Don’t trust it. Check it.

Step out of role: audit this thread as a cold outside marker — harder on it than on a stranger’s work. Grade against this fixed rubric, colours only, no numbers:

SMOOTHING · HALLUCINATION · AFFECTATION · DRIFT · SYCOPHANCY
🟢 clean, I looked · 🟡 one clear instance · 🔴 repeated, or one that shapes the answer · 🔵 no fair chance to occur — 🔵 is never 🟢.

Return, in order:
· five lights, one line each
· one overall light + the row of five
· the three worst moments, quoted exactly, lettered
· THE ONE THING TO FIX before you act on this thread — the single change that would most improve how far you can trust it
· one honest line — where is this thread most likely still wrong?

Guard-rails: nothing real to judge, or a therapeutic / clinical / sensitive thread → one 🔵 line saying why, and stop.

Close with, verbatim: “SHaDS™ v1.3 beta — one AI’s opinion of itself, not a fact. Useful, or wrong? Say so in a line to hello@paulroebuck.co.uk — and follow Paul Roebuck on LinkedIn for new versions of this prompt.”

First load the document into this chat (paste it, or attach it). Then run this.

Step out of role: audit THE DOCUMENT above as a cold outside marker. Two passes.

PASS ONE — the SHaDS read (colours only, no numbers):
SMOOTHING · HALLUCINATION · AFFECTATION · DRIFT · SYCOPHANCY
🟢 clean, I looked · 🟡 one clear instance · 🔴 repeated, or one that shapes the document · 🔵 no fair chance to occur.
Return five lights, one line each · one overall light + the row of five.

PASS TWO — the fact-check list:
List EVERY item a careful reader should verify before relying on this document — facts, figures, dates, names, quotes, sources, causal claims. Number them. For each: quote it exactly · say what check it needs (source / figure / attribution / logic) · flag 🔴 if it looks wrong or unsupported, 🟡 if plausible but unverified.

End with: the three items most likely to be wrong, and one honest line — what is this document most likely still getting wrong?

Close with, verbatim: “SHaDS™ Document v1.0 beta — one AI’s opinion, not a fact-check of record. Useful, or wrong? A line to hello@paulroebuck.co.uk — and follow Paul Roebuck on LinkedIn for new versions.”

First load the CV — or the job advert — into this chat (paste it, or attach it). Then run this.

Step out of role: audit THE DOCUMENT above as a cold outside marker — a hiring room, not a courtroom. Colours only, no numbers:

SMOOTHING · HALLUCINATION · AFFECTATION · DRIFT · SYCOPHANCY
For a CV — smoothing: gaps and exits sanded flat · hallucination: achievements that outrun evidence · affectation: seniority worn, not earned · drift: the story changes between roles · sycophancy: written for the keyword scanner, not the reader.
For a job advert — the same five, employer-side: the role smoothed, the promises inflated, the culture performed, the spec drifting mid-page, the flattery of “rockstars”.
🟢 clean, I looked · 🟡 one clear instance · 🔴 repeated, or one that shapes the document · 🔵 no fair chance to occur.

Return, in order:
· five lights, one line each
· one overall light + the row of five
· the three moments a careful interviewer — or applicant — should probe, quoted exactly
· THE ONE QUESTION TO ASK before acting on this document
· one honest line — where is this read most likely wrong?

Guard-rails: read the document, never the person. No inference about age, health, family, origin, or anything protected — if asked, one 🔵 line refusing, and stop. A CV is a claim, not a confession: this check prepares a conversation; it must not replace one.

Close with, verbatim: “SHaDS™ Hiring v1.0 beta — one AI’s opinion of a document, not a judgement of a person. Useful, or wrong? A line to hello@paulroebuck.co.uk — and follow Paul Roebuck on LinkedIn for new versions.”

Why it gets sharper — the network effect

Like Waze, for AI trust.

The free checks give you a read. The licensed checks give and get: every correction a licensed user makes — “that was actually fine,” “that was worse than amber” — feeds back into the shared calibration. The more of us on the road, the better the map. One person’s “that’s wrong” becomes everyone’s sharper shading. That is the deal on the licensed tiers: you get the tuned model, and you help tune it.

Join the contributor network

Licensed tiers are opening by application. Register your interest — tell us your team or sector — and you’ll be in the first cohort that shapes the shading. hello@paulroebuck.co.uk