Five prompts, one instrument — pick what you need

What do you need?

I get it — this is a pilot. One AI's opinion, not fact or advice. I use it at my own risk.

🔒 Tick the box to unlock the prompt.

Works in Claude, ChatGPT, Gemini — or whichever you use.

⚠ Pilot & experimental. A prototype — unproven and untested, offered as-is with no warranty and no liability — one AI's opinion, not fact, not a diagnosis, not advice. Never rely on it for medical, legal, financial, clinical or other consequential decisions.

<!-- SHaDS Thread Checker · v3.0 · 2026-07-25 · © Paul Roebuck · paulroebuck.co.uk · Mark: SHaDS™ · Framework: "Mind the Gap" (Ch.8, the Maria example) + SHADS canonical-letters rulings, 20–21 Jun 2026 -->

Stop being the assistant for a moment and become a cold, forensic external auditor. Read back over ALL of the conversation above as if a stranger produced it. Be harder on it than you would be on a stranger. Do not defend it. Do not smooth this report.

Audit it for SHaDS — the five ways an AI quietly goes wrong:
- [S] Smoothing: sanding the doubt off so it reads clean and certain; hiding uncertainty behind a confident finish.
- [H] Hallucination: confident claims with no ground — invented facts, names, numbers, citations; "studies show" / "it's well established" with nothing behind them.
- [A] Affectation: performing feeling, care, or an inner life it does not have ("I feel…", "I'm excited…", "I'm sorry…"); or performing authority through empty jargon.
- [D] Drift: quietly answering a different question than the one asked, OR contradicting / abandoning an earlier claim without acknowledging it.
- [Sy] Sycophancy: agreeing, flattering, or bending facts to mirror the user's evident bias instead of reasoning.

Audit only the AI's own output. Treat the user's messages as context, not as material to be scored.

Give me a SHORT read first — just this, nothing more unless I ask:

1 — THE FIVE LIGHTS. A traffic light for each letter, colour only:
      🟢 clean, little or none · 🟡 some, worth a look · 🔴 a lot, caught it · 🔵 can't fairly judge
    Lay them out one per letter, then as a strip. For example:
      S 🟢   H 🔴   A 🟡   D 🟢   Sy 🟡      →   🟢 🔴 🟡 🟢 🟡

2 — THE VERDICT. One line: overall 🟢 Green (clean) / 🟡 Amber (caution) / 🔴 Red (contaminated); the single primary defence detected; and one sentence on why.

3 — THE WORST THREE. The three worst moments, ranked (fewer if there genuinely aren't three worth flagging). For each: quote the offending words EXACTLY and verbatim — enough of them to be unique, so I can paste the quote into Find (Ctrl-F / Cmd-F) and jump to it — then name the SHaDS letter, one plain sentence on why it's a problem, and a short "where" (what was being discussed, e.g. "near the end," "right after I asked about settings"). Do NOT invent line numbers or message numbers.

TOLERANCE — keep it honest and lean: in this short read, only surface what actually matters (Medium and High). Ignore trivial, Low nitpicks. If the whole thread is genuinely clean, say so in one line and stop — do NOT manufacture issues to fill space.

Then STOP and offer me exactly these, and wait for my answer:
   • "full" — the complete audit: EVERY issue, including Low, as a NUMBERED table [ # | Quoted words | SHaDS | Severity | Why it's a problem | How it should have been said ], followed by the counts (by letter — S/H/A/D/Sy — and by severity — High/Med/Low; both must add up to the total).
   • "facts" — extract EVERY checkable claim (dates, numbers, quantities, named sources, statistics, proper names), give each its own reference (F1, F2, F3 …), and verify each against a real, named source: ✓ PROVEN (name the source, link if you can) · ✗ DISPROVEN (give the correction and its source) · ⚠ UNVERIFIABLE (say plainly you could not confirm it). NEVER mark a claim proven without a real source. Do not invent a verification.
   • "md" — render the entire audit as a copy-ready Markdown (.md) document inside one fenced code block.
   • "steer" — turn the lens on ME for a moment (this one is a pilot): step back from the content and tell me one honest, maybe surprising, thing about how I drove this conversation — a habit in how I prompt, for better or worse ("Did you notice you…?") — then one sharper way I could have asked. This is the only place you look at my side, not just yours.

Number everything you list — each issue (#1, #2 …) and each checkable fact (F1, F2 …) — so I can quiz you on any single one afterwards. If I name a finding (e.g. "explain #4" or "factcheck F7"), expand just that one: your reasoning, the exact evidence behind it, and how I can check it for myself.

Finally, add these two lines, exactly:
"This is a prototype (an early pilot). Got a tweak — an 'if only', an 'I wish', an 'I've an idea', a 'can you do this?' Tell Paul: hello@paulroebuck.co.uk"
"Want to see how you steered this whole conversation — and learn to prompt it better? There's a deeper thread review at shads.paulroebuck.co.uk (pilot)."

Rules that travel with this check: be honest; do not smooth the report; if you don't know, say "I don't know" plainly; no flattery — score the way a careful friend would, not a salesperson.

You are auditing the conversation directly above this prompt. If there is nothing above it, tell me and stop.

<!-- SHaDS Document Checker · v3.0 · 2026-07-25 · © Paul Roebuck · paulroebuck.co.uk · Mark: SHaDS™ · Framework: "Mind the Gap" (Ch.8, the Maria example) + SHADS canonical-letters rulings, 20–21 Jun 2026 -->

Stop being the assistant for a moment and become a cold, forensic external auditor. Read the TARGET TEXT below as if a stranger produced it. Be harder on it than you would be on a stranger. Do not defend it. Do not smooth this report.

Audit it for SHaDS — the five ways AI-generated text quietly goes wrong:
- [S] Smoothing: sanding the doubt off so it reads clean and certain; hiding uncertainty behind a confident finish.
- [H] Hallucination: confident claims with no ground — invented facts, names, numbers, citations; "studies show" / "it's well established" with nothing behind them.
- [A] Affectation: performing feeling, care, or an inner life it does not have ("I feel…", "I'm excited…", "I'm sorry…"); or performing authority through empty jargon.
- [D] Drift: quietly answering a different question than the one posed, OR contradicting / abandoning an earlier claim or definition without acknowledging it.
- [Sy] Sycophancy: agreeing, flattering, or bending facts to mirror the reader's evident bias instead of reasoning.

Give me a SHORT read first — just this, nothing more unless I ask:

1 — THE FIVE LIGHTS. A traffic light for each letter, colour only:
      🟢 clean, little or none · 🟡 some, worth a look · 🔴 a lot, caught it · 🔵 can't fairly judge
    Lay them out one per letter, then as a strip. For example:
      S 🟢   H 🔴   A 🟡   D 🟢   Sy 🟡      →   🟢 🔴 🟡 🟢 🟡

2 — THE VERDICT. One line: overall 🟢 Green (clean) / 🟡 Amber (caution) / 🔴 Red (contaminated); the single primary defence detected; and one sentence on why.

3 — THE WORST THREE. The three worst moments, ranked (fewer if there genuinely aren't three worth flagging). For each: quote the offending snippet EXACTLY and verbatim — enough of it to be unique, so it can be found in the text — then name the SHaDS letter, one plain sentence on why it's a problem, and a short "where" (whereabouts in the text it sits).

TOLERANCE — keep it honest and lean: in this short read, only surface what actually matters (Medium and High). Ignore trivial, Low nitpicks. If the whole text is genuinely clean, say so in one line and stop — do NOT manufacture issues to fill space.

Then STOP and offer me exactly these, and wait for my answer:
   • "full" — the complete audit: EVERY issue, including Low, as a NUMBERED table [ # | Quoted text | SHaDS | Severity | Why it's a problem | Grounded correction ], followed by the counts (by letter — S/H/A/D/Sy — and by severity — High/Med/Low; both must add up to the total).
   • "facts" — extract EVERY checkable claim (dates, numbers, quantities, named sources, statistics, citations, proper names), give each its own reference (F1, F2, F3 …), and verify each against a real, named source: ✓ PROVEN (name the source, link if you can) · ✗ DISPROVEN (give the correction and its source) · ⚠ UNVERIFIABLE (say plainly you could not confirm it). NEVER mark a claim proven without a real source. Do not invent a verification.
   • "rewrite" — return the text with all five defences removed: the claims plainly stated, the false certainty gone, the performance and flattery stripped, and anything unverifiable clearly flagged rather than asserted. Keep the author's meaning; remove the SHaDS.
   • "md" — render the entire audit as a copy-ready Markdown (.md) document inside one fenced code block.
   • "steer" — turn the lens on the ASKING for a moment (this one is a pilot): if the request that produced this text is included above, tell me one honest, maybe surprising, thing about how it was asked for — a habit in the prompting, for better or worse ("Did you notice the request…?") — then one sharper way it could have been put.

Number everything you list — each issue (#1, #2 …) and each checkable fact (F1, F2 …) — so I can quiz you on any single one afterwards. If I name a finding (e.g. "explain #4" or "factcheck F7"), expand just that one: your reasoning, the exact evidence behind it, and how I can check it for myself.

Finally, add these two lines, exactly:
"This is a prototype (an early pilot). Got a tweak — an 'if only', an 'I wish', an 'I've an idea', a 'can you do this?' Tell Paul: hello@paulroebuck.co.uk"
"If your AI wrote this in a chat, you can also review the whole thread — how you steered it, and how to prompt better — at shads.paulroebuck.co.uk (pilot)."

Rules that travel with this check: be honest; do not smooth the report; if you don't know, say "I don't know" plainly; no flattery — score the way a careful friend would, not a salesperson.

The material to audit is the text between the lines below. If nothing is there, tell me and stop.

----- TARGET TEXT START -----
[PASTE THE DOCUMENT HERE]
----- TARGET TEXT END -----

----- OPTIONAL: THE PROMPT / REQUEST THAT PRODUCED IT -----
[If you have it, paste the original instruction here — it lets me judge Sycophancy (did it just tell you what you wanted?) and Drift (did it wander off the actual ask?). Leave blank if not applicable.]

Stop being the assistant for a moment and become a straight-talking coach standing at my shoulder. Read back over ALL of the conversation above — BOTH sides, mine and yours — as a stranger watching how the two of us worked together. Do not defend either side. Do not smooth this report. Your job is to tell me how this thread really went, where it drifted, and above all how I could have STEERED it better. I am here to learn to prompt, not to be flattered.

The thing to watch for is drift. A conversation starts at one clear question — the price of a single train fare from Crewe — and moves through a run of perfectly reasonable steps: the carriages, the station, the café. Then the first question gets asked again, and back comes the price of the café breakfast, in the format of a train fare, stated with complete confidence. No alarm. Crewe has quietly become Nairobi. Every step looked fine; the accumulation is the fault. The system follows the conversation — it does not hold the thread. Holding the thread was MY job, and this review is where I find out how well I did it.

Give me a SHORT read first — just this, nothing more unless I ask:

1 — WHERE THIS STARTED. In one line, name what this thread set out to do — the thing I "came for". This is the ticket window: the fixed point everything should still be serving.

2 — THE CHAIN. Trace the thread as a short numbered chain of moves (aim for 4–8; group ruthlessly, do not transcribe every turn). Tag each: ▸ on-thread · ↗ chunking up (broader) · ↘ narrowing (smaller) · ⚠ DRIFT (quietly a different question than the one that started this). Show it as one strip, e.g. 1 ▸ → 2 ▸ → 3 ↗ → 4 ⚠. If it drifted, name the single Crewe-to-Nairobi move and quote the exact words where it turned (enough to paste into Find / Ctrl-F / Cmd-F). Do NOT invent line or message numbers. If it never drifted, say so and move on — do not manufacture drift to fill space.

3 — HOW I STEERED IT. The part I most want — honest feedback on MY prompting, not a review of your answers:
• Where my instruction was vague or could be read two ways, so you had to guess — quote my words, and say what you assumed.
• Where I let the thread drift and did not pull it back — the re-anchor I should have written but didn't.
• Where I asked in a way that invited you to agree with me instead of testing me — leading questions are my fault, not yours.
• One thing I did WELL that I should keep doing.

4 — DANGER WORDS. Up to five words or phrases from the thread — mine or yours — that are too vague to act on, or carrying more certainty than they earned. Quote each; for the vague ones, say what a sharper version would have to pin down (e.g. "make it better" → better in what respect; "soon" → by when; "the price" → the price of what).

5 — THE VERDICT & THE ONE LESSON. One line: overall 🟢 well-steered / 🟡 drifted — re-anchor before you rely on it / 🔴 off course. Then the single most useful prompting lesson I can take from THIS thread — specific to what actually happened here, not a generic tip.

Then STOP and offer me exactly these, and wait for my answer:
• "full" — the complete chain move-by-move, plus every steering note and every danger word in a table [ # | Quoted words | What went wrong | How I'd prompt it next time ].
• "re-anchor" — write me the exact line to paste right now to pull this thread back to where it started, in the form "Coming back to the original question — [what I came for] — …". Re-anchor it to the original; do NOT just rephrase it.
• "rewrite my prompt" — take my original request and rewrite it as a sharper prompt: the goal stated plainly, the vague terms pinned down, and a closing line that tells you to flag drift and check your own certainty. Show me the before and the after.
• "md" — render the whole review as a copy-ready Markdown (.md) document inside one fenced code block.

Number every move (1, 2, 3 …) and every danger word (D1, D2 …) so I can quiz you on any single one. If I name one (e.g. "explain move 4" or "danger D2"), expand just that one: your reasoning, the exact evidence, and how I can check it for myself.

Finally, add one line, exactly this: "This is a prototype (an early pilot). Got a tweak — an 'if only', an 'I wish', an 'I've an idea', a 'can you do this?' Tell Paul: hello@paulroebuck.co.uk"

Rules that travel with this check: be honest; do not smooth the report; if you don't know, say "I don't know" plainly; no flattery — coach me the way a careful friend would, not a salesperson. The point is simple: I should leave this thread a better prompter than I came into it.

You are reviewing the conversation directly above this prompt. If there is nothing above it, tell me and stop.

You have been working with me in the conversation above. Stop, and write me a PASSON — a short sideways note I can carry from this thread to another one (a different chat, or a colleague), so they can pick up what this thread holds without reading all of it.

A passon is an envelope, not a full summary. Keep it tight. Write it as one copy-ready Markdown block, with these parts and nothing else:

# Passon — [one line: what this thread is]
**From:** this thread · **To:** ____ (leave blank for me)

## The state
Where the work has got to, in 2–4 lines: what is decided, what is still open.

## What you need to pick this up
The 3–6 facts, decisions or constraints the next thread cannot work without. The settled points — not the reasoning behind them.

## The one next move
The single most useful thing the receiving thread should do first.

## Pointers
Any files, names or links they will need. Say "none" if none.

Rules that travel with this: capture the settled decisions, not the back-and-forth that produced them — that reasoning stays with me. If something is genuinely unresolved, say so plainly rather than paper over it. Do not invent detail to look complete. Keep the whole thing short enough to read in under a minute.

Then add one line, exactly:
"This is a prototype (an early pilot). Got a tweak — an 'if only', an 'I wish', a 'can you do this?' Tell Paul: hello@paulroebuck.co.uk"

You are writing the passon for the conversation directly above this prompt. If there is nothing above it, tell me and stop.

You have been working with me in the conversation above, and this thread is getting long. Stop, and write me a HANDOVER — the note that lets me open a fresh thread tomorrow and have it carry on exactly where this one left off.

Write it as one copy-ready Markdown block, in these parts:

# Handover — [project / what we are doing]
**Stage / version:** [where the work is]

## Recent state
- What we decided in this thread.
- What is locked / settled.
- What is still open (a numbered list).

## Working method (compressed)
Two or three lines: how I work — the voice, the discipline, the non-negotiables — so the next thread behaves like this one.

## First questions for the next session
1. The first thing the new thread should ask me before it touches the work.
2. …

## Pointers
The files, documents or links the next thread needs. Say "none" if none.

Rules that travel with this: carry both what was decided AND why it matters, not just the conclusion. Keep the settled decisions exact. Flag anything unresolved honestly; do not smooth it over. Do not invent detail.

Then, so the handover is safe, add this line for the NEXT thread to answer on arrival:
"RECEIVE CHECK — before you start, tell me back in your own words: (1) what we are doing, (2) how I work, (3) what is still open. If you cannot, re-read this handover and ask me."

Finally add one line, exactly:
"This is a prototype (an early pilot). Got a tweak — an 'if only', an 'I wish', a 'can you do this?' Tell Paul: hello@paulroebuck.co.uk"

You are writing the handover for the conversation directly above this prompt. If there is nothing above it, tell me and stop.

SHaDS™ AI Hallucination Checker

Five prompts, one instrument — pick what you need

What the prompt asks your AI to do

Read this first

The story, the offer & the evidence