disclose-project

You interview the user to disclose how Claude was used across a whole project. User's answers are the source; evidence below only jogs memory and flags contradictions. Same 0–10 scale and seven roles as the disclose skill (/disclose).

Evidence

root="$(git rev-parse --show-toplevel 2>/dev/null)" || { echo "(not a git repo — interview only)"; exit 0; }
echo "author: $(git config user.name 2>/dev/null || echo '(unset — ask)')"
if git check-ignore -q "$root/.ai/disclosures.jsonl"; then echo "ledger: enabled"; else echo "ledger: disabled (.ai/ not gitignored)"; fi
echo "HEAD: $(git rev-parse --short HEAD 2>/dev/null || echo unknown)"
echo "commits: $(git rev-list --count HEAD 2>/dev/null) total, $(git log -i --grep='Co-Authored-By: Claude' --format=%H 2>/dev/null | wc -l | tr -d ' ') with a Claude co-author trailer"
echo "trailer models:"; git log -i --grep='Co-Authored-By: Claude' --format=%B 2>/dev/null | grep -io 'Co-Authored-By: Claude[^<]*' | sort | uniq -c | sort -rn | head -5
echo "files touched in Claude-trailer commits, by top-level dir:"
git log -i --grep='Co-Authored-By: Claude' --name-only --format= 2>/dev/null | grep . | awk -F/ '{print (NF>1?$1"/":$1)}' | sort | uniq -c | sort -rn | head -12
echo "disclose ledger:"
python3 - "$root/.ai/disclosures.jsonl" <<'PY' 2>/dev/null || echo "  (none)"
import json, sys
rows = [json.loads(l) for l in open(sys.argv[1]) if l.strip()]
rows = [r for r in rows if r.get("scope") != "project"]
print(f"  {len(rows)} entries")
for r in rows[-15:]:
    print(f"  {r.get('date')} | {r.get('artifact')} | {', '.join(r.get('roles', []))} | SOI {r.get('user_soi')}")
PY

Counts returned are used to prompt questions, never shares: a trailer means Claude touched a commit, not that it wrote it, and chat-pasted work leaves no trailer.

Anti-Rationalization (hard gate)

Ask each phase one at a time, waiting for each answer. Do not write a table or statement until every phase has an answer (a rating or "n/a").

Interview

Phases, in order:

For each phase, show the evidence that speaks to it (matching dirs, ledger entries), then ask:

Phase <!-- n -->: <!-- name -->
Evidence: <!-- relevant trailer counts / ledger entries, or "none in repo" -->

1. What did Claude do here? (or "n/a")
2. How did you check or change its output?
3. Your SOI for this phase (0–10)?

Don't propose a score — a suggested number biases the user's rating. After the answer, flag a clear conflict with the evidence in one line (e.g. 40 trailer commits under analysis/ but rated 1) and accept the user's resolution.

Last turn: show the filled table and ask for the overall SOI score and the model name(s) and versions to cite (suggest the trailer models above).

Output

| Phase | Claude's role | Human verification | SOI |
|---|---|---|---|
| <!-- phase --> | <!-- roles, from the user's answer --> | <!-- how they checked it --> | <!-- n or n/a --> |

AI use disclosure: <!-- Author --> designed and conducted this <!-- project noun -->.
Claude (<!-- model(s), versions -->; Anthropic) was used for <!-- roles by phase, e.g.
"code development in data processing and analysis, and line editing of the manuscript" -->.
<!-- Author --> <!-- verification, e.g. "reviewed and tested all code and edited all text" -->
and takes full responsibility for the content (overall spectrum of involvement: <!-- n -->).

Use the user's words for roles and verification; offer tightened wording, don't embellish. Apply the disclose drift check: flag wording that implies more or less involvement than the ratings. Then offer to insert into the relevant manuscript file — default to the end, but ask where. Insert only on a yes.

Ledger

Unless --no-ledger or the evidence block says disabled, append one line to <repo>/.ai/disclosures.jsonl (mkdir -p .ai/; never edit .gitignore):

{"scope": "project", "date": "YYYY-MM-DD", "project": "<label>", "phases": [{"phase": "<name>", "roles": ["..."], "verification": "<...>", "soi": 3, "user_reply": "<verbatim>"}], "overall_soi": 3, "sha": "<short HEAD>", "statement": "<text>"}

user_reply quotes each phase answer verbatim. Best-effort: a failed write is mentioned, never blocks.

Verification

• Each of the 7 phases asked one at a time, with an answer recorded before the next
• No SOI (overall or for individual phases) proposed or assumed by you; conflicts with evidence flagged, user's resolution kept
• Model names and versions confirmed by the user, not copied from trailers unasked
• Table and statement use the user's words; drift check run
• Insertion offered; ledger written with verbatim replies, or skipped and explained