Asking Claude to make a report

Reports are not documents

A document presents what you already know. A report establishes what was unknown — by going to sources, synthesizing them, and recommending an action. The work the agent does in between is what separates a useful report from a confident-sounding essay.

If the agent never opened a browser, it didn’t write a report. It synthesized from training data, which is months stale, often wrong on numbers, and produces the polished-but-ungrounded output that gives AI reports their bad reputation.

The discipline: pin the methodology before the research starts, pin the output shape so the agent synthesizes instead of dumping, and pin the confidence flags so you can tell what to trust.

The reporting loop

Five stages, in order. Skipping any stage produces a different failure mode (see the patterns below).

The most-skipped stage is 5 (confidence). Operators read the report, agree with the recommendation, and never check what the agent was unsure about. The unflagged uncertainty is where the bad decisions hide.

Three ways AI reports fail

The patterns that produce confident-sounding reports you can’t actually act on. Hover any card to see the diagnosis.

The fix in all three: methodology pinned, output shape pinned, reasoning trace required. Without these the agent defaults to opinion-prose; with them it produces a report you can audit.

The methodology check

The diagnostic for reports. Paste the prompt before any research ask. The agent restates the methodology before searching.

Before you start researching, restate the methodology
back to me in this shape:

  Sources: <how many you'll pull, of what kind>
  Synthesis rule: <how you'll combine sources into findings>
  Output shape: <findings count, format, citation rule>
  Scope cap: <when you stop searching>
  Confidence flag: <how you'll express uncertainty per finding>

Wait for my confirmation before searching. If you can't
research it within these constraints, say so — don't start
and improvise.

Common correction at this stage: scope cap. Operators who don’t bound the search say “use as many sources as needed,” the agent uses 40, and the report becomes unreviewable. 8-12 sources is the right size for most reports. Past 12 you’re writing a literature review, not a report.

Annotated transcript

A real conversation: operator asks for an SBA-lending report, agent forces the methodology pin and the situation context, agent self-grades confidence at the end. Click any pinned annotation to expand the “why this turn matters” commentary.

Report brief template — copy and paste

Save in a snippets file. Paste before any report ask. Fill in the topic and your situation; the methodology block is reusable as-is.

Report brief: <one-line topic>

MY SITUATION (so the report is grounded, not generic)
-----------------------------------------------------
- <time horizon, scope, key numbers>
- <current state — what I have, what I don't>
- <constraints — what's off the table, what's required>

METHODOLOGY
-----------
- Use WebSearch for current state — topic is
   time-sensitive, training data is stale
- At least 3 sources per major claim
- Cap at 8-12 sources total
- Quote the source on every number / specific claim
- If a claim is helpful but unsourced, hedge or remove

OUTPUT SHAPE
------------
- 3 findings (one paragraph each)
- 1 recommendation paragraph
- Per-finding confidence flag (high / medium / low)
- One sentence: "what would change my recommendation"
- Source list at the bottom (URL + 1-line description)

RESTATE METHODOLOGY before searching, then research,
then deliver. After delivery, self-grade confidence
and flag the weakest source / claim.

The flip-condition sentence does most of the work. A report that says “recommend X, would flip if Y becomes true” gives you the future trigger to revisit. Reports without it become stale silently.

Prompt kit

Three prompts for briefing, running, and auditing reports. Save in your CLAUDE.md or a personal snippets file.

Before you start researching, restate the methodology
back to me in this shape:

  Sources: <how many you'll pull, of what kind>
  Synthesis rule: <how you'll combine sources into findings>
  Output shape: <findings count, format, citation rule>
  Scope cap: <when you stop searching>
  Confidence flag: <how you'll express uncertainty per finding>

Wait for my confirmation before searching. If you can't
research it within these constraints, say so — don't start
and improvise.

I need a report on: <topic>

My situation (so the report is grounded, not generic):
- <one line>
- <one line>
- <one line>

Methodology:
- Use WebSearch for current state — don't synthesize from
   training alone, the topic is time-sensitive
- Pull at least 3 sources per major claim
- Cap at 8-12 sources total — past that, returns diminish
- Quote the source on every number / quoted claim
- If a claim is helpful but unsourced, hedge or remove

Output shape:
- 3 findings (one paragraph each)
- 1 recommendation paragraph
- Confidence flag per finding (high / medium / low)
- One sentence: "what would change my recommendation"
- Source list at the bottom

Restate the methodology before researching so I can
confirm or correct.

Here's the report: <paste>

Audit it on five axes:
1. Sources — how many cited, how recent, how reputable?
2. Source-to-claim ratio — every number / specific claim
   tied to a source? List any unsourced.
3. Synthesis — do the findings combine sources or just
   restate them?
4. Recommendation — is the reasoning visible? Could you
   re-derive it from the findings?
5. Confidence — does the report flag where it's weak?

For any axis below 4, propose the smallest revision.
Don't redo the report — surgical only.

Apply this — your next report

20-minute exercise. One real question. Pinned methodology. Read the confidence flags carefully.