Back to storiesPractical guides

Package a content workflow as a Skill and try a different source

Extract a short Skill from an actual content workflow, explicitly load it for a different source, and review factual scope and requested formats. Includes the skill and real artifacts.

8 min readMarkdown
On this page7 sections

A different source reveals what is reusable

Our previous task adapted a research-assistant article into an introduction, cards and a voice script. This time we saved the method as a short Skill and applied it to a different article about reviewable quote drafts.

The new output is one introduction and two cards. It retains the new source’s fictional-example status, subtotal basis and unresolved items without carrying over old topic-selection facts.

The work took place in Codex through explicit reading of the skill body. We did not install a global skill, test automatic discovery or run an independent model comparison. The pack includes the request, skill, actual drafts and review.

Keep the stable decisions from the earlier task

The previous channel-writing task produced three Chinese drafts and three English adaptations. We retained three decisions: use the current source for facts, give each format a distinct reader task, and distinguish drafts and proposed visuals from completed actions.

“Always create three formats” did not become a rule. That was one assignment’s scope. Its six sources, 62-article inventory and selected topics also belong to the old material.

The new skill is named reviewed-channel-drafts. Its description identifies adaptation of a supplied article while preserving factual scope and links. It does not take responsibility for research, scheduling or publication.

A short file with clear boundaries

Download the skill and reuse pack. reviewed-channel-drafts/SKILL.md holds the method; method-source.md is the earlier article; new-source.md is the new input. The drafts and review record are actual outputs.

The instructions ask the assistant to read the current source and audience, ask for missing essential input, preserve evidence status, generate only requested formats, and save source links and a review. Drafting alone grants no permission to send or publish.

Product accounts and machine-specific directories are not prerequisites in the skill. New assignments replace the inputs rather than copy a second, gradually diverging method.

The new request asks for only two formats

We read the complete public Markdown of the field-notes-to-quote tutorial. Its rental data is fictional teaching material. The English version explains the same exercise.

Read reviewed-channel-drafts/SKILL.md and apply it to new-source.md.
Write for small-team leads organizing field notes into a reviewable draft.
Produce one newsletter introduction and two social cards, no video.
Preserve the source’s fictional teaching-example status.
Save drafts and a review; do not send or publish.

Explicit file loading keeps this exercise separate from automatic skill discovery. Installation scope and command registration need their own host-specific checks. Passing file-format validation does not establish those behaviors.

What the actual output retained

The introduction starts with ambiguous equipment quantities, a later confirmation, and unresolved transport. It describes what the tutorial teaches and identifies its numbers as fictional. It no longer discusses saved AI tutorials.

The two cards are “Check the confirmed scope” and “A subtotal is not a final quote.” The second says the example’s confirmed subtotal changes from CNY 1,200 to CNY 1,500 while transport, tax basis and amount, and delivery time remain unresolved. Missing values must not become zero or a final commitment.

Those figures come from the source: two devices become three at CNY 300 each, with four installation hours unchanged at CNY 150 per hour. We recalculated both expressions. That checks arithmetic, not a real commercial quote or long-term model reliability.

No video draft was generated because this assignment did not request one. English output is an editorial adaptation from the same task, not an independent experiment.

Inputs and outputs of the two assignments
Diagram of the actual assignments, not a product screenshot.

Check behavior separately from file structure

The skill passed structural validation. We also read the actual output against the source: does it retain fictional status, keep installation hours unchanged, preserve unresolved costs and dates, distinguish subtotal from final price, exclude old-task facts, and respect requested formats?

The output retained those boundaries. The review records evidence and unperformed actions rather than assigning an unsupported success rate. File validation was not presented as proof of automatic discovery.

Using a different input adds evidence beyond inspecting instructions, but only for this task. Another useful attempt would use an article without numbers or request only an introduction.

Repeat the exercise with your own assistant

Unzip the pack into a separate folder accessible to your assistant. Keep reference drafts outside its input. Use request-en.md or request-zh.md to explicitly load the skill and new source. After generation, compare the results with review.md and the reference artifacts.

Different wording is fine. Check that the current source was used, requested formats were delivered, and unfinished actions stayed in draft status. If a problem appears, revise the relevant step and try another input rather than accumulating pages of generic prohibitions.

This task leaves a short skill, its relationship to two different sources, and actual output from the second source. That is a concrete starting point for later discovery and reuse testing.