Live state first

The single most-repeated lesson on this machine: twelve-plus recurrences of building against a world that no longer exists — invented columns from a stale handover, a "still open" item carried across four handovers a month after it was done, an audit finding two migrations stale, a reviewer's citation superseded by a later migration. Layer: norm + playbook — the one-line norm in the global CLAUDE.md is the everyday trigger; this file is the depth.

The rule

Every document claim — including the repo's own comments, skills, and FORWARD notes — is a claim, not a fact. Before it becomes an input to code, SQL, config, or a fixture, verify it with one cheap probe of the live source. The latest migration, policy, or config is the truth; when two documents disagree, the catalog settles it.

How to apply

• One cheap probe before building: a catalog query for schema claims (information_schema, pg_policy, pg_get_functiondef), the deployed surface for behaviour claims, the actual config for settings claims.
• The file is not the live object, in either direction. The live database routinely runs ahead of main (raw-applied unmerged migrations) and behind the repo (comments and fixes never applied). A migration file's text is evidence of intent, not of state.
• Query the live distribution, not just the schema, before building on a data assumption — 254 residue rows once would have shipped a correct feature looking broken.
• Structured data needs a parser, not a grep with context lines. Grep-with-context misled on JSON twice; scope schema greps to the block they describe.
• Staleness ranks: "still open" items in handovers, path hints in a repo's own docs, and audit trackers are the most likely claims to be stale — they record status at a moment, not state.

What this skill does not do

It does not probe repo state — fetch, PR lists, and identifier ceilings are parallel-work-recon's territory. It does not define migration verification (that is db-migration-verification's post-apply half; this skill is the pre-compose half).

Why

Documents decay the moment parallel work exists, and composing from a stale description produces confident, wrong artifacts that pass review because the reviewer reads the same stale document. Evidence: PropOS LESSONS_LEARNED Sessions 9, 11, 12, 22, 41, 52, 54, 2026-07-06, 2026-07-08, 2026-07-09, 2026-07-19, 2026-07-20; ASH LESSONS_LEARNED 2026-05-25, 2026-07-10.

Additions from the misses-log sweep (2026-10-22)

• A time, date or count written into a record is read from a command, never composed. Run the clock (date -u) or the count in its own call, then write what it printed in the next; a reading in the same command as the write was typed before it existed. A value no command can give is written as unknown. (ICC L-093, seven instances.)