D1 · Prompting and Task Execution

This domain is roughly 8 of 60 items. It tests whether you can turn a vague business need into a prompt that reliably produces a usable result – and whether you know when to decompose, iterate, add examples, supply documents, or ask for clarification instead of accepting a mediocre first draft. The exam is not testing clever tricks; it is testing disciplined, repeatable prompting habits.

Learning objectives

By the end of this page you should be able to:

Construct a business prompt with the five load-bearing elements: role, context, task, format and constraints.
Decompose a large or multi-part task into ordered sub-tasks.
Iterate on a prompt using targeted follow-ups rather than starting over.
Adapt your prompting strategy to the task type – analysis, research, drafting, brainstorming, summarisation, extraction, translation, planning.
Use examples / few-shot prompting to lock in a pattern.
Specify format and length precisely.
Give Claude documents effectively and reference them in the task.
Recognise and resolve ambiguity by clarifying before generating.

1.1 Anatomy of an effective prompt

A weak prompt states a topic. A strong prompt states a job. Five elements do most of the work; you rarely need all five, but naming them prevents omissions.

Element	Question it answers	Example fragment
Role	From whose expertise should Claude answer?	“You are a procurement analyst…”
Context	What situation and inputs apply?	“We are choosing between two SaaS vendors for a 200-person team…”
Task	What exactly should Claude do?	“Compare them against cost, security and support…”
Format	What shape should the output take?	“…as a two-column table followed by a one-line recommendation.”
Constraints	What bounds apply?	“Under 300 words. Only use the two attached quotes. Flag any missing data.”

Compare:

Exam signal

When a stem shows a one-line topic prompt and a disappointing result, the correct fix is almost always to add the missing element (usually context, format or constraints) – not to change the model, raise a setting, or regenerate blindly.

1.2 Task decomposition

Claude handles a focused task far more reliably than a sprawling one. When a request bundles several deliverables, split it.

Signs you should decompose:

The prompt contains the word “and” joining unrelated deliverables.
The output must pass through stages (research → outline → draft → polish).
Part of the task depends on the result of an earlier part.
The first attempt drops or under-serves one of the parts.

Name the end deliverable and the stages that lead to it.
Prompt one stage at a time, feeding each output into the next. Review between stages so an early error does not propagate.
Keep a stable brief (in a Project or a pinned message) so later stages retain the original context.

Decompose vs. one prompt

Decompose when parts are dependent, when review between stages reduces risk, or when a single prompt keeps dropping a requirement. Keep it as one prompt when the parts are small, independent and low-stakes.

1.3 Iterating with follow-ups

The first output is a draft to refine, not a verdict. Iterate with targeted follow-ups in the same conversation – Claude retains the context, so you change only what is wrong.

Instead of…	Say…
Starting a new chat and re-typing everything	“Keep everything, but make the tone more formal.”
“Make it better”	“Tighten paragraph 2 and cut the last sentence; it repeats the intro.”
Regenerating and hoping	“The total in the table is wrong – recompute row 3 and the sum.”
Accepting a too-long answer	“Cut this to 150 words and lead with the recommendation.”

Good follow-ups are specific about what to keep and what to change. Vague follow-ups (“improve it”) produce random variation.

When to restart instead of iterate

If the conversation has drifted, contradicted itself, or accumulated so much history that Claude is confusing threads, start a fresh chat with a clean, consolidated brief. Endless follow-ups on a polluted context waste effort – this is “output drift”, covered in D7.

1.4 Adapting strategy to the task type

Different task types reward different prompt shapes. The exam expects you to recognise the type and pick the matching lever.

Task type	What to emphasise in the prompt	Common pitfall
Analysis	The criteria/dimensions, the decision to support, the data source	Asking “analyse this” with no criteria
Research	Scope, recency, that sources are required and must be checked	Trusting unsourced claims
Drafting	Audience, tone, length, format, key points to include	Vague “write something about…”
Brainstorming	Quantity, diversity, constraints to respect, “no filtering yet”	Asking for one “best” idea too early
Summarisation	Length, what to keep vs. drop, audience, structure	Not stating target length
Extraction	Exact fields/schema, what to do with missing values	Free-form output that is hard to parse
Translation	Target language, tone/register, glossary of fixed terms	Inconsistent rendering of key terms
Planning	Timeframe, resources, dependencies, output as a table/steps	No timeframe or dependencies

Exam signal

The stem usually names the task type (“draft an email”, “extract the invoice fields”, “brainstorm campaign ideas”). Match the prompt lever to that verb. Answers that ignore the task type – e.g. asking for one idea when the task is brainstorming – are distractors.

1.5 Examples and few-shot prompting

When you need a specific pattern – a tone, a structure, a labelling scheme – showing one or two examples is more reliable than describing it.

You describe the task and hope Claude infers the pattern.

Fine for simple, unambiguous tasks.

You show 1–3 worked examples so the pattern is unmistakable.

Use when the categories are subtle, the format must match exactly, or a zero-shot attempt was inconsistent.

Few-shot examples also fix format drift: if Claude keeps varying the output shape, one perfect example pins it down.

1.6 Specifying format and length

“Be concise” is not a specification. State the shape and the size.

Vague	Specific
“Give me a summary.”	“Give me 5 bullet points, one line each.”
“Make a table.”	“A 3-column table: Risk \| Likelihood \| Mitigation.”
“Keep it short.”	“Under 150 words.” / “Exactly 3 sentences.”
“List the steps.”	“Numbered steps; each starts with a verb.”
“Return the data.”	“Return CSV with columns name,email,tier in that order.”

Being explicit about format is also what makes an output checkable later (D2) and machine-consumable (structured data). If another system or a spreadsheet will consume the output, state the exact schema and column order.

1.7 Giving Claude documents

Claude works best when the source material is in front of it. Paste or attach the document, then anchor the task to it.

Attach or paste the document rather than describing it from memory.
Reference it explicitly: “Using only the attached report…”, “Based on section 3…”.
Constrain to the source when you want grounded answers: “If the document does not state it, say so – do not guess.”
Point to the part that matters for long documents: “Focus on the pricing table on page 4.”
For recurring documents, put them in a Project’s knowledge so every chat can use them without re-pasting (D5).

Exam signal

“Using only the attached…” and “if it is not in the document, say so” are the hallmarks of grounded prompting. Correct answers prefer supplying the document and constraining to it over relying on the model’s general knowledge for facts that live in a specific source.

1.8 Clarifying ambiguity

If the request is ambiguous, resolving the ambiguity before generating beats generating and discovering the mismatch later. You can either ask Claude to surface its assumptions or state them yourself.

Ambiguity	Resolution
Audience unknown	State it: “for non-technical executives”.
Scope unclear	Bound it: “only EMEA, only FY2026”.
Multiple interpretations	Ask Claude: “List your assumptions before you start; I will confirm.”
Missing input	Provide it or tell Claude to flag gaps rather than invent.
Undefined success	Define “good”: “a good answer names the top 3 risks with a mitigation each.”

A useful move: “Before you answer, tell me what you’d need to know to do this well.” Claude will surface the ambiguities, and you resolve them in one exchange rather than three regenerations.

1.9 A repeatable prompting workflow

State the job using role/context/task/format/constraints as needed.
Supply the source material and anchor the task to it.
Decompose if the task bundles dependent deliverables.
Read the first output critically – what is missing or wrong?
Follow up with targeted edits, keeping what works.
Add an example if a pattern will not hold.
Lock the shape with an explicit format and length.

1.10 Prompt structure and delimiters

When a prompt mixes several kinds of content – instructions, a source document, examples, and the actual question – Claude does better when the parts are visibly separated. Business users do not need code, but they should label sections so the model never confuses the material to work on with the instructions about it.

Technique	What it looks like (plain language)	Why it helps
Labelled sections	“INSTRUCTIONS: … / DOCUMENT: … / QUESTION: …”	Stops Claude treating the document text as a command
Quotation fences	Wrap a pasted email in quotes or triple-dashes	Marks where the source starts and ends
Ordered instructions	Number the steps you want, in order	The model follows a sequence instead of guessing priority
Put the ask last	State the material first, the request at the end	The final instruction is the one most reliably honoured
One question per block	Separate unrelated asks	Avoids the model answering only the first

Exam signal

When a stem shows a prompt where the source text and the instruction are jumbled – and Claude “answered the document” or ignored half the ask – the fix is to separate and label the parts and put the instruction last, not to change the model.

1.11 Prompting by business task – a worked gallery

Item writers draw stems from everyday office work. For each task type below, study the before (topic-only) and after (job-specified) prompt and the resulting shift in output.

Task	Before (weak)	After (strong)
Meeting summary	“Summarise the meeting.”	“From these notes, write minutes: Decisions, Action items (owner + due date), Open questions. Bullet form, under 200 words.”
Sales follow-up	“Write a follow-up email.”	“Warm 90-word follow-up to a prospect who liked the demo but worried about price; reference their concern; propose a 20-min call; one clear CTA.”
Policy FAQ	“Make an FAQ.”	“From the attached expenses policy, write 6 FAQ pairs a new joiner would ask; plain English; cite the section after each answer.”
Data request	“Give me the numbers.”	“Return a table: Region \| Q1 \| Q2 \| Growth %, one row per region, ‘n/a’ where a figure is missing, from the attached sheet only.”
Prioritisation	“Help me prioritise.”	“Score these 8 tasks on Impact (1–5) and Effort (1–5) in a table, then list the top 3 quick wins (high impact, low effort).”

Before/after in practice – a job description:

Exam signal

The correct answer names the specific missing elements for that task type (owner+due-date for actions, CTA for a follow-up, schema for a data request). Options that add generic effort (“make it detailed”, “try harder”) are distractors.

1.12 Common misconceptions

Misconception	Reality	Why it matters on the exam
“A more capable model fixes a vague prompt.”	The model cannot supply context or format you never gave it.	The constraint-blind distractor: swapping the model instead of fixing the prompt.
“Longer prompts are always better.”	Relevance beats length; padding adds noise, not specifics.	Tests whether you add the right elements, not more words.
“Regenerating is the same as iterating.”	Regeneration varies wording randomly; iteration changes a named defect.	‘Try again’ / ‘regenerate’ options are wrong.
“One big prompt is more efficient than several.”	Bundled dependent tasks drop requirements and vary in quality.	Decomposition is the correct pattern for multi-part asks.
“‘Be concise’ specifies the length.”	It is not a specification; state a word/sentence/bullet count.	Format-vagueness distractor.
“Few-shot examples are only for developers.”	Business users use worked examples to lock tone, format and subtle categories.	Recognising when to show rather than describe.
“Claude remembers a document from a previous chat.”	New chats do not retain prior context unless it is in a Project/Memory.	Ties D1 grounding to D5 configuration.
“Asking for the ‘best’ idea first is efficient brainstorming.”	It collapses the divergent phase; generate many, then converge.	Task-type mismatch distractor.

1.13 Scenario walkthrough – rescuing a stalled deliverable

Scenario. Priya, an operations associate, needs a supplier-risk briefing for Thursday’s steering meeting. She pastes a 25-page supplier audit and types: “Read this and tell me what’s important, make it good.” Claude returns three paragraphs of generic prose that mention “various risks” and invent a compliance statistic. Priya tries again with “be more thorough and accurate”; the output gets longer but no more useful, and still has no page references. She is tempted to switch to the most expensive model.

Expert reasoning trace.

Name the symptom. Generic, unfocused, one invented figure, no provenance. That maps to ambiguity + missing task/format spec – not a model-capability problem. So switching models (the tempting move) is rejected: a bigger model still lacks a defined job.
Reject “be more thorough/accurate.” These are not actionable instructions (they name a wish, not a specification), which is why the second attempt failed. This is the constraint-blind, “prompt-as-effort” distractor.
Supply the missing job. State the audience (steering committee), the decision it supports (which suppliers to flag for review), the criteria (financial stability, compliance, delivery, concentration), and the format (a table plus a 5-bullet summary, recommendation first).
Anchor to the source and block fabrication. “Use only the attached audit; cite the page for each risk; if a figure is not stated, write ‘not stated’ – do not estimate.” This kills the invented statistic at the root.
Right-size and decompose if needed. If one pass still drops criteria, split: first extract per-supplier risks into a table, review it, then draft the summary from the confirmed table.
Lock it in. Because this briefing recurs monthly, save the instructions and the audit location as a Project so next month starts from the working prompt (D5).

Exam-correct decision: rewrite the prompt with audience, decision, criteria, format and a grounding rule; decompose only if a single pass still drops parts. Not a bigger model, not “be more thorough”, not blind regeneration.

Exam traps in this domain

Trap	Why it is wrong
“Switch to a bigger model to fix a vague prompt”	The prompt is the problem; a bigger model still lacks the missing context/format
“Regenerate until it looks right”	Blind regeneration is not iteration; fix the specific defect with a follow-up
“One mega-prompt is more efficient”	Bundled, dependent tasks drop requirements; decompose and review between stages
“‘Be concise’ specifies the length”	It does not; state a word/sentence/bullet count
“Describe the document to Claude from memory”	Supply the actual document and anchor the task to it
“Ask for the answer, not the format”	Unspecified format is hard to check and to consume; specify shape and schema
“Start a new chat to make a small change”	Follow up in-context; you lose the working context by restarting
“Guess the audience/scope and generate”	Clarify ambiguity first; generating on a wrong assumption wastes iterations
“Jumble the document and the instruction together”	Label and separate the parts and put the ask last, or Claude may ‘answer the document’
“‘Be more thorough / accurate’ will fix it”	These name a wish, not a specification; add concrete criteria, format and grounding
“Pad the prompt with more words to improve it”	Relevance, not length, drives quality; add the right elements, not more text
“Reuse a document Claude saw in a previous chat”	New chats do not retain it; re-supply it or persist it to a Project

Practice questions

Each item states how many responses to select. Attempt before revealing.

Q1 · A manager types 'Write about our new product' and gets a generic, off-target paragraph. What is the BEST fix? (Select one)

A. Switch to a more capable model and resend the same prompt. B. Regenerate several times and keep the best result. C. Add role, context, audience, key points, format and length to the prompt. D. Lower the response length setting.

Answer: C. The output is generic because the prompt is generic – it names a topic, not a job. Supplying the missing elements (who it is for, what to say, how long, what shape) fixes the root cause. A bigger model (A) still lacks the context. Regenerating (B) varies wording, not relevance. A length setting (D) does not address relevance.

Q2 · An analyst asks Claude to 'research competitors, write a brief, draft three headlines, and build a timeline' in one prompt, but the timeline is missing. What is the MOST appropriate approach? (Select one)

A. Repeat the same prompt with ‘and do not forget the timeline’. B. Decompose the request into ordered steps, reviewing each output before feeding it into the next. C. Ask for a longer response. D. Use a cheaper model to save cost.

Answer: B. The request bundles four dependent deliverables, so parts get dropped. Decomposing into stages with review between them is the reliable pattern. Nagging in one prompt (A) is fragile. Length (C) and model tier (D) are unrelated to the omission.

Q3 · Claude produced a solid draft memo but the tone is too casual for the board. What is the BEST next action? (Select one)

A. Start a new conversation and rewrite the brief from scratch. B. Reply ‘make it better’. C. Reply ‘Keep the structure and content; rewrite in a formal tone suitable for a board audience’. D. Regenerate the response.

Answer: C. A targeted follow-up that says what to keep and what to change is efficient and precise. Restarting (A) throws away good context. ‘Make it better’ (B) is vague and yields random change. Blind regeneration (D) may lose the good structure.

Q4 · You need Claude to classify 500 support tickets into three subtle categories and the zero-shot attempt is inconsistent. What is the BEST improvement? (Select one)

A. Ask Claude to ‘try harder’. B. Provide two or three labelled examples per category (few-shot) to pin the pattern. C. Increase the response length. D. Split the tickets into two chats.

Answer: B. Few-shot examples lock in subtle category boundaries and format far better than description. ‘Try harder’ (A) is not actionable. Length (C) does not address category confusion. Splitting chats (D) does not improve consistency.

Q5 · Which prompt will produce the MOST checkable, machine-consumable output for importing into a spreadsheet? (Select one)

A. ‘Summarise the contacts nicely.’ B. ‘List the contacts.’ C. ‘Return CSV with columns name,email,company in that exact order; use “unknown” for missing values.’ D. ‘Give me the contacts as a paragraph.’

Answer: C. A specified schema, column order and a rule for missing values makes the output directly importable and checkable. The others leave shape and completeness undefined.

Q6 · An operations lead wants figures extracted from a 30-page report but is not sure Claude will avoid guessing. Which TWO prompt techniques BEST reduce fabrication? (Select two)

A. Attach the report and instruct ‘use only the attached report’. B. Add ‘if a figure is not stated, write “not stated” — do not estimate’. C. Ask for a longer answer. D. Tell Claude ‘be accurate’. E. Use a more expensive model.

Answer: A and B. Grounding the task to the supplied document and giving an explicit rule for missing values are the two levers that most reduce fabrication. ‘Be accurate’ (D) is not actionable, length (C) is irrelevant, and model tier (E) does not enforce grounding.

Q7 · A user asks Claude to 'summarise this' with no other guidance and gets a summary that is too long and misses the point. What is the BEST fix? (Select one)

A. Specify the audience, target length and what to keep vs. drop. B. Ask Claude to summarise the summary. C. Change the model. D. Paste the document again.

Answer: A. Summaries need a target length, an audience, and a keep/drop rule to be useful. Re-summarising (B) compounds loss, model change (C) is irrelevant, and re-pasting (D) does not add guidance.

Q8 · A request is ambiguous: 'prepare the quarterly update'. Which approach BEST prevents wasted iterations? (Select one)

A. Generate immediately and fix problems afterward. B. Ask Claude to list what it would need to know (audience, scope, format, timeframe) and confirm before generating. C. Generate three versions and pick one. D. Use the longest possible response.

Answer: B. Resolving ambiguity up front – audience, scope, format, timeframe – prevents generating on a wrong assumption. Generating first (A, C) risks solving the wrong problem. Length (D) is unrelated.

Q9 · Which of the following is the STRONGEST example of a well-formed drafting prompt? (Select one)

A. ‘Write a customer email.’ B. ‘Write a warm, plain-language email to a customer whose order shipped late, under 120 words, apologising, giving the new delivery date of 12 May, and offering a 10% discount code.’ C. ‘Write a really good customer email, make it professional.’ D. ‘Draft something for the late order.’

Answer: B. It names audience, tone, length, the situation, the exact facts to include and the call to action – everything needed to produce a usable draft in one pass. The others omit most load-bearing elements.

Q10 · A translation of a policy renders the defined term 'data controller' three different ways. Which prompt technique would have PREVENTED this? (Select one)

A. Ask for a longer translation. B. Supply a glossary of fixed term translations and instruct Claude to use them consistently. C. Use a more expensive model. D. Translate paragraph by paragraph in separate chats.

Answer: B. Translation of documents with defined terms needs a glossary and a consistency instruction. Length (A) and model tier (C) do not enforce term consistency; splitting chats (D) makes consistency harder.

Q11 · During a brainstorming task, a user asks Claude for 'the single best campaign idea' on the first prompt. Why is this suboptimal, and what is the better approach? (Select one)

A. It is fine; always ask for the best idea first. B. Brainstorming rewards quantity and diversity first; ask for 10–15 varied ideas without filtering, then narrow in a follow-up. C. Brainstorming should use extraction formatting. D. Brainstorming needs a lower length limit.

Answer: B. Asking for one ‘best’ idea too early collapses the divergent phase that makes brainstorming valuable. Generate many diverse options first, then converge. Extraction formatting (C) and tighter length (D) work against divergence.

Q12 · A recurring monthly report always needs the same company background and formatting. What is the MOST efficient way to avoid re-pasting it every time? (Select one)

A. Keep a text file on your desktop and paste it each session. B. Store the background and formatting rules as custom instructions and knowledge in a Project so every chat uses them automatically. C. Ask Claude to remember it verbally each time. D. Use a longer prompt each month.

Answer: B. Recurring context belongs in a Project’s custom instructions and knowledge files, so it is applied automatically without re-pasting (see D5). Manual pasting (A, D) is error-prone; verbal ‘remembering’ (C) does not persist across new chats reliably.

Q13 · A prompt says 'Analyse this vendor data' and returns an unfocused essay. Which addition would MOST improve it? (Select one)

A. State the decision it supports and the specific criteria to analyse against, plus the output format. B. Ask for more words. C. Add ‘please analyse thoroughly’. D. Attach an unrelated example.

Answer: A. Analysis needs criteria and a decision to serve, plus a defined output shape. ‘Thoroughly’ (C) and more words (B) do not add focus; an unrelated example (D) misleads.

Q14 · An associate pastes a customer email and types 'reply to this and tell me if we breached SLA' in one run-on line; Claude drafts a reply but never answers the SLA question. What is the BEST fix? (Select one)

A. Switch to a more capable model. B. Label the parts (DOCUMENT / SLA / TASKS), number the two tasks, and put the request after the source so both are answered. C. Regenerate until it answers both. D. Tell Claude to ‘be thorough’.

Answer: B. The two asks were jumbled with the source, so one was dropped; separating and labelling the parts and ordering the tasks fixes the root cause. A bigger model (A) still faces the jumble; regenerating (C) is blind variation; ‘be thorough’ (D) is not actionable.

Q15 · A new joiner asks Claude to 'continue the analysis from yesterday's chat' in a brand-new conversation, and Claude has no idea what they mean. What went wrong and what is the fix? (Select one)

A. The model forgot on purpose; upgrade it. B. New chats do not retain prior context; re-supply the material (or keep it in a Project) so it is available. C. Increase the response length. D. Ask ‘are you sure you remember?’.

Answer: B. A fresh chat has no memory of a previous one unless context is persisted in a Project or Memory; the fix is to supply or persist it. Model tier (A), length (C) and a self-check (D) do not restore missing context.

Q16 · Which TWO additions would MOST improve the prompt 'draft our quarterly newsletter'? (Select two)

A. Name the audience, the 3–4 stories to feature, tone and word count. B. Specify the structure (headline, intro, sections, CTA) and any must-include facts. C. Ask for it to be ‘engaging’. D. Choose a more expensive model. E. Request maximum length.

Answer: A and B. A drafting task needs audience, content, tone, length and structure to produce a usable result. ‘Engaging’ (C) is not actionable, model tier (D) does not add specifics, and maximum length (E) is the wrong lever.

Q17 · An associate needs 15 varied event-theme ideas but keeps getting three safe, similar ones. What prompt change helps MOST? (Select one)

A. Ask for ‘better ideas’. B. Explicitly request 15 distinct ideas across different styles, with no filtering yet, then narrow in a follow-up. C. Switch to Haiku for speed. D. Ask for the single best idea.

Answer: B. Brainstorming rewards explicit quantity and diversity before convergence; stating ‘15 distinct, different styles, no filtering’ widens the divergent phase. ‘Better ideas’ (A) is vague, model tier (C) is irrelevant, and asking for one idea (D) collapses divergence.

Q18 · A prompt asks Claude to 'extract all action items' from meeting notes, but the output misses owners and dates. What is the BEST refinement? (Select one)

A. Specify the exact fields: a table of Action | Owner | Due date, with ‘unassigned’ where no owner is named. B. Ask for a longer answer. C. Use extended thinking. D. Paste the notes twice.

Answer: A. Extraction needs an explicit schema and a rule for missing values, which makes it complete and checkable. Length (B), thinking (C) and re-pasting (D) do not define the missing fields.

Q19 · Claude keeps formatting a recurring report slightly differently each time despite a written description of the format. What is the MOST reliable fix? (Select one)

A. Describe the format in even more words. B. Provide one perfect worked example of the exact layout for Claude to match (few-shot). C. Increase the temperature. D. Ask it to ‘be consistent’.

Answer: B. When a described format will not hold, one concrete example pins the pattern down far more reliably than more description. More words (A) and ‘be consistent’ (D) are not actionable; temperature (C) increases variation.

Q20 · A four-part request (research, brief, headlines, timeline) keeps producing a strong brief but a weak, generic timeline. What is the BEST response? (Select one)

A. Add ‘make the timeline better’. B. Decompose: run the timeline as its own step after the brief, specifying weeks, milestones and dependencies as a table. C. Switch to Opus for the whole thing. D. Ask for a longer overall answer.

Answer: B. Uneven quality on a bundled ask is the overloaded-prompt pattern; isolating the weak part as its own well-specified step fixes it. ‘Make it better’ (A) is vague, a bigger model (C) does not address bundling, and length (D) is the wrong lever.

Key takeaways

A strong prompt states a job, not a topic: role, context, task, format and constraints.
Decompose bundled, dependent tasks and review between stages.
Iterate with targeted follow-ups that say what to keep and what to change – do not restart or regenerate blindly.
Match the prompt lever to the task type; the stem’s verb usually names it.
Use few-shot examples to lock in a subtle pattern or a required format.
Specify format and length explicitly; it makes output checkable and consumable.
Supply documents and anchor the task to them: “use only the attached… ; if not stated, say so.”
Resolve ambiguity before generating rather than after.
Separate and label the parts of a mixed prompt (instructions vs. source vs. question) and put the ask last.
Add the right elements, not more words; “be thorough/accurate” is a wish, not a specification.
New chats do not remember previous ones; re-supply or persist recurring material in a Project.