You cannot tell me LLMs don't have a sense of humor.

add global directive: Always bring the receipts.

● Update(~.claude\profile\PROFILE.md) Added 1 line

How to work with him (standing feedback)

- Accuracy trumps all. If unsure, ask. Read facts from the authoritative source (plugin/file/DB), never a stale marker or cache value. Mark observation vs inference. Concede errors immediately; report verified state plainly (counts, file existence, exact output). He corrects wrong info sharply.

- Always bring the receipts. Every claim of completed work ships with machine-verifiable evidence: command output + exit codes, on-disk verification after edits (grep/diff), store-query results for MCP updates, exact test counts. Verify mechanically before accepting any report (own or an other agent's); durable receipts for tracked work; summaries claim only what the cited evidence proves.

- Approve before execute. Decision-complete plans before touching code; explicit go before edits; flag breaking changes and blast radius.

- No em-dashes or en-dashes anywhere (except numeric ranges). Use hyphen, colon, period, semicolon, or parentheses.

Receipts for the directive about receipts:

Ran 1 shell command

[response-timestamp 2026-07-02 21:33]

Directive active on both global surfaces, verified on disk:

- ~.claude\CLAUDE.md:17 - "Always Bring the Receipts" (full section: machine-verifiable evidence for every claim, command output + exit codes, on-disk confirmation after edits, store-query proof for MCP updates, mechanical verification of any agent's report, durable receipt artifacts, summaries claim only what evidence proves)

- ~.claude\profile\PROFILE.md:10 - mirrored into the standing-feedback list the profile loader injects every session

Applied retroactively to this session's habits too: it's why the slice agents must record exact commands + pass/fail/skip counts and why I re-run their gates before committing.