interview mode: "Let me ask you a few quick questions instead."
From the harvested conversations, cluster the work into 5 to 10 recurring workflows. Give each a plain-language name the person would recognize. Then count sessions per workflow across the window. Show the person your clustering before continuing and invite corrections, they know their work better than the transcript does. This confirmation step matters, do not skip it.
For each workflow, you still need one thing the transcripts cannot tell you: what the work actually produced, in one sentence. Ask for the ones that are not obvious from the conversation content.
Mode 2: Interview
Ask, in this order, keeping it brief:
- What period should this cover? (default: last 30 days)
- What do you use Claude for? Prompt for 4 to 8 recurring things. If they are stuck, offer categories: drafting and editing, research and summarizing, code, meeting notes, spreadsheet or data work, client communication, brainstorming.
- For each one: roughly how often, and in one sentence, what does it produce?
- What is your hourly rate, or a rough estimate of what your time is worth? If they decline, skip all dollar figures and build the report on hours alone.
- What do you pay for Claude, if anything? (Pro is $20/month, Max 5x is $100, Max 20x is $200, or an employer-provided seat.)
Convert stated frequency into a session count for the window: daily is about 20 sessions a month, a few times a week is about 12, weekly is about 4, occasional is about 2. Say that you are estimating rather than presenting these as counted.
Mode 3: Import
If the person has a usage export, run:
python scripts/analyze_usage.py <path-to-csv> --out workflow_analysis.csv
See references/export-guide.md for how to pull one, and note that most export
paths require an admin, billing, or developer role. Use the real session counts
and costs from the export in place of estimates, then continue with mapping
below exactly as in the other modes.
Map every workflow (all modes)
Read references/jtbd-taxonomy.md and assign each workflow four things:
- Business Process: the function it belongs to, in the person's own organizational language where they offer it
- Job to be Done: what the work was hired to accomplish, phrased as an outcome starting with a verb, never as a tool
- Value Pillar: one of Revenue Growth, Cost Reduction, Risk Mitigation, or Transformation
- Category: one of
analysis,document,email,meeting,communication,code,data,other, which sets the time-saved band
Propose your mapping and let the person correct it. Push gently on pillars: most people default everything to Cost Reduction, but review work that catches problems is Risk Mitigation, and work they could not do at all before is Transformation. Those two are usually undercounted and carry the most weight in a budget conversation.
If a workflow has no clear outcome, keep it in the report with the person's actual words, including "not sure." An honest receipt includes the parts nobody can explain yet, that gap is the most useful finding in the whole exercise.
Build the report
Write the payload to JSON and render:
python scripts/build_report.py --data payload.json --out what-claude-did-for-me.html
Payload shape is documented at the top of build_report.py, and
assets/sample_payload.json is a complete working example. Required per
workflow: name, category, sessions. Strongly recommended: process,
jtbd, pillar, frequency, outcome.
Save the report where the person can download it and present it to them. The report includes a live hourly-rate control that recalculates every dollar figure, and a Save as PDF button.
After the report
Offer one closing line they could actually use in a renewal or review conversation, built from their own words rather than generic phrasing. Point out the single largest workflow with no clear outcome, if there is one, as the thing worth watching next period.
Notes
- Time saved is an estimate of expert-equivalent time, not a measurement, and the report says so in its methodology section. If the person starts treating the figures as audited, restate that plainly.
- The category bands in the taxonomy reference are directional defaults, not cited research. If the person's experience differs, adjust the band and rerun, that is a feature, not a workaround.
- Keep the tone collaborative throughout. This helps someone build a case for their own work, it is not an audit of them.