| name | attention-ledger |
| description | Use this to account for where AI-assisted attention actually went and what it bought - including the uncomfortable parts. |
Attention Ledger
Work Classes
Classify episodes approximately into:
PRODUCT VALUE / MAINTENANCE / LEARNING / RESEARCH / INFRASTRUCTURE /
AGENT-META WORK / ADMINISTRATION / REWORK
Estimate two shares per class: share of sessions and share of downstream
outcomes. The gaps between those two columns are the finding.
Name work that FEELS high-leverage but historically produced little
downstream benefit. Name low-visibility work with unusually high leverage.
This section should be evidence-driven and possibly uncomfortable.
Explore vs Exploit
EXPLORE: alternatives, technologies, architectures, models, strategies.
EXPLOIT: commit, implement, ship, polish, scale.
Measure mode switching. Hunt for:
- premature commitment before options were generated
- excessive exploration with no commitment
- architecture churn, tooling churn, model churn
- repeated redesigns of the same subsystem
- unfinished promising work
Do not call exploration waste unless outcome evidence supports it.
Exploration is R&D until proven otherwise.
Risk vs Deliberation
Compare deliberation depth against reversibility and downside:
| Decision type | Observed deliberation | Appropriate? |
|---|
| cheap + reversible | hours of comparison? | over-deliberation |
| expensive + irreversible | minutes? | under-deliberation |
Find both rows in the real data.
Metrics Discipline
Never equate messages, code volume, commits, tokens, or session length
with productivity. Use completions, regressions, correction loops, rework,
deployment success, later reversals, maintenance burden. State cycle-time
uncertainty honestly.
One-Line Memory
Attention follows interest. Audit whether value followed.
