Throughout my career, I have seen people look busy without producing corresponding value. I have also seen engineers adjust their behavior after learning their organization tracked contribution counts. The visible metric became something to optimize.
Counts can rise without the work becoming more valuable. Productivity depends on the value created or preserved relative to the effort spent.
Activity is still useful context. An unexpected drop can prompt a conversation about changing responsibilities, blockers, or work the count misses. It gives me a reason to ask, but no explanation of its own.
A contribution graph is an incomplete picture
GitHub’s contribution graph includes eligible commits, pull requests, reviews, issues, and other actions, subject to GitHub’s contribution criteria. The result can look reassuringly busy without showing whether the work helped anyone.
Investigating a failure may reveal that no code change is needed. Design, mentoring, incident response, and unblocking colleagues can create value with few commits. A careful review may prevent a bad change; the review count says nothing about what it prevented.
I released my contribution-history generator to show how little those counts tell us on their own. It creates dated commits containing generated JSON records, with configurable activity patterns. The commits are real; they do not represent software delivered on those dates. AI makes the distinction more important: generating more code does not establish that anything useful has been built.
Consider a hypothetical engineer adding unnecessary commits or splitting a pull request solely to increase the count. Small pull requests can improve review and delivery. The question is whether the split helps the work or merely helps the number.
Measuring work can change the work
People may change their behavior when they know it is being observed, even before a metric becomes a target. This is related to the Hawthorne effect, although the evidence does not establish a universal or predictable response. A dashboard can influence the work it describes.
Making the number a target adds another incentive. Goodhart’s law describes the risk of optimizing a measure at the expense of what it was meant to represent.
That can mean deliberate gaming, or simply choosing the more visible of two worthwhile tasks. If counted activity earns recognition while investigation and mentoring barely register, the visible work becomes easier to justify.
AI tokens bring the same temptation
Token consumption measures AI usage. It cannot, by itself, tell us what the work accomplished or how much human effort it required.
Rewarding token volume could encourage unnecessary requests, repeated generations, oversized context, or redundant agent work. Consumption could rise while the result stayed the same.
High usage may support a difficult investigation or extensive verification. Low usage may reflect an efficient solution. Comparisons need context about the models, tasks, and workflows.
The cost includes the human time spent checking, correcting, and maintaining the result. A quick generation followed by hours of rework may save little.
Usage data can help manage costs, plan capacity, and understand adoption. Low adoption may prompt a discussion about access, training, or suitability. Ranking engineers by that usage tells us little about their impact.
What I would look at instead
I find it useful to separate three things:
- Activity: the actions involved in doing the work, such as commits, reviews, or AI requests.
- Output: what the work produces, such as a feature, a migration, or a repair.
- Outcome: the impact on the people, business, or systems the work is meant to help.
Consider a hypothetical account setup change. Opening pull requests is activity; shipping a new screen is output. The outcome question is whether people complete setup more successfully or need less help.
Before building, I would agree with Product on the customer or business improvement worth the effort. We would record the current completion rate and support requests per person attempting setup, set a review date, and decide what evidence would justify keeping, revising, or rolling back the change.
After release, I would compare the results with that baseline and check customer feedback, new errors, and changes in who is trying the product. Fewer total support requests mean little if fewer people attempted setup. If the evidence is inconclusive, agree on what to learn next and when to revisit the decision.
Replacing commit counts with feature counts repeats the mistake. Adoption and other customer or business measures can also become distorted targets; no single number removes the need for judgment.
DORA’s guidance on software delivery performance considers delivery speed alongside failed changes and unplanned rework. It also warns against turning metrics into targets or comparing dissimilar services. These measures describe delivery performance; I still need evidence that a change was worth the effort.
Useful outcomes include better reliability, performance, maintainability, and less operational burden. These gains can take time to appear, and several teams or external factors may contribute. Maintenance may prevent deterioration without moving a product measure upward. A test that reproduces a failure before a fix and passes afterward provides evidence that the specific defect was addressed. The absence of incidents, by itself, does not establish what a change prevented.
Turn evidence into decisions
The harder decision is what to prioritize. Suppose recurring failures in that setup flow compete with a promised feature. If customers cannot get started, I would argue for a targeted repair. The discussion with Product and engineering leads should cover customer harm, operational burden, and the cost of delaying the feature. The agreement should name who owns the work, what moves in the plan, and when to review the decision. Choosing the feature instead means being explicit about the risk being accepted.
I would assess shared outcomes at team level, then look at each person’s responsibilities and contribution. That means examining decisions, reviews, and help that unblocked others. Precise individual scores can imply certainty that shared results do not support.
A product change that misses its goal does not automatically mean someone performed poorly. Accountability needs to consider the responsibilities within their control: testing assumptions, raising blockers, and following through on agreed work. When those expectations are missed, I would clarify the gap, address blockers, agree on support, and set a review date. If the gap persists, I would use a focused improvement plan with clear expectations and consequences. Context should lead to a fair decision, not an indefinite delay.
For me, the review comes back to three questions:
- What problem was this intended to solve, and for whom?
- What evidence would show it helped, and when can we check?
- What did it cost the team to deliver and maintain it?
A reason to ask, not a verdict
A dashboard should start a conversation. The manager still has to understand the work, set fair expectations, and act on the evidence. The question remains: what value did the work create or preserve, and at what cost?