---
name: look-back-effort-accounting
description: >
  Weigh completed agent-assisted software work after it ships. Produces an
  anonymized human-effort-equivalent activity measure, cleanup credit, attended
  wall-clock, agent labor when available, and calibration rows for later
  forecasts. Use for retrospective accounting, not prediction or valuation.
---

# Look-Back Effort Accounting (public draft)

Measure a completed git range or finished run. This is a retrospective
activity-volume model: it estimates how much implementation work the delivered
change represents under a consistent set of coefficients.

It does **not** measure business value, code quality, developer worth, dollars
saved, or what a team “should” have taken. Lines changed are a size proxy, not
an outcome.

## Credit and adaptation

This draft descends from
[Todd Saunders’s public `/cost-estimate` prompt](https://x.com/toddsaunders/status/2029594318361571497):
lines of code plus complexity translated into senior-developer hours. I first
used it publicly on March 5, 2026 with both hours and dollar
return-on-investment framing.

This version changes the purpose and mechanics:

- retrospective activity accounting rather than replacement-cost valuation;
- hours only, with no dollar or return-on-investment claim;
- a small five-tier file model;
- deletions reported as separate cleanup credit;
- attended wall-clock kept separate from summed agent labor;
- anonymized ledger rows feeding a separate forward-estimation draft.

The current fork retires the dollar layer and keeps the hours-only accounting
skeleton. Julien Barbier’s later open-source
[`/cost-estimate`](https://github.com/jbarbier/claude-code-cost-estimate) is a
parallel expansion of Todd’s idea, not the direct parent of this fork. The
adaptations and any mistakes here are mine.

## Inputs

Collect:

1. a closed git range;
2. per-file insertions and deletions with rename detection;
3. session start and verified finish;
4. summed worker durations, if the harness exposes them;
5. completed task count and task types;
6. exclusions for generated, vendored, lock, snapshot, and minified files.

Do not publish repository names, branches, home paths, emails, account names,
customer data, or raw agent transcripts.

## Real coefficients used by this draft

These are the current coefficients from one personal workflow. They are knobs,
not validated industry constants.

| Tier | Category | Examples | Human-equivalent rate |
|------|----------|----------|----------------------:|
| T1 | Trivial | Config, boilerplate, CI configuration | 60 lines/hour |
| T2 | Low | Simple UI, tests, docs, routine scripts | 35 lines/hour |
| T3 | Medium | Business logic, data access, API integration | 22 lines/hour |
| T4 | High | Architecture, migrations, state machines | 14 lines/hour |
| T5 | Specialist | Parsers, compilers, systems, cryptography, GPU | 10 lines/hour |

Deletion-only cleanup is credited separately at **120 deleted lines/hour**.

For the companion forward forecast, the current anonymized trace calibration
uses an overall human-equivalent → attended-agent coefficient of **43.3×** for
the window **2026-06-17 through 2026-07-14**. Do not apply that coefficient to
someone else’s workflow without collecting local actuals.

## File classification

Use the file’s responsibility, not merely its extension.

| Signal | Starting tier |
|--------|---------------|
| Configuration, manifests, simple generated-adjacent edits | T1 |
| Tests, documentation, simple views, routine scripts | T2 |
| Services, repositories, controllers, transformations | T3 |
| Schemas, migrations, protocol boundaries, state machines | T4 |
| Parsers, compilers, system extensions, cryptography, GPU code | T5 |

When uncertain, round up one tier and record the judgment.

## Calculation

For each included file:

```text
creation_h = inserted_lines / tier_rate
```

Across the range:

```text
human_equivalent_h = Σ creation_h
cleanup_credit_h = total_included_deletions / 120
attended_wall_h = verified_finish - attended_start
agent_labor_h = Σ worker durations  # only when measured
```

Keep all four values separate. In particular, do not blend cleanup credit into
the insertion-derived headline and do not call summed parallel labor
“wall-clock.”

Optional descriptive ratios:

```text
attended_multiplier = human_equivalent_h / attended_wall_h
labor_multiplier = human_equivalent_h / agent_labor_h
```

These ratios are diagnostics for the observed run. They are not general claims
about agent productivity.

## Required output

```markdown
## Look-back: [anonymized run description]

Range: [start]..[end]
Tasks completed: N
Commits: N | +X inserted | -Y deleted | Z files

| Tier | Files | Insertions | Rate | Human-eq h |
|------|------:|-----------:|-----:|-----------:|
| T1 | ... | ... | 60 | ... |
| ... | ... | ... | ... | ... |
| **Total** | **N** | **X** | | **H.H** |

Cleanup credit: D / 120 = C.C h (reported separately)
Attended wall-clock: W.W h (measured | reconstructed | assumed)
Agent labor: A.A h (measured | unavailable)

Interpretation: activity-volume proxy only; not value, quality, savings, or a
universal speed claim.
```

## Ledger row

Append an anonymized row to a private ledger:

```json
{
  "id": "anon-001",
  "work_type": "chore",
  "tasks_completed": 4,
  "insertions": 1100,
  "deletions": 40,
  "human_equivalent_h": 36.2,
  "cleanup_credit_h": 0.33,
  "attended_wall_h": 0.42,
  "agent_labor_h": 0.90,
  "walltime_source": "measured",
  "rates_version": "personal-v1"
}
```

The public artifact may contain rounded aggregates. The private ledger should
retain enough provenance to recompute the result.

## Guardrails

- Exclude generated and vendored material, but retain an auditable exclusion
  count.
- Use rename-aware git statistics so moves are not counted twice.
- Do not infer active human attention from the first and last commit alone.
- Record failed attempts and rework when the trace supports it.
- Treat zero-line work items as missing size data, not zero effort.
- Label reconstructed or assumed clocks; do not present them as measured.
- Keep forecasting in the companion forward skill so hindsight does not leak
  into a supposedly prospective estimate.
- Revisit the coefficients as the model, harness, repository, and working style
  change.

## Closing the loop

Compare the completed result with the forecast recorded before execution:

```text
schedule_error = actual_attended_wall_h / forecast_schedule_wall_h
labor_error = actual_agent_labor_h / forecast_agent_labor_h
human_eq_error = actual_human_equivalent_h / forecast_human_equivalent_h
```

Keep the misses. A calibration that remembers only favorable runs is a story,
not an estimate.

---

*Portable draft · real coefficients from one anonymized personal workflow ·
retrospective activity accounting, not valuation*