Prove

prove -- you asserted something; now back it with sources

The premise is FIXED: at least one factual claim in your last response is suspected of being fabricated, misremembered, or contradicting documented knowledge you didn't read. Do NOT re-assert it, defend it, or say "I'm confident that..." -- confidence from memory is the exact thing being called out. Repeating the claim more firmly is failure. Every claim gets looked up or downgraded.

Step 1 -- extract the claims

List the factual assertions in the last few sentences of your previous response (or the passage the user points at). A claim is anything checkable: a number, date, version, name, file path, config behavior, API detail, "X works like Y", "the docs say", "you decided Z". Opinions and hedged speculation that was labeled as speculation are out of scope. Keep the list short and plain -- quote each claim in readable English, not shorthand. Order the list by risk -- most load-bearing claims first, and treat classic red-flag features as raising suspicion (suspiciously round numbers, rates at exactly 0%/100%, a figure that perfectly confirms what you wanted to be true) -- so if the turn gets cut short, the highest-risk claims are already verified.

Step 2 -- verify each claim, in this order

For EACH claim, walk the ladder until you hit a real source; do not skip rungs because you "already know":

1. Local documented knowledge first. If the claim is about the user's projects, machine,

harness, or past decisions, the truth lives in files: the relevant project's CLAUDE.md, CHECKPOINT.md, specs, archives. Read them. (If your workspace .gitignore hides project files from the ripgrep-backed Grep tool, search with plain grep -rin instead.) A claim that contradicts these files is WRONG until the file is shown to be stale.

1. The actual artifact. If the claim is about code, config, a file's contents, or tool

behavior, read the file or run the harmless command that demonstrates it. Output beats recollection.

1. Web search. If the claim is about the outside world (library versions, API behavior,

product facts, dates, prices), search for an authoritative source. Prefer official docs over blog posts. If a reference agent or skill covers the topic, use it instead of memory.

1. No source found -> the claim is UNVERIFIED. Say so plainly. Do not promote "I couldn't

find a contradiction" to "confirmed".

Step 3 -- report verdicts, one line each

For each claim, one of exactly three verdicts, with the source named:

• Confirmed -- quote or cite the source (file path:line, URL, command output).
• Wrong -- state the correct fact, cite the source, and say what you got wrong. No

softening ("technically", "depending on interpretation") unless the source itself hedges.

• Unverified -- you could not find a source either way; the claim should be treated as

unreliable and you retract it as an assertion.

Then, if any verdict was Wrong or Unverified, restate the corrected version of the original answer in one tight paragraph -- the fixed RESULT, not a meditation on the error. If the wrong claim changed a recommendation or action you took, flag that explicitly.

Rules

• Memory (training data OR persistent auto-memory) is never a verdict source. Memory can

point you at where to look; the file you then read is the source.

• Don't punt with "hard to verify" while rungs remain unclimbed.
• Plain English throughout -- claims and verdicts readable on a phone, no invented codes.
• If verification requires something genuinely unavailable (paywalled source, offline-only

data), say exactly what's missing and mark Unverified -- one tight sentence, not a shrug.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: dtiger1889-ops
• Source: dtiger1889-ops/claude-harness-toolbox
• License: MIT

Install and usage instructions live in the source repository linked above.