How to give AI better context: write down why you decided

Six · 9 min read · July 16, 2026

How to give AI better context: write down why you decided

Say your company decided a while back to prefer annual agreements for recurring services. The finance lead put it into the purchasing guidelines as one line: "Prefer annual agreements for recurring services."

Now a team lead wants to try a new scheduling tool for three months before committing to it. She asks an AI assistant to draft the purchase request. It finds the guideline, quotes it, and recommends signing up for a year.

The assistant read the right document. What it didn't have was the reason behind the rule. Annual terms made sense for services the company already knew it would use all year, because the terms were predictable and the need was settled. A three-month trial is exactly the case the rule was never meant to cover. Nobody wrote that down, because everyone in the room when the decision was made already knew it.

The usual advice is to give the AI more context. Sometimes that's the fix. Here it wouldn't help much. Ten more procurement files won't tell the assistant that the rule stops at trials if none of them say so.

First, work out what was missing

Before you add anything, figure out which kind of gap you're looking at. Each one has a different fix, and only one of them calls for writing down more of your reasoning. Walk the trial request through these questions in order:

  1. Could the AI reach the source it needed? Here it could, because it quoted the guideline. If it couldn't, the fix is access.
  2. Is that source current, and is it the one that governs? If finance replaced the guideline last quarter, or a vendor policy takes priority over it, the person who owns purchasing has to say which one counts.
  3. Do two sources disagree? If the guideline says annual and a department budget memo says nothing longer than three months without sign-off, someone has to settle that. An AI picking between them on its own is guessing.
  4. Does the source state the conclusion without the judgment around it? This is the trial problem. The rule was there. Its purpose and its limits weren't.
  5. Was everything already there, and the AI still got it wrong? Then stop adding documents and look at the instructions it was given, what context it was handed and how its work was checked.

Anthropic's engineering guide to context for AI agents pushes back on the "just add more" reflex. Models recall information less reliably as their context fills up, so the guide's advice is to find "the smallest possible set of high-signal tokens," adding that minimal doesn't necessarily mean short. For the trial request, the high-signal piece is one sentence about where the rule stops, and it isn't in any file yet.

A paper presented at ICLR 2025, "Sufficient Context", drew the line that separates question 5 from the four before it. Its tests used question-answering datasets, not company records, so the comparison is ours. The researchers sorted cases by whether the retrieved text held enough to answer. When the text was enough, smaller models often made something up or declined anyway, which is question 5. When it wasn't, the larger models they tested often gave a wrong answer instead of declining. That's what the assistant did with the trial: the guideline didn't cover it, and it answered anyway.

What the guideline kept, and what it lost

The one-line guideline kept the conclusion and lost most of what the finance lead knew when making it. The point was predictable terms for services already in steady use. Monthly terms were considered and turned down for those services, because once the need was settled, predictability mattered more than flexibility. Trials and one-time purchases were never meant to be covered. The finance lead could approve an exception, and a real change in how the company used its tools would be a reason to look at the rule again.

Software teams have dealt with this for a long time. In 2011, Michael Nygard wrote that "one of the hardest things to track during the life of a project is the motivation behind certain decisions." Someone who arrives later without that motivation, he said, can only blindly accept the decision or blindly change it. An AI working from the one-line guideline is in the same position. With the trial request, it blindly accepted.

A short record of why

We use a short record for this and call it a Decision DNA card. The idea isn't new. It's a close relative of the Architecture Decision Record Nygard described, which holds a decision's context, the decision itself, its status and its consequences. Martin Fowler's write-up on ADRs adds one piece the purchasing guideline lost: a record should "explicitly list all the serious alternatives that were considered, together with their pros and cons." If your team already writes ADRs, keep doing that. The card takes the same habit to purchasing, pricing, hiring or which customers you serve.

It has seven fields:

  1. Decision: one choice, not a broad topic. What question was answered, and what did you choose?
  2. Purpose: what outcome, constraint or risk was the choice meant to handle?
  3. Evidence: which sources or observations mattered, and what was still uncertain?
  4. Trade-offs: what did you optimize for, and what cost did you accept?
  5. Alternatives: which option did someone seriously consider and turn down, and why?
  6. Boundaries: where, when and for whom does the decision apply, and where doesn't it?
  7. Authority and review: who owns it, who can override it, which source controls, and what should reopen it?

Here's a version to copy:

Decision:
Purpose:
Evidence and uncertainty:
Trade-offs:
Alternatives:
Boundaries:
Authority and review:
Controlling source link:

Keep the card short and link the source documents instead of pasting them in, so the card holds the judgment and the guideline stays where it lives. If nobody knows a reason or a boundary, write "unknown" and go ask. A plausible explanation someone filled in to complete the form will be read by the next person, and the next AI, as fact.

The card for the annual-agreement rule

Here's what the finance lead might write for the guideline.

Decision: For recurring services we already rely on, prefer an annual agreement. For a new service, prefer one once a trial has shown we need it.

Purpose: Predictable terms for services we know we'll use all year.

Evidence and uncertainty: Last year's renewals, where the services we already used ran all year. We don't know yet how many new tools outlast their trial.

Trade-offs: We give up flexibility in exchange for predictable terms.

Alternatives: Monthly or short terms. Turned down for established services, but still the right choice when demand is uncertain.

Boundaries: Applies to services in established, ongoing use. Trials, one-time purchases and anything with unsettled requirements are outside it.

Authority and review: The finance lead owns this and can approve exceptions. Reopen it if our usage, a vendor's terms or our risk changes materially.

Controlling source link: (the purchasing guideline)

Run the trial request again with the card linked. The assistant still finds the guideline. Now it can also see that trials sit outside it, that short terms are the option kept for uncertain demand, and that the finance lead approves exceptions. It has what it needs to propose a three-month term, note that the annual preference applies if the team keeps the tool, and flag anything it's unsure about for the finance lead.

Keep the reasons open to question

A card doesn't make the original decision right, and it can go stale like any other document. Say the company moves most of its tools to monthly plans and nobody updates the card. Its reasons keep the annual rule looking justified long after the situation behind them has changed.

There's a second risk. An AI working from your written reasons may stay inside them. A small study presented at CHI 2025, "AI, Help Me Think, but for Myself", points to it. Twenty-one people built an investment portfolio, a task with generally no objective right or wrong answer, using two AI assistants. One built on the reasons each person had written down. The other made recommendations. The first fit better into people's thinking and led to slightly better outcomes. The recommender offered more new insights and took less effort. The study looked at people making one decision, not at company records, but its trade-off is the one a card has to plan for.

That's why the card keeps the losing alternative and the review condition. Short terms aren't rejected forever. They're on record as the right choice when demand is uncertain, and a change in how the company uses its tools is on record as a reason to reopen the rule. So when the assistant is asked about another annual renewal after that move to monthly plans, it has grounds to point out the change to the finance lead, rather than quietly following the old rule or quietly overturning it.

Name who decides

A link to the guideline shows where the rule came from. It doesn't show whether the rule still governs, which exception takes priority, or who can override it. People have to answer those questions, and the card is where their answers get written down.

For the annual-agreement rule, the card names the finance lead as the owner and the person who approves exceptions. It names the purchasing guideline as the source that controls, so when a budget memo says something different, nobody has to guess which one wins, and neither does the AI. And it says what should reopen the decision: a material change in how the company uses its tools, in a vendor's terms or in risk. The card holds those answers, but someone still has to watch for the change and run the review.

When a card is worth writing

Write one for a decision that recurs or matters, that other people and their AI tools will keep reusing, and where knowing the reason would change how the rule gets applied. The annual-agreement rule qualifies, because it comes up every time someone buys software.

Skip it for one-off choices and anything easy to reverse. Skip it for the other gaps from the start of this piece too: a missing fact needs the source, and a conflict needs a decision before anything gets recorded. If an existing ADR or policy note already captures the reasoning, use that instead of writing a second record. And if nobody can state the reason without guessing, don't write a card yet. Mark it unknown and ask the person who made the call.

Start with one decision

You don't need to convert an archive. Pick one rule your team, or your AI tools, apply often.

  1. Choose one recurring or consequential decision that people keep reusing.
  2. The next time an AI gets it wrong, check which gap it is: access, an outdated or overruled source, a conflict, missing judgment, or a record that was fine and used badly.
  3. If it isn't missing judgment, fix the gap you actually found.
  4. Ask the person who owns the decision what they know. Don't record a guess.
  5. Fill in the seven fields and link the controlling source.
  6. Use the card in one real request, then check whether the boundary and the owner came through. Revise the card, or drop it if it didn't help.

For the purchasing guideline, that's one conversation with the finance lead and one rerun of the trial request.

Say your team's purchase requests ran on Supanova. You'd write a project brief and put the card's boundary and owner into it: annual terms are for services already in steady use, trials stay short, and the finance lead approves exceptions. Your atoms do the work from that brief, and the finished work comes back for your review. That's where you check whether the trial was treated as a trial, and whether the card needs another line.

Now think of the rule your team quotes most often. Could anyone write down where it stops and who can make an exception? If not, that's your first card.

You can see how Supanova works at supanova.team.

Sources

AI ContextOrganizational KnowledgeDecision DNADecision Records