Sources

Knowledge base of 2026-11-12 · 28 techniques in use · 0 retired

When First Try rewrites your draft it applies the techniques on this page and no others. Each one lists where it comes from. The score is a formula, not an opinion: it starts from the strongest source (our own tests 0.90, a research paper 0.70, a vendor's guide 0.60, a house rule 0.45, a community source 0.30), rises by 0.10 for each further independent publisher, falls by 0.20 for each source that contradicts it, and is not used below 0.40.

We keep findings in our own words with a short quotation and a link. We do not copy source documents.

Wording added for the kind of task

Be specific about the output wanted; when more than the minimum is wanted, ask for it explicitly.

Evidence: Strong (0.80) · other · any model

fabric (community), improve_prompt pattern (community) In order to get a highly relevant response, make sure that requests provide any important details or context.
Google, Prompt design strategies (Gemini API) (vendor guide) If you need a more conversational or detailed response, you must explicitly request it in your instructions.
Anthropic, Prompting best practices (vendor guide). Advises asking explicitly when more than the task strictly requires is wanted.

Produce distinct alternatives and name the idea behind each in a few words.

Evidence: Limited (0.45) · creative · any model

First Try, First Try house rule (no outside source yet) (house rule)

State the reasoning behind the main conclusion in two or three lines so it can be checked.

Evidence: Limited (0.45) · data_analysis · any model

First Try, First Try house rule (no outside source yet) (house rule)

Include as many relevant features and interactions as possible. Go beyond the basics to create a fully-featured implementation.

Evidence: Good (0.60) · frontend_design · any model

Anthropic, Prompting best practices (vendor guide) Include as many relevant features and interactions as possible. Go beyond the basics

Avoid the generic AI aesthetic: overused fonts (Inter, Roboto, Arial, system fonts), purple-gradient-on-white colour schemes, identical rounded cards, predictable layouts. Commit to one cohesive aesthetic with dominant colours and sharp accents.

Evidence: Good (0.60) · frontend_design · any model

Anthropic, Prompting best practices (vendor guide). The frontend design section names these defaults and asks for one committed aesthetic.

Explain at the stated level: adjust depth, never accuracy.

Evidence: Good (0.70) · research_learning · any model

arXiv, Principled Instructions Are All You Need (Bsharat, Myrzakhan & Shen) (research paper). Principle 5: ask for an explanation at a stated level of the reader.

Teach, then test: include a short self-check without giving the answers up front.

Evidence: Good (0.70) · research_learning · any model

arXiv, Principled Instructions Are All You Need (Bsharat, Myrzakhan & Shen) (research paper). Principle 15: teach the topic and include a test, without giving the answers first.

Be specific: amounts, durations and schedules, not general advice.

Evidence: Limited (0.45) · personal_life · any model

First Try, First Try house rule (no outside source yet) (house rule)

Give the one-line reasoning behind each recommendation so it can be challenged.

Evidence: Limited (0.45) · marketing_sales · any model

First Try, First Try house rule (no outside source yet) (house rule)

State the audience explicitly and write for them.

Evidence: Good (0.70) · marketing_sales · any model

arXiv, Principled Instructions Are All You Need (Bsharat, Myrzakhan & Shen) (research paper). Principle 2: integrate the intended audience in the prompt.
arXiv, Less Back-and-Forth: A Comparative Study of Structured Prompting (research paper). Audience was one of three checklist items in prompts that scored 7.50 against 5.67 raw (0–8 rubric).

Give the reasoning behind the recommended option in two or three lines, including what was ruled out.

Evidence: Limited (0.45) · planning_strategy · any model

First Try, First Try house rule (no outside source yet) (house rule)

Implement the change rather than describing it; when something is ambiguous, state the assumption and continue.

Evidence: Good (0.70) · software · any model

First Try, First Try house rule (no outside source yet) (house rule)
Anthropic, Prompting best practices (vendor guide) Change this function to improve its performance.

Solve the general case: do not special-case tests or hard-code expected values to make them pass.

Evidence: Good (0.70) · software · any model

First Try, First Try house rule (no outside source yet) (house rule)
Anthropic, Prompting best practices (vendor guide) Implement a solution that works correctly for all valid inputs, not just the test cases.

Write for the stated reader and build on what they already know.

Evidence: Good (0.70) · writing · any model

arXiv, Principled Instructions Are All You Need (Bsharat, Myrzakhan & Shen) (research paper). Principle 2: integrate the intended audience in the prompt. The authors report +57.7% quality on GPT-4 across the 26 principles together.

Match the user’s own voice and wording where it is given.

Evidence: Good (0.70) · writing · any model

arXiv, Principled Instructions Are All You Need (Bsharat, Myrzakhan & Shen) (research paper). Principle 26: match the language and style of a provided sample.

Rules built into every rewrite

Phrase constraints as what to do, not what to avoid.

Evidence: Strong (0.80) · any task · any model

arXiv, Principled Instructions Are All You Need (Bsharat, Myrzakhan & Shen) (research paper). Principle 4: use affirmative directives.
Anthropic, Prompting best practices (vendor guide). Advises telling the model what to do instead of what not to do.

Give four to eight checkable quality criteria drawn from the checklist for this kind of task.

Evidence: Good (0.70) · any task · any model

arXiv, Principled Instructions Are All You Need (Bsharat, Myrzakhan & Shen) (research paper). Principle 25: clearly state the requirements the model must follow.
arXiv, Less Back-and-Forth: A Comparative Study of Structured Prompting (research paper). Checklist-improved prompts scored highest (7.50 of 8) and used the fewest tokens.

Ask a clarifying question only when the draft is ambiguous; never on a clear one.

Evidence: Strong (0.95) · any task · any model

First Try, First Try eval, 37 conversations, 9 November 2026 (our own test). The gate asked on 60% of ambiguous drafts and on 0% of clear ones (37 cases).
arXiv, CLAM: Selective Clarification for Ambiguous Questions (Kuhn et al.) (research paper). Clarification helps on ambiguous inputs and penalises unambiguous ones.
arXiv, Less Back-and-Forth: A Comparative Study of Structured Prompting (research paper). Clarifying-question prompts scored 6.67/8 vs 7.50 for checklist prompts, and used more tokens. Checklist prompts also produced the best quality-effort tradeoff, using fewer average tokens than both raw and clarifying prompts.

Say who the result is for and carry over the relevant facts from the conversation.

Evidence: Strong (0.80) · any task · any model

arXiv, Principled Instructions Are All You Need (Bsharat, Myrzakhan & Shen) (research paper). Principle 2: integrate the intended audience.
arXiv, Conversational User-AI Intervention: A Study on Prompt Rewriting (Sarkar et al.) (research paper). Rewriting with conversation history improves responses, more so in longer conversations.
arXiv, Less Back-and-Forth: A Comparative Study of Structured Prompting (research paper). Checklist prompts covering context and audience scored 7.50 against 5.67 raw (0–8 rubric). Small study, single rater.
OpenAI, Prompt engineering guide (vendor guide) It is often useful to include additional context information the model can use to generate a response within the prompt you give the model.

End with two to four checkable criteria for when the work is done.

Evidence: Good (0.70) · any task · any model

OpenAI, Prompt generation (vendor guide). Steps and explicit requirements keep multi-step work from skipping parts.
Anthropic, Prompt engineering overview (vendor guide). Success criteria are set before prompt engineering starts.

Name the exact output: format, length, language and deliverable.

Evidence: Good (0.70) · any task · any model

OpenAI, Prompt generation (vendor guide). The meta-prompt always carries an Output Format section.
Anthropic, Prompting best practices (vendor guide). Advises being specific about the output format and constraints.
Anthropic, Improve your prompts in the developer console (vendor guide). Vendor-reported: 100% word-count adherence on summaries and +30% accuracy on a multilabel classification test after the improver enforced format.

State the goal in one sentence and say why it matters.

Evidence: Good (0.60) · any task · any model

Anthropic, Prompting best practices (vendor guide). Clear, explicit instructions and the motivation behind them give more targeted responses.

Split a long prompt into labelled sections or tags.

Evidence: Strong (0.90) · any task · any model

fabric (community), improve_prompt pattern (community) Delimiters like triple quotation marks, XML tags, section titles, etc. can help demarcate sections of text to be treated differently.
arXiv, Principled Instructions Are All You Need (Bsharat, Myrzakhan & Shen) (research paper). Principles 8 and 17: sections and delimiters.
Google, Prompt design strategies (Gemini API) (vendor guide) Employ clear delimiters to separate different parts of your prompt.
OpenAI, Prompt engineering guide (vendor guide). Recommends Markdown or XML delimiters and an order of identity, instructions, examples, context.
OpenAI, Prompt generation (vendor guide) [Optional sections with headings or bullet points for detailed steps.]
Anthropic, Prompting best practices (vendor guide). XML tags help the model parse complex prompts unambiguously.

Never turn a gap into an invented fact: mark each default (assumed), list the assumptions once, and leave unknowns as [placeholders].

Evidence: Strong (0.90) · any task · any model

First Try, First Try house rule (no outside source yet) (house rule)
arXiv, Conversational User-AI Intervention: A Study on Prompt Rewriting (Sarkar et al.) (research paper) LLMs often need to -- and inherently do -- make \emph{plausible} assumptions about a user's intentions and goals when interpreting prompts
OpenAI, Prompt generation (vendor guide). The meta-prompt preserves user content and does not add facts.

Do not add "think step by step"; ask instead for the reasoning behind the main recommendation.

Evidence: Strong (0.90) · any task · any model

arXiv, Better Zero-Shot Reasoning with Role-Play Prompting (Kong et al.) (research paper) role-play prompting acts as a more effective trigger for the CoT process
Google, Prompt design strategies (Gemini API) (vendor guide) it's generally not necessary to have the model outline, plan, or detail reasoning steps in the returned response itself.
OpenAI, Prompt engineering guide (vendor guide). Reasoning models do better with high-level goals than with prescribed thinking steps.
Anthropic, Prompting best practices (vendor guide) A prompt such as “think thoroughly” often produces better reasoning than a hand-written step-by-step plan.

Include examples only when the user supplied them, clearly delimited; otherwise ask for one with a placeholder.

Evidence: Strong (0.80) · any task · any model

arXiv, The Prompt Report (Schulhoff et al.) (research paper). Few-shot prompting was among the most consistently effective of 58 techniques surveyed.
Anthropic, Prompting best practices (vendor guide). Calls examples one of the most reliable ways to steer format, tone and structure.

Open with one or two sentences naming the specific practitioner for this task: specialty, what they are known for, what they judge by.

Evidence: Limited (0.40) · any task · any model · Not for: Factual lookups and follow-ups: personas do not improve factual accuracy.

arXiv, ExpertPrompting (Xu et al.) (research paper). Answers from a task-specific expert identity were preferred 48.5% against 23% for vanilla prompting (GPT-4 judge, 500 samples); a generic "imagine you are an expert" line gave no gain.
arXiv, Better Zero-Shot Reasoning with Role-Play Prompting (Kong et al.) (research paper). Role-play lifted ChatGPT on 12 reasoning benchmarks, e.g. AQuA 53.5% to 63.8%.
Anthropic, Prompting best practices (vendor guide). Recommends giving the model a role.
Against: arXiv, Is "A Helpful Assistant" the Best Role for Large Language Models? (Zheng et al.) (research paper). 162 roles across 2,410 factual questions: no or slightly negative effect on accuracy.
Against: arXiv, Prompting Science Report 4: Playing Pretend (Wharton Generative AI Labs) (research paper). In-domain expert personas had no significant impact on GPQA Diamond and MMLU-Pro accuracy across six models (except Gemini 2.0 Flash). Expert personas showed no consistent benefit across models, with few exceptions.

Every line must be checkable or change the answer; do not repeat context.

Evidence: Strong (0.80) · any task · any model

arXiv, The Prompt Report (Schulhoff et al.) (research paper). Duplicated context hurt accuracy in a case study.
arXiv, Less Back-and-Forth: A Comparative Study of Structured Prompting (research paper). The highest-scoring prompts were also the shortest.
OpenAI, Prompt generation (vendor guide) Clarity and Conciseness: Use clear, specific language. Avoid unnecessary instructions or bland statements.

House rules

Entries marked "house rule" come from our own practice and have no outside source yet. They carry the lowest score that is still used, and they are the first candidates for testing.

Our own test

We send the same requests to an AI twice, once as typed and once rewritten, and a separate judge compares the two answers without knowing which is which. The current figures are on the home page.

Retired

Nothing has been retired yet. When advice is contradicted, goes out of date or lowers results in our own test, it moves here with the reason.