Skip to content
ENDA

Practical AI

Better answers from AI, with less effort and less waste

Portrait of Thomas Bjerre BojsenBy Thomas Bjerre BojsenDirector at CornerstonesLinkedIn

Most disappointing results from AI come down to a handful of habits. Six short animations show what happens inside the model when you work with it, and what to do differently. Below them are three tools: for seeing where AI is heading, choosing the right model and keeping AI within safe bounds.

Explore the landscapeAI Trend Atlas Pick the right modelAI Model Index Keep AI in boundsThe Guardrail Ring

Tip 01

Brief it like a new colleague

A one-line request leaves the AI to guess what you want, so you get a spread of plausible but wrong answers. A short brief gives it something to aim at.

  • Role who it should act as: “You are our B2B copywriter.”
  • Goal the concrete output: “Write the launch email for Loop.”
  • Constraints length, tone, audience, format, what to avoid.
  • Example something good to match. One example beats a paragraph of adjectives.

Try thisNot sure what to include? Ask the AI to interview you with three questions before it starts.

Tip 02

Keep chats short and focused

The model itself has no memory between messages. Every time you send one, the whole conversation is sent and read again. A chat of eight turns means 36 turns of reading. Even with caching, which makes re-reading cheaper, long threads fill up the context window and get slower and less precise as they near its limit.

  • New topic? Start a new chat. Old context you no longer need is re-read on every turn.
  • Long thread? Ask for a short summary of decisions and open points, then continue in a fresh chat with that summary.
  • Going in circles? Restart with a better brief rather than correcting the tenth answer.

Try this“Summarise this conversation in ten bullet points I can paste into a new chat.”

Tip 03

Let it fetch, don’t paste everything

Pasting a whole spreadsheet or report fills the AI’s working memory with rows it never needed, and the answer you wanted gets buried. When the AI can reach the source itself, it asks for just the row that matters.

  • Connect your sources Use the connectors (Claude), plugins (ChatGPT), file search or database tools your AI assistant offers, so it can look things up.
  • Point, don’t pour Share a link or a file and say where to look, instead of pasting the full text into the chat.
  • Paste with purpose If you must paste, trim it to the part the question is about.

Try this“Look up order #4821 in the orders sheet and tell me what Nordic Tools paid.”

Tip 04

Send scouts for the digging

Research is messy: dozens of documents, searches and dead ends. Do it in your main chat and all of that clutters every turn that follows. Hand it to a separate chat or a sub-agent instead, and bring back only the answer.

  • Split the work Run research in its own chat, project or agent, with a clear question to answer.
  • Ask for the essence Request a short summary with sources, not the full trail of reading.
  • Keep the main line clean Paste only that summary into the chat where you make decisions.

Try thisIn tools with sub-agents, say so outright: “Use a sub-agent to research this and report back in five bullets.”

Tip 05

Get a plan first, then build

Start with the brief from tip 1, but don’t ask for the finished product straight away. Ask for a plan first: an outline, a structure, a design specification. A plan is quick to read and cheap to correct. Fix the plan and the AI builds the right thing; skip it, and every correction means rebuilding, and paying for, the whole thing again.

  • Ask for the plan before the work an outline for a report, a structure for a campaign, a specification for code. Coding tools such as Claude Code, Codex and Cursor have a Plan mode for exactly this.
  • Fix the plan, not the product changing a line in a plan takes seconds; changing a finished document means generating it all again.
  • Then build and fine-tune short, specific feedback, two or three passes. If it’s still off, restart with a better brief (tip 2).

Try this“Before you start, give me a short plan: the sections, the key points and what you’ll leave out. Wait for my OK.”

Tip 06

Match the effort to the task

Many AI tools let you choose how hard the model thinks before it answers. In Claude the setting is called Effort and runs from Low to Max; in ChatGPT you pick Instant, Medium, High or Extra High in the model picker. Higher effort makes it plan, check its work and try more angles. That costs tokens and time: each step up can use several times as many tokens, while the improvement gets smaller. Bigger isn’t better by default; for most tasks a lower level gives the same answer, faster and cheaper. (The numbers in the animation are illustrative; real ones vary by model and task.)

  • Low lookups, translation, reformatting: the answer is known and only needs writing.
  • Medium a good starting point for summaries, reports and everyday work.
  • High debugging, design trade-offs, data analysis: where it pays to check the work.
  • Max stubborn problems and high-stakes decisions only. Expect it to be slow and costly.

Try thisStart lower than you think. If the answer misses, ask why: missing information calls for a better brief or the right source (tips 1 and 3), not more effort.

Tools on this site

Know the landscape, pick the right model, set the guardrails

A network of AI topics grouped into eight coloured clustersExplore the landscapeAI Trend AtlasAn interactive map of AI trends in industry and business, with a focus on European SMEs and mid-market companies. Follow the links between technologies, use cases and the forces driving them.Open the atlas AI models plotted by intelligence against cost, coloured by how open they arePick the right modelAI Model IndexWhich model should you use for the job in front of you? Get a clear pick for 12 everyday jobs, or compare models on price, openness and test results.Open the index The Guardrail Ring: twelve threats around seven checkpoints and eighteen controlsKeep AI in boundsThe Guardrail RingWhich controls keep an AI assistant or agent safe, where they sit on the path of every request, which threats they stop and who owns them. Pick a threat and see every layer that stands in its way.Open the ring