# Align

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

- Skill: `tony/align` (Agent Skill)
- Install (CLI): `npx skillmds@latest add tony/align`
- Raw SKILL.md: https://api.skillmd.com/api/skills/tony/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/align

---



# `/double-check:align`

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.