Commence with the End in Mind

Put the description of the finished object in its own tag - <finished_output> or <target_artifact> - separate from <source_material>, <constraints> and the <task> block that comes last. Then use the end description as something to audit against: ask for a claims audit, a scansion note, a check on over-certainty. Claude will report honestly that four lines do not scan or that a bullet overstated a hedge, which is the real payoff of having defined the target precisely.

• Give the finished object its own tag, separate from the source.
• Material first, <task> last.
• Ask it to audit the output against your stated target.
• Ask where your brief's own tensions pushed the draft.
• Ask for the weakest line, named.

<source_material>
This is a transcript of an interview with a leading climate scientist. It covers rising sea levels, carbon capture technology, and policy recommendations. It is spoken speech, so there are digressions, false starts and interviewer questions of no substance.

<transcript>
[Paste 3000-word interview transcript]
</transcript>
</source_material>

<finished_output>
What I need at the end: exactly five bullet points, each a complete sentence of 25 to 35 words, serious in tone. The test they must pass is that someone who has not read the interview could, from the five bullets alone, explain the scientist's position to a colleague accurately. Not a list of topics touched on - the positions taken.
</finished_output>

<method>
Read the whole transcript before writing anything. Find the scientist's strongest claim on each of the three threads, then the two most significant things said outside them. Do not divide the five bullets evenly across the threads if the interview does not weight them evenly - follow the interview.
</method>

<task>
Produce the summary.

Before the bullets, give me one paragraph under <the_through_line>: what you take the scientist's central position to be. If my five bullets are going to misrepresent them, it will show up here first, where I can catch it.

Then <summary> with the five bullets.

Then <left_out>: the most significant thing you dropped, and why.

Then <over_certainty>: anything the scientist said tentatively that your bullets state more firmly than they did. Scientists hedge, summaries flatten, and the flattening is the actual risk in this task - so check your own bullets against the transcript's hedging and report what you find even if the answer is that you overstated something.
</task>

<target_artifact>
A 14-line sonnet in Shakespearean form: three quatrains and a couplet, rhyming ABAB CDCD EFEF GG, in iambic pentameter. Early modern diction throughout - thou, thy, doth - used consistently, not sprinkled for flavour. The volta at line 9. The closing couplet must resolve the poem's conceit rather than restate the opening.

It should survive being read aloud by someone who knows the form.
</target_artifact>

<subject>
A modern smartphone, as a marvel worth praising. Two features to celebrate: the camera as an "all-seeing eye", and the long-lasting battery as a "tireless heart".
</subject>

<constraints>
Shakespeare had no word for a phone, a battery, a photograph or a screen. Those words may not appear. The device must be rendered entirely through metaphor and period-plausible language. If you reach for "device", the persona has already failed.
</constraints>

<task>
Write the sonnet.

Then, outside the poem:

<scansion>Go line by line and name every line where the iambic pentameter is imperfect or forced. Be exacting here. Presenting a metrically rough poem as clean is the failure I am trying to avoid, and I would rather have four honest problems named than a claim that all fourteen lines scan.</scansion>

<anachronism_check>Any word or concept in your own poem that Shakespeare could not have used.</anachronism_check>

<weakest_line>The one line you would rewrite, and what is wrong with it.</weakest_line>

If the form and the subject genuinely fight each other at some point - if a tireless heart cannot be got into iambic pentameter without strain - say where, rather than forcing it and hoping I do not read closely.
</task>

<role>
You are a financial journalist writing for a business publication.
</role>

<data>
<company>Innovate Hyderabad</company>
<quarter>Q3 2025</quarter>
<units_sold>50,000</units_sold>
<qoq_growth>+20% against Q2</qoq_growth>
<competitor name="Zenith Corp">45,000 units</competitor>
</data>

<finished_output>
A news report of approximately 150 words. Optimistic in tone, factual in substance - the optimism must come from how the figures are framed, not from adjectives added to them. Structured as a journalist would: a lead carrying the news, supporting detail, then competitive context. Headline above it.
</finished_output>

<data_discipline>
- Use nothing outside <data>. No invented revenue, market share, headcount, product names or executive quotes. Where a real report would carry a quote, leave a marked placeholder.
- Q2 units are not in the data. Do not state or infer them.
- "Market leadership" is not supported. 50,000 units against one named competitor's 45,000 tells us nothing about the market as a whole. You may write that the company outsold Zenith Corp this quarter. You may not write that it leads the market.
</data_discipline>

<task>
Write the headline and report.

Then <claims_audit>: list every factual claim in your report alongside the data point it rests on. Any claim without a data point behind it, flag as unsupported.

Then <where_the_brief_pushed_me>: I asked for optimism and gave you data that does not support the strongest optimistic reading. Tell me where that pressure showed - which sentence you were tempted to overstate, and what the overstated version would have said. I would rather see the temptation named than find it in the copy.

Keep the report in an Artifact; I will revise the wording across a few turns.
</task>