Best Practices

Connect high-signal sources first

Not all sources are equally useful. Prioritize the ones where you spend decision-making time:

• Email (Gmail or Outlook) is usually the richest single source — it captures your relationships, commitments, and priorities.
• GitHub or GitLab gives Jackdaw a precise picture of your technical work.
• Linear or Todoist captures what you're actively working on.

Add health and lifestyle sources like Strava once the work layer is established.

Capture constraints, not just preferences

Preferences are things you usually want. Constraints are things you always need — and the distinction matters for agents:

• "I prefer TypeScript" → a preference; an agent may use something else.
• "Never use any in TypeScript — it's a project rule" → a constraint; the agent should follow it every time.

The easiest way to capture either is to tell the assistant: "Never generate SQL without parameterized queries." Jackdaw records it as a typed fact — a constraint — and surfaces it to future work when relevant. Direct, non-conflicting facts can land immediately; identity details and changes that conflict with the current state wait in your inbox so you can see what would be replaced.

Correct early, correct explicitly

When an AI gets something wrong about you — wrong employer, wrong stack, wrong preference — correct it immediately and specifically:

"That's wrong. I'm not using React — this project is all SvelteKit."

Jackdaw captures the correction as a typed fact that overrides the original. Vague corrections ("that's not right") don't produce reliable updates — be specific.

You do not have to go through the assistant, either. Open the relevant Wiki page and edit the fact directly — inline edits save as you go. Telling the assistant is fastest in the flow of a conversation; editing the page yourself is best when you are reviewing the current state and spot something wrong.

Review your inbox

Inferred claims, identity details, corrections, and AI-proposed facts follow the inbox review rules before they change the Wiki. Check it regularly — it is how you keep the current state trustworthy. An approved fact becomes available to later people and agents; a rejected one is logged so the same claim does not keep coming back.

Keep the wiki readable

The wiki is read by AI, but authored by you. Write it the way you'd brief a sharp new colleague:

• Use short, declarative sentences.
• Link people and projects with [[wikilinks]] — this builds a graph, not just a list.

A cluttered wiki with contradictory facts is worse than a sparse one. Less, but accurate, beats more, but noisy.