How to Create a Reusable Astria Skill
To create a reusable skill, describe the kind of result you want, write a first set of instructions, and test it with real inputs. Review the output, then change the instructions that should apply to the next run.
This guide draws on three actual Claude Desktop Code sessions: Match-cut video skill, Song skill creation (which also developed Music Duo), and UGC try-on video skill. They show different parts of that process: defining the format, choosing a generation approach, preparing references and audio, reviewing failures, and deciding which findings belong in the skill.
The examples below come from those sessions on November 4–5, 2026. Review findings are attributed to the chats; they are not claims that every instruction worked or that the skills were exhaustively tested.
Examples
Four outputs from the development sessions. Each had useful results and remaining problems. Use the controls to play sound.
Music Duo
Music Duo: the drum-led test from the Song session. Its review still found synchronization and clothing issues.
Song
Song: the third test film, with explicit lyrics in the video prompt. Framing and the end of the sung line still needed work.
Match Cut
Match Cut: the motion-master version. Its cut review found better stride continuity, with framing and garment differences still present.
UGC Try-on
UGC Try-on: draft 9, with numbered changing shots and approved music. Some complete outfits were still shown too briefly.
1. Define the format before writing the prompt
Describe what the workflow should make and what gives it its structure. Include the inputs you expect, the important constraints and where the workflow should stop to ask you something.
The three opening briefs were quite different:
| Skill | Format requested | What needed to stay consistent |
|---|---|---|
| Match Cut | A short film switching between looks and backgrounds | The same person continues the same action through each cut |
| Song | Take in references, generate a song for the brief, then generate video with that audio reference | The person, voice, lyrics and video performance belong together; no separate muxing step |
| UGC Try-on | A casual try-on video from one camera position, with cuts between garments | The person and room remain consistent while the clothing changes |
Those are reusable directions. The particular person, four outfits, bicycle or song lyrics used in a test are inputs or creative choices for that run.
Keep the first version small. In the UGC brief, the creator asked:
start with only one draft first
And:
see match-cut skill for reference - should be a simple skill. stay concise and do not invent/over-engineer steps.
An existing skill can supply a useful structure. It does not mean every workflow needs that skill’s steps: Song explicitly asked for audio first and video second, without a separate soundtrack-muxing stage.
2. Decide what the skill needs to know
Write down how the agent should get its inputs, which choices it can make, and which choices require a question.
In these sessions, existing Astria prompt IDs supplied outfit references. The skills needed to keep both saved reference tokens and garments described only in the source prompt’s text. A tool-generated restyle prompt in the Match Cut test dropped some of those described garments, so the workflow was changed to preserve them explicitly.
Person references needed their own rule. UGC’s first draft described a person in words without a saved face reference. The creator corrected that:
the video is missing a model face. we always need a model face
Music Duo later gained an intake step for two faces and clothing assigned to each performer. That is a different requirement from choosing one person for a Match Cut film.
A useful intake section answers:
- What references are required, and which input formats can supply them?
- What should happen when a required reference is missing?
- Which details belong to this run, and which are defaults for the format?
- Which decision should the user make before generation?
The last question mattered in Match Cut. Claude first made a single-prompt video, then started a motion-master version after finding stride discontinuities. The creator asked whether the first video was already good enough. The resulting skill asks about the approach upfront, rather than automatically escalating to more generations after its own frame check.
3. Write the instructions, then run a small test
A skill needs enough detail to execute: intake, creative decisions, preparation, generation and review. Write those decisions in ordinary language, using tool names only where the agent needs them.
The real files contain instructions such as these:
Match Cut — what changes in the prompt:
State the shared action once, then only what changes per shot.
Song — how to handle timing:
Timing: state times only for what you measured. Otherwise order the shots by lyric line and let the audio set the pace.
UGC Try-on — how to describe the sequence:
Number the shots, without timestamps. A request for the prompt alone leaves a placeholder for any face or place token that does not exist yet.
These are different rules for different formats. Exact cut times helped define the Match Cut test. Measured lyric timing mattered for Song. Numbered garment actions replaced timestamps in UGC after its earlier shot structure became too crowded.
Test the first version on a small set of real references. Keep the skill version, inputs and output IDs so the result can be traced back to the instructions that produced it. A successful submission only shows that the job was accepted; wait for the output and review it.
4. Check the parts that matter to this format
Use the brief to decide what to inspect. A general “looks good” review misses the behavior the skill is supposed to repeat.
| Format | Useful checks from the sessions | What those checks did not establish |
|---|---|---|
| Match Cut | Compare frames on either side of each cut: body position, stride phase, camera movement and garment fidelity | Frame alignment alone did not settle whether a small jump was distracting at playback speed |
| Song | Measure the audio excerpt, locate sung lines, inspect framing during those lines and check the generated film’s audio | Moving mouths and waveform measurements did not prove perceptually correct lip sync |
| UGC Try-on | Inspect the changing actions, room layout, camera position and visibility of complete outfits | A readable change did not mean the garment got enough time on screen |
| Music Duo | Check who performs each part, visible instrument contact, reference fidelity and timing against the track | An attractive two-person scene did not establish a synchronized performance |
Say what you actually checked. In these chats, Claude repeatedly distinguished sampled-frame or audio measurements from listening. It asked the creator to judge the sound and sync rather than claiming it had heard them.
Keep failures alongside improvements. The Match Cut motion-master version aligned the stride more closely, but inherited a clipped bicycle wheel from its master and changed some garment details. UGC’s numbered shots made the actions clearer, but left too little time to see the first three outfits whole.
5. Change the rule that caused the problem
Feedback is most useful when it identifies an observable problem and the decision behind it. The sessions produced several kinds of revision:
Change the preparation. Song’s first test returned a longer track with an instrumental introduction. The assumed singing rate and a silence check did not locate a usable vocal passage. The workflow was revised to find the actual sung lines and cut an excerpt around them. The later Music Duo tests varied again, so a fixed words-per-second estimate was not a reliable substitute for measuring the recording.
Add an approval step. UGC’s room was initially chosen inside the video prompt. After the creator asked to see it first, the skill generated room options and saved the selected still as a background reference. Song and Music Duo presented the audio excerpt before generating the video.
Rewrite the structure. UGC progressed from garment hold-ups to physically described, numbered changing shots. The creator wanted each garment going on or off to have its own shot. That required replacing the earlier sequence, not just appending another instruction to it.
Clarify the performance. Song gained the lyrics in the video prompt, placed in their corresponding shots, plus more explicit lip-sync wording. The next film measured closer to the audio reference, but its framing also changed. The session correctly stopped short of proving that the wording alone caused the improvement.
Move a choice earlier. Match Cut’s single-render and motion-master paths became an upfront choice. The skill also added a check of the master’s framing, because later shots inherit that framing.
After an edit, distinguish the rule is saved from the rule has been tested. Some final wording changes in these sessions were committed without another render.
6. Test a change without turning one result into a universal rule
When comparing two approaches, keep the person, products and other settings as similar as the question allows. Record what else changed. Otherwise, a better result may not tell you which edit helped.
The Song session reused the audio excerpt across several tests. One retry changed the person and expanded to four looks; another retained those inputs while changing both lip-sync wording and framing. Those runs answered different questions, and none isolated every variable.
The Music Duo follow-up tested a drum-led track after a sparse harmonica version. The measured sound match improved, but synchronization problems remained. Claude proposed requiring a steady beat in every duo track. The creator replied:
Not yet
The rule was left out. One comparison was useful evidence for another test, not enough to restrict every future duet. The session also corrected several unverified comparisons from its earlier report.
Use the same restraint with failures. If a model ignores an existing framing instruction, repeating that instruction in five places may not help. Consider changing the shot, reducing the number of actions, or testing another approach. Keep the skill readable.
7. Save the workflow and test it with new inputs
These sessions used Claude Desktop’s Code mode to edit Markdown files in Astria’s skills repository. Saving meant reviewing and committing those files. Making the updated skill available in Astria involved a separate sync; a Git push alone was not proof that the current version was live. These chats did not demonstrate the personal/workspace skill-saving flow inside Astria chat.
Keep the reusable decisions in the skill: how to take references, choose an approach, prepare assets, generate and review. Keep this run’s person, clothing IDs and lyrics out of the general instructions unless the skill is deliberately specific to them.
Then try another input set. Song did include a retry with a different person and more outfits. Match Cut compared two approaches on the same four looks. UGC changed the face and room, but did not test its final instructions with a second product set. These are different levels of testing; none establishes reliability across every collection or brief.
Read the real files
The excerpts above come from the actual match-cut, song and ugc-try-on source files reviewed alongside the chats. For complete examples:
- UGC Try-on SKILL.md (.zip) — the unedited file at revision
64438ce. - Match Cut SKILL.md (.zip) — revision
8b2d9e2, with two relative-cost phrases omitted. Use the current pricing page for generation pricing.
These files require Astria tools and the helper skills named in their frontmatter. Their instructions describe intended behavior; the development videos show why checking the result remains part of the workflow.