Weight: 15% of overall score · How the overall score is calculated
Definition
Complexity Progression measures whether the difficulty of tasks you delegate to Claude is increasing over time. An engineer who only ever delegates simple edits and one-liners is not growing their AI collaboration — they have found a comfort zone and stayed in it. This dimension rewards those who push the ceiling of what they trust Claude to handle.
Progression is not about volume. It is about trajectory.
How it's measured
Each session is assigned a complexity score from 1 to 5 by an LLM reviewer:
| Score | Description |
|---|---|
| 1 | Typo fix, trivial edit, single-line change |
| 2 | Small isolated change, simple bug fix |
| 3 | Standard feature, moderate refactor, multi-file change |
| 4 | Complex feature, significant debugging, cross-system work |
| 5 | Expert-level architecture, novel system design, deep debugging |
Sessions are sorted chronologically and a linear regression slope is computed over the complexity scores:
slope = (n·Σxy − Σx·Σy) / (n·Σx² − (Σx)²)
Where x = session index (0, 1, 2…) and y = complexity score (1–5). A positive slope means complexity is rising over time. The Anthropic cohort benchmark slope — moving from avg 3.2 to 3.8 over ~25 sessions — corresponds to approximately 0.024 per session (score 8).
The slope is mapped to a 1–10 dimension score:
| Slope | Score |
|---|---|
| ≥ 0.05 | 10 |
| ≥ 0.024 | 8 |
| ≥ 0.01 | 7 |
| ≥ 0 (flat) | 5 |
| ≥ −0.01 | 4 |
| ≥ −0.024 | 3 |
| < −0.024 | 1 |
At least 2 classified sessions are required. Fewer than 2 returns a neutral score of 5.
What high vs low looks like
High (score 8–10)
- Early sessions: debugging, small refactors, isolated scripts
- Later sessions: architecture discussions, multi-service features, system-level changes
- Visible upward trend in session complexity over weeks or months
- Willingness to delegate work that feels uncomfortable or unfamiliar
Low (score 1–4)
- Complexity is flat across all sessions — always at level 2 or 3
- Or complexity is declining — reverting to simpler tasks over time
- Treating Claude as a productivity tool for existing easy tasks rather than a capability expander
Behavioural patterns in real sessions
The Anthropic work study tracked average task complexity across its engineering cohort from February to August 2025. The cohort average rose from 3.2 to 3.8 on the 1–5 scale — a 19% increase in six months.
More telling was the shift in task type distribution. Feature implementation grew from 14.3% to 36.9% of all sessions. Design and planning sessions — the highest-complexity category — grew from 1.0% to 9.9%. Engineers were not just doing more; they were delegating progressively harder work.
This was not uniform across teams. The pre-training team showed 54.6% of sessions classified as feature building — the highest of any team. Security team sessions were 48.9% code understanding — a pattern reflecting domain-specific needs rather than lower maturity.
The study also found that over half of surveyed engineers reported they could "fully delegate" only 0–20% of their work at the start of the study period. Complexity Progression is the dimension that tracks whether that ceiling is being pushed upward over time.
How it affects your overall score
Complexity Progression carries 15% of your total score.
A one-point improvement in this dimension adds 0.15 points to your overall score.
This dimension responds most slowly to habit changes — it requires sustained behavior change across multiple sessions over time. It is also the most forward-looking dimension: a rising slope now predicts higher scores across multiple dimensions in future analyses.
It interacts with Task Breadth (broader task types tend to include higher-complexity work) and Delegation Intelligence (complex tasks must still be appropriate delegation targets).