Prompting Insights

Prompt wording matters, but it matters less than the system around the prompt: shared context, clear intent, access to the right tools, verification, and a feedback loop. A clever sentence cannot compensate for missing project knowledge or an ambiguous definition of success.

Lead with intent

For ambiguous or consequential work, explain what you are trying to achieve before prescribing the implementation:

"I'm considering doing X to achieve Y. Assess the approach, identify the important tradeoffs, and recommend what to do."

That gives the AI room to challenge the route while preserving the goal. For bounded execution where the approach is already settled, a direct command is better. The distinction is not polite versus forceful prompting; it is exploration versus execution.

Give it the operating context

Do not repeat your entire world in every prompt. Give the agent a reliable bootstrap that loads the shared constitution, identity, project memory, and local instructions. The prompt can then focus on what is different about this task.

For long or autonomous work, define:

• the desired outcome
• the boundaries and prohibited actions
• the authoritative sources
• how success will be verified
• what should happen when one part fails
• the final report you expect

Collaborative self-improvement

After finishing a task, I'll often ask: "What should we change so this works better next time?" The useful result is not the reflection itself; it is promoting a durable lesson into the right shared surface: project memory, a skill, a runbook, model policy, or occasionally the constitution.

Without that promotion step, the insight disappears with the conversation.

Define agent-native jobs in plain language

Agent-native scheduled work still needs a strong natural-language instruction:

"Every hour, send me the top 10 latest headlines on [topic] since your last update"

But the prompt is only one field in a reliable job. The job should also define its owner, schedule, executor, delivery route, mutable state, failure behavior, and model role. Be specific about what counts as new and when silence is the correct result.

If the task can be completed deterministically, write a script instead. Natural language is valuable when judgment is required, not as a replacement for ordinary software.