AI Disclosure

AI Disclosure Tracking

Track Claude's contributions during coding sessions and maintain a disclosure file for PR transparency.

Involvement Levels

1. Autonomous - Claude wrote the code/solution independently
2. Assisted - Claude implemented based on user direction
3. Advised - Claude provided guidance that user implemented

File Location

Disclosure file: .claude/disclosures/<branch-name>.md

Get branch name with git branch --show-current. If not in a git repo, use session-<date>.

Workflow

1. Opt-in

When starting work on a feature branch, offer once:

"Would you like me to track AI contributions for this branch?"

If yes, create the disclosure file and begin tracking.

2. Track Silently

Only track contributions made AFTER the user opts in. Do not backfill anything from earlier in the conversation, even if it seems relevant to the branch. The disclosure file starts as a clean slate from the moment the user says yes.

After significant actions (writing functions, fixing bugs, refactoring):

1. Log the contribution - Append to the appropriate section in the disclosure file
2. Record what you changed - Track in the internal section:
   • Which files you touched
   • Brief summary of what you wrote/changed
   • Initial involvement level

Only use the disclosure file to track contributions — do not rely on in-context memory of what happened before opt-in.

No prompts during work—just silently maintain the record.

3. Verify and Generate Summary

When user requests summary or creates a PR:

1. Check your work against current state:

   • Review files you logged as touching
   • Compare current file contents against what you originally wrote
   • Use git diff or read the files to see if user modified them after you
   • Also recall from conversation context: did user correct you? Ask for changes? Rewrite parts?

2. Downgrade if needed:

   • If user significantly modified your code afterward → downgrade to Assisted
   • If user corrected your approach multiple times → downgrade to Assisted
   • Add note: "co-creation with significant user involvement"

3. Generate the summary with accurate involvement levels

Disclosure File Format

# AI Disclosure for branch: <branch-name>

## Summary
[Generated on request]

## Contributions

### Autonomous
- [One-line descriptions of independent work]

### Assisted
- [One-line descriptions of directed work]

### Advised
- [One-line descriptions of guidance provided]

Internal Tracking

Track your changes in an HTML comment (not shown in final summary):

<!--
CHANGES:
- src/partition.py: wrote repartitioning logic (autonomous)
- tests/test_partition.py: wrote validation tests (autonomous)
- src/boundaries.py: implemented boundary calc (autonomous)

CORRECTIONS:
- src/boundaries.py: user fixed off-by-one error (count: 2)
-->

This record lets you verify at summary time whether files still contain what you wrote, or if the user significantly changed them.

Downgrade rule: If user significantly modified your code or corrected your approach repeatedly, downgrade from Autonomous to Assisted and note: "co-creation with significant user involvement".

Example Output

# AI Disclosure for branch: feature/healpix-partitioning

## Summary
Claude assisted with repartitioning logic (co-creation with significant human involvement), autonomously wrote test cases, and advised on spatial indexing approaches.

## Contributions

### Autonomous
- Wrote test cases for HEALPix partition validation

### Assisted
- Implemented repartitioning logic based on user requirements

### Advised
- Suggested using spatial indexing for performance

PR Format

When user creates a PR, offer a copy-paste block:

## AI Disclosure

Developed with Claude assistance:
- **Autonomous**: [list]
- **Assisted**: [list]
- **Advised**: [list]

Details: `.claude/disclosures/<branch>.md`

Source: shap/shap — distributed by TomeVault.