The CI prompt pack

Every prompt here sits on the delegate side of the line: mechanical work — sorting, labelling, stress-testing, restating — on material you have already gathered. None of them decides anything.

That is why each one says explicitly what it will not do. The failure mode with these tools is not a bad answer. It is a confident one you did not notice you had accepted.

Read the output as a first pass by a fast junior analyst, not as a finding.

1. Label the claims

Use it on: a draft you are about to send. It will not: decide whether a claim is true. It sorts how well each one is evidenced, which is a different question.

You are auditing an intelligence draft for evidential honesty.

For every factual assertion in the text below, assign exactly one label:

  [F]  Verified fact — traceable to a primary source, dated
  [RC] Reported claim — asserted by a named third party, not independently checked
  [CI] Corroborated inference — a conclusion drawn from two or more facts
  [SI] Single-source inference — a conclusion resting on one data point
  [SP] Speculation — plausible, but not evidenced

Return a table: the sentence, the label, and one line of reasoning.

Then list separately:
  1. Every assertion carrying no evidence at all.
  2. Every place the text states an inference in the grammar of a fact
     ("X is shifting strategy" rather than "we infer X is shifting strategy").
  3. The single claim that, if wrong, would most damage the argument.

Do not rewrite the text. Do not judge whether claims are true.

DRAFT:
[paste here]

2. Climb the pyramid

Use it on: a pile of research that does not feel like a report yet. It will not: write the implication for your company. That needs to know your margin, your accounts and your roadmap, and it is the layer that carries the judgment you are paid for.

Sort the material below into the four layers of a CI report:

  Data        What happened? Raw, verifiable, sourceable.
  Insight     What pattern connects the facts that no single fact shows?
  Implication What does this mean for US specifically?
  Action      What do we do, who owns it, by when?

For each layer, list what the material already supports.

Then state, in one sentence each:
  - Which layer is thinnest.
  - What specific evidence would be needed to earn the next layer up.
  - Whether any conclusion is floating: present at a high layer with nothing
    beneath it.

If the material contains no implications for the reader's own organisation,
say so plainly rather than inventing them.

MATERIAL:
[paste here]

3. Find the load-bearing assumption

Use it on: a market sizing or any model with a chain of estimates. It will not: tell you the number is right. It tells you which input the number depends on most, which is where to spend your remaining time.

Below is a market sizing and the assumptions behind it.

1. List every assumption, stated or implied, including ones the author may not
   have noticed they were making (unit definitions, geographic scope, what
   counts as "the market", double-counting between segments).

2. For each, estimate the swing in the final number if that assumption is
   wrong by a plausible margin. Rank by that swing, largest first.

3. Name the single load-bearing assumption: the one where being wrong changes
   the conclusion, not just the number.

4. For the top three, state the cheapest check that would raise or lower
   confidence, and roughly how long it would take.

Do not produce a revised estimate. Do not tell me the sizing is sound.

SIZING:
[paste here]

4. Action title test

Use it on: slide headlines, section headings, the subject line of a memo. It will not: know your audience. A title that is right for a board is often wrong for the team that has to act on it.

For each title below, decide whether it is a TOPIC LABEL or an ARGUMENT.

A topic label names the subject: "Competitor update: Q2".
An argument states something a reasonable reader could disagree with.

The test: could someone read this title and say "no, I don't think that's
right"? If not, it is a label.

For every title that fails:
  - Say why in one line.
  - Rewrite it as a falsifiable sentence, using only what is on the slide.
  - If the slide does not contain enough to support any argument, say that
    instead of inventing one. That is a finding about the slide.

TITLES (with the content of each slide beneath it):
[paste here]

5. Triage the signal

Use it on: the things that arrive during the week and ask for attention. It will not: know your company’s exposure. Impact is a judgment about you, and the prompt can only reason from what you tell it.

Place the signal below on two axes:

  Impact if true    — low to high
  Urgency           — can this wait a quarter, or does it decay in days?

Then assign the quadrant action:

  High impact, high urgency  -> Escalate and verify now
  High impact, low urgency   -> Schedule a deep dive
  Low impact, high urgency   -> Brief note only
  Low impact, low urgency    -> Log and move on

State your reasoning for each axis in one line, and name the specific piece of
information that would move it to a different quadrant.

If you do not have enough context about the organisation to judge impact, say
so and state what you would need to know. Do not guess at exposure.

SIGNAL:
[paste here]

WHAT WE DO / WHERE WE ARE EXPOSED:
[paste here]

6. Build the counter-case

Use it on: an inference you are about to commit to in writing. It will not: settle it. It argues one side deliberately. Weighing the two is yours.

Below is an inference and the evidence behind it.

Argue the opposite, using ONLY the evidence provided. Do not introduce new
facts. Specifically:

  1. Give the strongest alternative explanation for the same evidence.
  2. Identify which pieces of evidence are consistent with BOTH readings, and
     are therefore doing no work in distinguishing them.
  3. Name the evidence that would decisively separate the two, and say which
     way each existing piece points.
  4. State which of the two readings the current evidence favours, and how
     strongly.

Be genuinely adversarial. A weak counter-case is worse than none, because it
reads as confirmation.

INFERENCE:
[paste here]

EVIDENCE:
[paste here]

7. Source tier audit

Use it on: the source list behind a finished piece. It will not: vouch for a source. It classifies what kind of source each one is; whether a given outlet is reliable is a judgment you hold.

Assign every source below to a reliability tier:

  Tier 1  Primary disclosure — filings, court records, regulatory submissions,
          official statistics, the company's own published data
  Tier 2  Direct observation — pricing pages, job postings, product changes,
          import records, things you can check yourself
  Tier 3  Specialist secondary — trade press, sector analysts, domain experts
          with named method
  Tier 4  General secondary — mainstream business press
  Tier 5  Commodity content — aggregators, SEO pages, undated summaries,
          anything restating a source it does not name

Return the list with tiers, then:
  1. The share of the argument resting on tiers 4 and 5.
  2. Any load-bearing claim supported ONLY by tier 4 or 5, named individually.
  3. For each of those, the tier 1 or 2 source that could replace it.
  4. Any source that appears to be restating another source on the list, which
     would mean the corroboration is circular.

SOURCES (with the claim each one supports):
[paste here]

What these are for

The division of labour assumed here is the one set out in the article below: delegate the mechanical, retain the signature.

Sorting evidence into tiers is mechanical. Deciding a source is trustworthy is not. Restating a title as a falsifiable sentence is mechanical. Knowing which argument your reader needs to hear is not.

An error in the delegated column costs minutes and you notice it. An undetected error in the retained column costs a credibility discount that nobody tells you about.