The instructions worth keeping

The same correction, made twice, is a good reason to write something down.

For Leaf Lane, that might be a reminder to keep the source link beside a claim, preserve my rough notes, or save an article as a draft for me to review. If those instructions stay in a chat, I have to remember to bring them into the next one.

A skill gives that instruction a place to live. It is a small package of directions for a recurring job, with examples or scripts when they help. OpenAI’s skills documentation describes the format; the useful part is deciding what belongs in it.

Give each skill a clear job

I’ve split the Leaf Lane content workflow into writing, images, and saving drafts. The writing skill handles structure, sources, and the edit for generic AI language. The image skill handles the visual idea, crop, and image description. The content manager saves the work into the site and checks that it is there.

They share an editorial guide. That matters because a rule copied into several places can drift: one version says to preserve the original notes, another forgets them, and a third has an old description of the editor.

The handoffs are concrete. Writing produces a title, excerpt, body, sources, and notes about anything still uncertain. Images produces a usable file with a description and a record of how it was made. Saving produces a draft I can open, read, and edit.

Publication stays with me.

Fix the instruction behind the correction

Suppose a draft loses a source link. Adding the link back fixes that article. The next run needs something more specific: keep the source beside the claim it supports, and check the rendered draft before handing it over.

The same applies to an image that crops badly on a phone. Keep the visual subject away from the edges and inspect the small version. “Make a better image” gives the next agent very little to work with.

A useful instruction names the input, the expected result, and the check that tells you the work is finished. It also says what to do when something is missing. An agent should leave a factual question for review rather than fill the gap with a plausible story.

Start with one repeated job

You can use the same approach for preparing an estimate or assembling a weekly report. Write down the details someone otherwise has to explain each time: where the inputs come from, what the output contains, and who checks it.

Try it on a few real examples with someone reviewing the results. When a correction repeats, improve the instruction. When the task changes, update it.

That is the test I care about: can the next run use what the last run taught us, without someone having to remember the whole conversation?