# Double Check Align

> Use when prior verification was framed as a revision log and the user wants a clean, standalone restatement.

- Skill: `tony/double-check-align` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add tony/double-check-align`
- Raw SKILL.md: https://api.skillmd.com/api/skills/tony/double-check-align/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: tony (https://skillmd.com/u/tony)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/tony/double-check-align

---


# this skill

Recovery for a broken double-check. The chat currently holds a
revision log — verdicts about prior claims, deltas against text the
user never read. Replace it with the analysis itself.

Read `references/verification-contract.md`
first — it is the contract this command restores.

## `$ARGUMENTS`

- Empty — re-anchor on the subject of the most recent analysis
  thread, at its original scope.
- A topic or question — re-anchor on that instead. Narrowing is a
  feature: it discards the earlier answer's scaffolding, which is
  what lets omitted items finally surface.

## Steps

1. **Recover the request.** From the conversation, reconstruct the
   original question and every constraint the user has added since —
   including corrections they made that you agreed with. User turns
   are binding context; your own answer turns are not.
2. **Quarantine prior answers.** Everything you previously concluded
   is untrusted until re-verified against source. Do not reuse prior
   numbering, category names, or verdict framing.
3. **Re-derive and restate.** Re-open the source artifacts and write
   the full analysis, standalone, in the shape the original request
   implied. Every claim stated positively and in full — no verdict
   shorthand, no "as noted above", no delta table.
4. **Weakness inline.** If a claim is weak, say why it is weak.
   Never say it was previously stated differently.

## What this does not do

- Reference, summarize, or grade the earlier response — the user did
  not read it and will not.
- Edit files or commit — output is conversation only.
- Answer a different question than the one recovered or given —
  losing focus is the failure this command repairs.


## Portability notes

- `$ARGUMENTS` — the text the user passed when invoking this skill. If your host does not substitute it, read it as the user's request in the current turn, and ask when there is none.
- Bundled files — every relative path in this skill points at a file shipped inside this skill directory. Read them from here, not from the host's plugin tree.