lucidpaths/claude.md-workflow/proof
Claude Code and agent skill by lucidpaths/claude.md-workflow
- community
Checking that the SKILL.md file is still in the repository.
Verify a data operation without being fooled by your own checks — the discipline for migrations, dedups, reconciliations, mass-edits, backfills, and "is it actually done?" claims over real data. Chase mismatched numbers, verify the verifier, match by natural key, prove on a clone, confirm-before-delete. Use when the user runs /proof or is about to run or trust a data operation that deletes or overwrites data, or any migration, dedup, reconciliation, mass-edit or backfill.
Install
npx vskill@latest install lucidpaths/claude.md-workflow/proofSKILL.md
The same SKILL.md was published earlier as lucidpaths/claude.md-workflow/proof.
Proof — Verify a Data Operation Without Being Fooled by Your Own Checks
The thesis: "the check passed" is not "the thing is true." Most data-op disasters aren't a missing check — they're a check that silently lied. This is the discipline for any migration, dedup, reconciliation, mass-edit, backfill, or "is it really done?" verdict over real data.
The one reflex
When a number doesn't match what you expected, that is the most valuable thing on the screen. Chase it — don't round it off, don't explain it away, don't move on. The discrepancy is the finding.
Example: every bug worth catching first showed up as a count that was slightly wrong: 130 vs 98, 39 edits vs 4, 112 rows vs the real ~4.
Verify the verifier — before you trust a passing check
A check that can only pass is not a check.
- Confirm the query ran against what you think it did: real column / identifier names (a clean zero from the wrong column name is worse than no query at all), the specific table or scope (not "the field exists somewhere"), the right database / pod / cluster.
- A null join, an empty result, a "skipped" is NOT a pass. Make checks assert loudly. Before trusting that a check passed, prove it can fail — feed it a known-bad case in an isolated fixture or clone and watch it catch. Never plant a fault in production to prove a check.
- Count with real `count(*)`, never an estimate (
reltuples/n_live_tupare stale by design).
Match reality, not your assumptions
- Identify rows by a stable natural key (name / email / code), never by position or a mutable property. "
id == id" across two systems is the assumption that overwrites the wrong rows. Before matching, prove the key is unique on both sides:count(distinct key)equalscount(*)in each system; a key that fails it is not a key. - Never validate a snapshot against a moving source. Freeze the source, then compare the whole frozen window — two coincidentally-unchanged tables look like perfect parity and aren't.
- Exclude re-serialized or intentionally-divergent columns from parity hashes, or you'll "find" edits that were never made — and a gate must understand intended difference, not flag it as drift.
Prove it on a clone before prod
Rehearse the operation on a throwaway clone that matches the target's real config — a clone missing prod's settings hides exactly the bug that only fires in prod. Assert every invariant on the clone, re-run once to prove idempotency, then run for real under sign-off.
Before you destroy anything
- Back up the exact predecessor first (the version you inspected, not the newest backup), verify it is non-empty, and restore it once on an isolated target. If the source changed since you inspected it, stop and inspect again.
- A failed read is not an empty state. When reading the predecessor fails, abort the write. Write through a transaction or a same-directory atomic replace; never truncate the predecessor before the replacement is complete (atomic replace does not serialize two writers).
- Read back the exact target after the mutation. A command that exited 0 is not proof the intended state reached its consumer.
- Canonical-confirmed-then-delete: confirm the correct entry is in place, then remove the wrong one. Deletion after confirmation isn't dataloss — but hoarding tagged zombies isn't cleanup either.
Done when
One invariant over the full set (never a sample) is provably 0 / N — zero rows lost, zero dead links, zero mismatches you can't explain. Treat docs, prior session notes, and your own earlier claims as stale until re-verified against live. In the result, keep separate what you measured against live data versus what you're trusting from a doc or estimate — they are not the same confidence. "Looks done" is not done; show the invariant.
Closing block (required) — State the close in this form — otherwise you asserted it, you did not prove it
FAIL-PROOF is the load-bearing line. A check with no demonstrated failure mode is the exact thing this skill exists to catch, and it is invisible in a passing result.
Security
- Security scanPassed, 90 of 100Pattern checks on the skill's files.
- Model reviewPassed
Quality evaluation
- Task evaluationNot runNo evaluation yet. Run one in Skill Studio.
Source
- SKILL.md in the repositoryCheckingclaude-starter-kit/.claude/skills/proof/SKILL.md
Popularity
- 7-day trend score
- 0
Not refreshed yet.
Works with
Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.
- Claude Code
- Cursor
- GitHub Copilot
- Windsurf
- Codex
- Gemini CLI
- Cline
- Amp
- Roo Code
- Goose
- Continue
- Kilo Code
- Junie
- Trae
- OpenHands
- Qwen Code
- AI Agent
- Tabnine Chat
- Sourcegraph Cody
- Amazon Q
- Firebase Genkit
- Sweep
- Devin
- Phind
- Mentat
- Claude Desktop
- Aider
- GPT Pilot
- Plandex
- bolt.new
- v0
- Replit Agent
- Anse
- TypingMind
- BetterChatGPT
- LibreChat
- Warp AI
- Pieces
- Fig AI
Embed badge
[](https://verified-skill.com/skills/lucidpaths/claude.md-workflow/proof)