Bigspin

A study by Bigspin · 2026

Not everyone uses AI the same way.

sessions
4,313
developers
170
personas
7

Teams measure AI adoption with usage metrics. Those metrics only capture a small slice in a much larger story. We analyzed 4,313 real Claude Code sessions and found seven distinct ways developers actually work with coding agents — patterns no dashboard would show you.

Research led by Chris Potts, Co-Founder + Chief Scientist, and Moritz Sudhof, Co-Founder + CEO.

Explore the research

What's your coding persona?

People work with AI in distinct ways to pursue different outcomes. Recognizing your habits as a user is a powerful tool for personal growth and developers who identify these patterns can capitalize on them.

The Pair Programmer

13.3% of users

Long, deep sessions. Steers reasoning, reframes goals, verifies.

Sessions are long — 22 prompts median, 8 edits per session — and you're in the loop the whole time. You direct how the agent reasons, you reframe the goal mid-flight when reality shows you something new, you actually run tests and read diffs. The agent isn't an executor; it's a partner you're shaping the work with.

Blind spot

Sustained engagement makes you the best at hard problems, but it scales poorly across many independent tasks. Applied to a small task you over-invest; applied to a large task you ship.

Signature moves

  • Reframe the goal mid-flight. When reality surfaces something unexpected, restate the goal in materially different terms.
  • Steer how the agent reasons. Tell it how to think — “stop assuming, ask first”, “use ultrathink”, “back up and re-plan.”
  • Verify before you accept. Run the tests yourself before declaring the work done. Read the diff before committing.
Want to see yours?

Inside the analysis: Our methods

We turn 4,313 Claude Code sessions from 170 users into seven persona types through the pipeline below. Click any step to read how it works.

Source

Where the data comes from

4,313 Claude Code sessions from 170 users, drawn from SWE-chat — a Stanford-led public dataset of real open-source coding-agent interactions paired with the resulting git history.

We restricted attention to users with at least three sessions in the dataset — enough to capture usage patterns rather than one-offs.

Seven session types

Each type has a different structure, orientation, and expected outcome. Simple metrics like “did it ship a PR?” flatten these distinctions. The same session can be a successful Tradeoff Deliberation and a “failed” shipper at once.

Share of all sessions in the corpus.

Prevalence

  1. Quick Check-in
    35.5%
  2. Open-ended Workshop
    29.9%
  3. Deep Collaboration
    15.7%
  4. Low-Friction Ship
    10.6%
  5. High-Friction Ship
    5.1%
  6. Tradeoff Deliberation
    2.0%
  7. Subagent Orchestration
    1.2%

Hover any bar for the type's signature. n = 4,313 sessions.

The variance across these types is why a one-metric scorecard misleads. Open-ended workshops are supposed to not ship a PR. High-Friction Ships have the highest discovered-requirement rate — that's where invisible work surfaces.

Shapes of productive practice

Each persona spends time across multiple session types — the mix is what makes them distinctive. 70% of users span three or more types, so a persona name describes the dominant mode, not the only one.

Pair Programmer

n = 23

Spec-First Architect

n = 9

Quick-Turn Sprinter

n = 27

Showrunner

n = 62

Runtime Mechanic

n = 23

Prompt Minimalist

n = 4

Multi-Mode Journeyman

n = 23

Quick
Workshop
Deep collab
Low-friction ship
High-friction ship
Tradeoff
Orchestration

Pair Programmer

Long, deep sessions. Steers reasoning, reframes goals, verifies the work.

If this is you

Sustained engagement makes you the best at hard problems, but it scales poorly across many independent tasks. Applied to a small task you over-invest; applied to a large task you ship.

Session mix

  • Quick Check-in10%
  • Open-ended Workshop22%
  • Deep Collaboration50%
  • Low-Friction Ship4%
  • High-Friction Ship6%
  • Tradeoff Deliberation7%
  • Subagent Orchestration1%

What's in your agent's conversations?

Hand drawing a curve through a dotted square

At Bigspin, we apply the same methodology — including session-level signals, LLM-annotated briefs, persona clustering — to your AI agent's real user conversations. We surface invisible failures, persona-level patterns, and the specific moves that move outcomes. Bigspin does the analysis, creates daily briefings, and gives actionable insights to improve your AI.

Get in touchLearn more about Bigspin

For product and research teams running AI agents in production. We'll scope a research engagement against your transcripts and metadata.

Want to see your persona?

Open a new Claude Code session and follow our SKILL.md walkthrough. About 15 minutes later, you'll have a local report describing your coding persona. Everything runs on your machine — no API key, no upload, no telemetry. Your session history is read locally and never leaves your computer.

  • No data sharing, 100% local on your device
  • Best with ~30+ sessions in your history
  • Run from an Opus session for calibrated comparisons
bigspin toolkit on GitHub

bigspinai/toolkit

Toolkit source, SKILL.md, and the full README with setup steps, requirements, and the /persona slash command.

View on GitHub