The Language of Human–AI Collaboration
gaiish
Models answer the prompt you actually wrote. gaiish is the practice of writing that prompt well —
saying what you mean, supplying the background the model cannot see, and naming the outcome you
expect.
- Five moves take an idea from your head to a result you can use: clarity, context, intent, precision, result.
- Four things change when you make them: you communicate, collaborate, optimize and empower.
Start with clarity
The technology behind it
Principle 1 of 5
Clarity
Say one thing at a time, in plain words, with no hidden assumptions.
- One request per instruction — a prompt asking for three things gets a shallow attempt at all three.
- Plain words instead of internal shorthand: name the concrete change you want.
- Every fact the answer depends on is stated, or the model invents one.
Read the full page, with examples
Principle 2 of 5
Context
Give the model the documents, constraints and audience it cannot infer.
- Paste the material rather than describing it — a model with your context beats a bigger model without it.
- Name the audience, the constraints and the decisions already made.
- When the corpus is large, retrieval puts the right passage in the prompt for you.
Read the full page, with examples
Embeddings & RAG
Principle 3 of 5
Intent
State the goal and the format, not just the topic.
- A topic gets you an essay; a goal gets you the thing you needed.
- Name the artefact — email, table, patch, JSON — and its length.
- Say what the output will be used for, so the model can judge trade-offs.
Read the full page, with examples
Principle 4 of 5
Precision
Tighten the wording, bound the scope, and say what to leave out.
- Numbers instead of quantifiers: "three bullets, 20 words each", not "keep it short".
- Bound the scope — which file, which period, which product.
- Exclusions are instructions too: say what not to cover.
Read the full page, with examples
Principle 5 of 5
Result
Check the output against the goal, then refine rather than restart.
- Judge the answer against the criteria you set, not against a vague feeling.
- Keep what works and correct the specific part that does not.
- Verify facts, numbers and citations before you rely on them.
Read the full page, with examples
The technology
Generative AI
Generative AI describes models that produce new content — text, images, audio, video and code —
rather than only classifying what already exists. Trained on very large corpora, they learn the
structure of language and media well enough to continue, translate, summarise and invent.
- Transformers and LLMs, diffusion models, multimodal systems.
- Embeddings and retrieval, fine-tuning and alignment, agents and tool use.
- A reference table of the widely used language model families.
Technologies and language models
Transformers & LLMs
Outcome
Communicate
Express your ideas with purpose.
- Most disappointing output is a communication failure, not a capability failure.
- A brief that names audience, length and format usually lands in one attempt.
- Stating the unstated is the cheapest hallucination control there is.
Read the full page, with examples
Clarity
Outcome
Collaborate
Work together with AI smarter.
- Collaboration is what happens after the first answer: draft, then react to specifics.
- Supply the room to work in — documents, data, prior decisions.
- The model generates and does the legwork; you keep taste and judgement.
Read the full page, with examples
Context
Outcome
Optimize
Get better results every time.
- Turn a lucky prompt into a reliable one by measuring it against written criteria.
- Ten to twenty real inputs catch far more regressions than intuition.
- Change one variable at a time — wording, context, model — or you learn nothing.
Read the full page, with examples
Precision
Outcome
Empower
Augment human potential with AI.
- The model removes the work that never needed a human: drafting, summarising, legwork.
- With tools it can search, query, run and file — not just answer.
- Judgement, taste and accountability stay with the person who owns the outcome.
Read the full page, with examples
Agents & Tool Use
