Track Three Numbers

Complete The Codify-Loop first; this lesson builds on the loop so the three numbers you track reflect actual codification cycles, not sporadic edits.

You have been using AI for two months. Sometimes you accept the output. Sometimes you rewrite half of it. You can't tell if it's getting better or you're just used to it. Three numbers fix this.

Here is the before and after: Acceptance rate is the percentage of AI outputs you accept without rewriting. Measure it by counting: across your last 10 sessions, how often did you take the output as-is versus edit it substantially or discard it? A baseline acceptance rate of 30-40% is common in the first month. After 3 months of consistent codification, 60-70% is achievable in your highest-frequency task types.

Iteration count is the number of exchanges before you get usable output on a typical task. One well-structured task with a good rules file should resolve in 1-3 exchanges. If you are consistently at 6-8, either the task is genuinely complex or the rules file is not doing its job.

Codification rate is the fraction of sessions that produce at least one new rule. It should be between 20-40%. Higher than 50% means you are in early habit formation, which is fine for the first month. If it stays above 50% after month two, your rules are likely too granular and not compounding.

If acceptance rate is not trending up after 10 sessions, you have an accumulation problem: rules are going in but model behavior is not improving. The usual cause is vague rules, conflicting rules, or a rules file that has grown too large to be loaded cleanly.

Paste this into Claude:

I want to measure my compounding baseline. Here is a summary of my last 5-10 AI-assisted sessions: [describe roughly what tasks you worked on and how the sessions went. Include any sessions where you rewrote output, any where you accepted it quickly, and whether you added rules after any of them]. Calculate my approximate: (1) acceptance rate as a percentage, (2) average iteration count per task, and (3) codification rate as a percentage of sessions that produced a new rule. Then tell me what each number suggests about whether my sessions are compounding or just accumulating.

What good looks like:

• You have a specific percentage for acceptance rate, even if approximate
• You have an average iteration count for your most common task type
• You have a codification rate number, or you identified that you have been adding zero rules and that is the problem
• You understand which of the three numbers is your biggest lever right now
• You set a target for each number over the next 4 weeks

What a good response looks like:

Based on your session summaries, here are your three numbers.

Acceptance rate: approximately 35%. Across 8 sessions, you described accepting output directly in 3 sessions (the email drafts, the meeting summary, and the data cleanup). In the other 5 you rewrote substantially or discarded and restarted. 3/8 = 37.5%, call it 35%.

Average iteration count: 5.2 exchanges per task. Your highest-friction task type is code review summaries, where you described going back and forth 7-8 times. Your lowest is email drafts at 2-3 exchanges. The gap tells you something: your rules file is probably optimized for email-style tasks and weak on technical summarization.

Codification rate: 12.5%. You added a rule in exactly 1 of your 8 sessions. That is your biggest lever. At 35% acceptance rate after 2 months, you should be at 60-70% by month 3 if codification is consistent. At 12.5% codification rate, you are accumulating sessions without compounding them.

Targets for the next 4 weeks: acceptance rate from 35% to 50%, iteration count from 5.2 to 3.5, codification rate from 12.5% to 30% (at least 1 rule per 3 sessions). The codification target unlocks the other two.

Paste this into Claude:

I want to track my three numbers over a longer window. For the next 4 weeks, I will log each session with these fields: date, task type, outputs accepted vs rewritten, exchanges to completion, and whether I added a rule. Here is my log after [N] weeks: [paste or describe your log entries]. Calculate my acceptance rate, average iteration count, and codification rate week by week. Show me the trend. Tell me whether any number is moving in the wrong direction and what the most likely cause is.

What good looks like:

• You logged at least 4 weeks of sessions in a consistent format
• You have week-by-week numbers for all three metrics, not just a single average
• You identified at least one metric that is trending in the wrong direction
• You diagnosed the most likely cause of any declining metric (vague rules, conflicting rules, no codification, wrong task type mix)

What a good response looks like:

Week-by-week analysis of your 4-week log.

| Week | Acceptance Rate | Avg Iterations | Codification Rate |
|------|----------------|----------------|-------------------|
| 1    | 32%            | 5.8            | 20%               |
| 2    | 41%            | 4.2            | 40%               |
| 3    | 38%            | 4.9            | 60%               |
| 4    | 36%            | 5.1            | 65%               |

Weeks 1-2: healthy trajectory. Acceptance rate up 9 points, iteration count down 1.6 exchanges, codification rate rising.

Week 3: codification rate spiked to 60% and acceptance rate dipped 3 points. This pattern is the early warning: you started writing rules faster than they could improve output quality. When codification rate exceeds 50%, you are usually adding rules that are too narrow or that conflict with existing ones.

Week 4: the problem compounded. Acceptance rate dropped again to 36%, iterations climbed back to 5.1. Codification rate at 65% is unsustainable.

Diagnosis: rule bloat. You added 8-10 new rules in weeks 3-4. Likely some conflict with earlier rules, and the file may now exceed what loads cleanly in context. Recommended action before week 5: run the pruning audit from L6-03. Delete or consolidate any rules you cannot trace to a specific session failure. Target: remove at least 30% of rules added in weeks 3-4.