refactor(verify): narrow trigger to claim+evidence; drop read-only mode

[bensonwong]

Summary

• Skill was firing too broadly — description and §1 Orient included OCR, extract, summarize, read, parse as triggers, so /verify ran even when the user only wanted file contents
• Dropped read-only mode from §1 — skill now requires both a claim and an evidence source; plain document reading uses prepare directly and answers normally
• Removed Tool precedence section ("ALWAYS use this skill for PDFs…") — overclaim replaced with a single clear note: use prepare for reading, /verify for citing
• Dropped §2 Read-only fast path (~13 lines) — no longer a concern of this skill
• Triage table 8→6 rows — removed read-only row and the self-correction loop ("you prepared the claims file as evidence")
• Collapsed Format 2 k≠claimText subsection (8 lines) into a 2-line gotcha note — concept still enforced by field table and STOP AND CHECK
• Updated frontmatter description to verify-only scope
• Net: 309 → 271 lines (−38)

Test plan

• Prompt with only a PDF and "summarize this" — skill should NOT fire; model uses prepare and answers directly
• Prompt with a claim + evidence doc — skill fires, full §1→§4 pipeline runs
• /verify in prompt — skill fires regardless
• HTML embed case — §2 triage routes to HTML annotation path correctly
• Existing verified HTML re-run — triage routes to verify --html correctly

[claude]

Review: refactor(verify): narrow trigger; drop read-only mode

Overall the refactor is sound. Narrowing the trigger is correct — the old description fired the skill on plain reads, burning steps unnecessarily. The changes are internally consistent.

---

What works well

Frontmatter description is now precise. The old version was a catch-all; the new one correctly scopes to verify/cite/fact-check.

citation-anchors.md reference block (lines 96-103) is now a structured when-to-consult list rather than an afterthought at the end of the Format 2 subsection. Clearer decision criteria for when to open that rules file — a genuine improvement.

Field table k note (line 166) captures both formats in one place. Cleaner than the old 8-line subsection, and STOP AND CHECK (line 208) reinforces it identically.

Triage table simplification is justified. The "you prepared claims as evidence" row was a self-correction path that the narrower trigger now prevents upstream.

---

Minor gaps worth addressing

1. Follow-up verification: no guidance on re-using an existing .txt

Line 15 introduces: "If the user first asks a question about a document, answer it using prepare. If they then ask you to verify that answer, run this skill..."

But there is no hint about whether to re-run prepare or reuse .deepcitation/name.txt if it still exists. The removed read-only fast path had this explicitly: "If the same document was used and .deepcitation/name.txt still exists on disk, no need to re-run prepare."

Worth preserving — either appended to line 15 or as a note in S2 under the file/URL row. Without it, an agent will re-run prepare unnecessarily on a doc it already processed.

2. Format 2 gotcha example (low priority)

"converts automatically" reads like a SAFE agreement. AGENTS.md favors generic, domain-neutral examples (invoices, leases, contracts, memos). A SAFE is a contract so it is technically fine, but something more self-evidently neutral — e.g. "expires automatically" from a lease — would better match the spirit of the invariant.

3. Test plan gap

No coverage for the follow-up scenario introduced on line 15 (user reads first, then verifies). Worth adding: prepare a doc, get an answer, then ask to verify — skill fires, prepare is skipped if .txt exists.

---

No issues with

• Removal of the Tool precedence section ("ALWAYS use this skill") — overclaim, rightly gone
• Preamble + prepare in same turn (unchanged, still correct)
• HTML annotation path (untouched)
• STOP AND CHECK (untouched)
• Invariants section (untouched)

Net: approve. The re-prepare optimization is the one substantive thing worth adding before merge; the rest are suggestions.

[bensonwong]

@claude review the latest changes