agentcut 0.2.25 → 0.2.26
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +29 -0
- package/dist/config.d.ts +0 -1
- package/dist/config.js +0 -5
- package/dist/config.js.map +1 -1
- package/dist/studio/chat.js.map +1 -1
- package/dist/studio/cli.js +19 -0
- package/dist/studio/cli.js.map +1 -1
- package/dist/studio/host.d.ts +1 -2
- package/dist/studio/host.js +28 -8
- package/dist/studio/host.js.map +1 -1
- package/dist/studio/mcp-guides.js +1 -18
- package/dist/studio/mcp-guides.js.map +1 -1
- package/dist/studio/proc.js +16 -6
- package/dist/studio/proc.js.map +1 -1
- package/dist/studio/proxies.d.ts +4 -0
- package/dist/studio/proxies.js +15 -6
- package/dist/studio/proxies.js.map +1 -1
- package/dist/studio/schema.d.ts +0 -1
- package/dist/studio/schema.js +0 -1
- package/dist/studio/schema.js.map +1 -1
- package/dist/studio/tools.d.ts +0 -4
- package/dist/studio/tools.js +25 -31
- package/dist/studio/tools.js.map +1 -1
- package/dist/studio/voice/kokoro-catalog.d.ts +56 -28
- package/dist/studio/voice/kokoro-catalog.js +56 -28
- package/dist/studio/voice/kokoro-catalog.js.map +1 -1
- package/dist/studio/voice/ming-catalog.d.ts +19 -0
- package/dist/studio/voice/ming-catalog.js +40 -0
- package/dist/studio/voice/ming-catalog.js.map +1 -0
- package/dist/studio/voice/ming.d.ts +12 -0
- package/dist/studio/voice/ming.js +397 -0
- package/dist/studio/voice/ming.js.map +1 -0
- package/dist/studio/voice/previews.d.ts +5 -0
- package/dist/studio/voice/previews.js +50 -0
- package/dist/studio/voice/previews.js.map +1 -0
- package/dist/studio/voice/registry.js +2 -0
- package/dist/studio/voice/registry.js.map +1 -1
- package/dist/studio/voice.d.ts +1 -1
- package/dist/studio/voice.js +5 -6
- package/dist/studio/voice.js.map +1 -1
- package/package.json +1 -1
- package/template/runtime/composition.tsx +5 -1
- package/template/skill/SKILL.md +98 -262
- package/template/skill/references/craft.md +97 -53
- package/template/skill/references/feature-launch.md +33 -29
- package/template/skill/references/how-to.md +146 -161
- package/template/skill/references/launch-direction.md +89 -0
- package/template/skill/references/mcp-external.md +19 -0
- package/template/skill/references/product-launch.md +134 -131
- package/template/skill/references/release-capture.md +53 -0
- package/template/skill/references/sound.md +11 -0
|
@@ -1,30 +1,34 @@
|
|
|
1
|
-
# Feature Launch
|
|
2
|
-
|
|
3
|
-
Use this guide for the Feature Launch chat action or a request to announce one new feature or capability. Also read `product-launch.md` for brand, capture, visual, sound and final review standards. Feature Launch has its own brief and story; it is not a whole-product introduction or a step-by-step lesson.
|
|
4
|
-
|
|
5
|
-
## Identify the feature before work
|
|
6
|
-
|
|
7
|
-
For a bare request, ask which feature is new and what people can now do. A product URL or an old launch alone does not identify the feature. Ask locally before an agent run when possible. Do not scan other projects or shared libraries to guess. Preserve deliberate attachments, current brief, selected style and music. With supplied context, inspect it first and ask only for missing facts that affect the video.
|
|
8
|
-
|
|
9
|
-
Use the assigned canvas video. A populated source gets a separate target, timeline and chat; keep the source unchanged. Reuse a truly empty video. Explicit edits keep their target. This is the standard creative workflow, not the limited release-job API.
|
|
10
|
-
|
|
11
|
-
## Plan the story
|
|
12
|
-
|
|
13
|
-
Write the audience, one new benefit, one verified claim and a shot list in notes.md. Default flow:
|
|
14
|
-
|
|
15
|
-
1. **New benefit.** Open with what people can now do. Name the feature when it helps recognition.
|
|
16
|
-
2. **Feature in action.** Show a real action and its visible result, with enough context to understand what changed.
|
|
17
|
-
3. **Outcome.** Make the benefit visible. Use a supported before/after only when the evidence exists. Do not invent time savings, performance figures or customer claims.
|
|
18
|
-
4. **Next step.** End with one useful action: try the feature, open its location, or read the announcement. State verified availability or plan restrictions when relevant.
|
|
19
|
-
|
|
20
|
-
The problem statement is optional. Existing users often know the problem; do not delay the feature with a repeated product introduction. For a new audience, add only the context required to understand the benefit. Keep one main feature; do not turn the video into a feature list. Choose duration from the material and channel rather than a fixed formula. Respect requested length, aspect ratios, voice, captions and music.
|
|
21
|
-
|
|
22
|
-
For example, a canvas-versions announcement could show making two independent versions, then switching between their timelines and chats. The visible result supports “explore another version while keeping the original.” A How-to would instead explain each step needed to complete that task.
|
|
23
|
-
|
|
24
|
-
## Build and verify
|
|
25
|
-
|
|
26
|
-
|
|
27
|
-
|
|
28
|
-
|
|
29
|
-
|
|
1
|
+
# Feature Launch
|
|
2
|
+
|
|
3
|
+
Use this guide for the Feature Launch chat action or a request to announce one new feature or capability. Also read `product-launch.md` for brand, capture, visual, sound and final review standards. Feature Launch has its own brief and story; it is not a whole-product introduction or a step-by-step lesson.
|
|
4
|
+
|
|
5
|
+
## Identify the feature before work
|
|
6
|
+
|
|
7
|
+
For a bare request, ask which feature is new and what people can now do. A product URL or an old launch alone does not identify the feature. Ask locally before an agent run when possible. Do not scan other projects or shared libraries to guess. Preserve deliberate attachments, current brief, selected style and music. With supplied context, inspect it first and ask only for missing facts that affect the video.
|
|
8
|
+
|
|
9
|
+
Use the assigned canvas video. A populated source gets a separate target, timeline and chat; keep the source unchanged. Reuse a truly empty video. Explicit edits keep their target. This is the standard creative workflow, not the limited release-job API.
|
|
10
|
+
|
|
11
|
+
## Plan the story
|
|
12
|
+
|
|
13
|
+
Write the audience, one new benefit, one verified claim and a shot list in notes.md. Default flow:
|
|
14
|
+
|
|
15
|
+
1. **New benefit.** Open with what people can now do. Name the feature when it helps recognition.
|
|
16
|
+
2. **Feature in action.** Show a real action and its visible result, with enough context to understand what changed.
|
|
17
|
+
3. **Outcome.** Make the benefit visible. Use a supported before/after only when the evidence exists. Do not invent time savings, performance figures or customer claims.
|
|
18
|
+
4. **Next step.** End with one useful action: try the feature, open its location, or read the announcement. State verified availability or plan restrictions when relevant.
|
|
19
|
+
|
|
20
|
+
The problem statement is optional. Existing users often know the problem; do not delay the feature with a repeated product introduction. For a new audience, add only the context required to understand the benefit. Keep one main feature; do not turn the video into a feature list. Choose duration from the material and channel rather than a fixed formula. Respect requested length, aspect ratios, voice, captions and music.
|
|
21
|
+
|
|
22
|
+
For example, a canvas-versions announcement could show making two independent versions, then switching between their timelines and chats. The visible result supports “explore another version while keeping the original.” A How-to would instead explain each step needed to complete that task.
|
|
23
|
+
|
|
24
|
+
## Build and verify
|
|
25
|
+
|
|
26
|
+
Read `references/launch-direction.md` before building a new film or a major revision. Use its
|
|
27
|
+
concrete visual settings, shot table, early proof captures and transition checks with this feature's
|
|
28
|
+
brief. Keep an approved script or existing direction when the request is a small edit.
|
|
29
|
+
|
|
30
|
+
Follow the full Agentcut skill: inspect real assets and the component/3D libraries, plan intentional motion and camera changes, use editable tracks, coordinate music and sound effects, and retain final full-picture and audio review. Do not replace the process with a fixed animation recipe. Use the user's own footage or safe real-app capture. If proof or access is missing, ask for it before building a false demonstration.
|
|
31
|
+
|
|
32
|
+
Use `sound.md` for voice and music. Generic voice-over uses local Kokoro; a paid voice requires explicit user selection. A silent announcement is valid when requested. If voice is used, inspect the take, check its words and use transcription timings for accurate captions. Music must support the action and leave the message clear.
|
|
33
|
+
|
|
30
34
|
Before delivery, verify that viewers can identify the feature, understand its new benefit, see evidence in the actual demonstration, and know the next step. Check small-screen readability, output formats, timing, asset paths, picture and sound. Deliver the editable result only after the normal final review passes.
|
|
@@ -1,161 +1,146 @@
|
|
|
1
|
-
# How-to
|
|
2
|
-
|
|
3
|
-
Help a viewer complete one real task and recognize success. Use this guide for the How-to action,
|
|
4
|
-
tutorial requests and software walkthrough videos. The user's brief, uploaded material, existing edit,
|
|
5
|
-
selected style and music take priority. This is not a fixed scene sequence or a promotional film.
|
|
6
|
-
|
|
7
|
-
## Establish the task before searching
|
|
8
|
-
|
|
9
|
-
The studio asks locally when the request is empty. If a bare request reaches the agent without a
|
|
10
|
-
task, brief, recording or relevant project material, ask immediately: "What should viewers learn to
|
|
11
|
-
do? Add a URL, recording, or a short description of the task." Stop before tool calls, source or
|
|
12
|
-
library searches, planning or rendering. A project name, generic earlier request, music choice or
|
|
13
|
-
unrelated shared stock does not establish a task. A known product alone may still lack a procedure.
|
|
14
|
-
|
|
15
|
-
This also applies when How-to is clicked while viewing a finished launch or another populated video.
|
|
16
|
-
The old edit, its incidental assets and prior launch conversation do not identify the new tutorial's
|
|
17
|
-
task. Ask locally before creating a destination or sending to the agent. An explicit current task,
|
|
18
|
-
genuinely established tutorial task, or material deliberately attached for the tutorial can proceed.
|
|
19
|
-
Keep project uploads available for inspection after the user answers; do not discard them.
|
|
20
|
-
|
|
21
|
-
With material, inspect the current project and attachments first. Read `references/format.md`,
|
|
22
|
-
`references/footage.md` for recordings, and `references/website-howto.md` for website execution and
|
|
23
|
-
tool recovery. Use media information, source frames and transcripts to find complete actions and
|
|
24
|
-
their results. Do not ask the user to repeat facts already in the brief or a visible recording.
|
|
25
|
-
If one crucial task detail remains missing, ask that one question and continue work that is clear.
|
|
26
|
-
When a user reply is required, state the exact missing input plainly, for example
|
|
27
|
-
"Question for you: Which voice should this use?" or "Still blocked: I need a recording of the task."
|
|
28
|
-
Do not bury the needed action beneath a long completion update or imply that generation continues.
|
|
29
|
-
|
|
30
|
-
## Obtain the real UI from a task-only brief
|
|
31
|
-
|
|
32
|
-
A request such as "how to add a style to Agentcut" supplies a task. Do not make the user also ask
|
|
33
|
-
for screenshots, captions or an edited video. Proactively obtain authentic evidence:
|
|
34
|
-
|
|
35
|
-
1. Check relevant captures or recordings already in the target project's assets and its supplied
|
|
36
|
-
source material. Reuse them only if they show the current labels, complete actions and success state.
|
|
37
|
-
2. Use the supplied or already-known product URL or an already-running local preview. Read the
|
|
38
|
-
relevant routes and control labels from configured source folders if that helps locate the task.
|
|
39
|
-
A source-code read locates controls; it is not a visual or proof that an action works. Do not
|
|
40
|
-
search the home folder for identity or start/restart the user's services as part of discovery.
|
|
41
|
-
3. Use an available permitted browser capture tool. Prefer a connected browser automation tool
|
|
42
|
-
when it supports actual screenshots/recording. Otherwise the helper includes `agent-browser`:
|
|
43
|
-
read `agent-browser skills get core` and its video-recording reference for the installed version,
|
|
44
|
-
use a named session, open the known URL, inspect the live controls, and take real screenshots or
|
|
45
|
-
a recording into the new video's `assets/`. Inspect before each action and re-inspect afterward.
|
|
46
|
-
The installed CLI supports `screenshot <path>` and `record start <path.webm> --cursor`, followed
|
|
47
|
-
by planned actions and `record stop`. Use only supported commands from its local documentation.
|
|
48
|
-
4. Demonstrate the requested task in prepared safe demo state, retaining the real pointer/action,
|
|
49
|
-
response and final success state. Do not modify the user's original video to make tutorial footage.
|
|
50
|
-
For Agentcut, use a disposable safe demo video rather than changing the completed launch. Reuse
|
|
51
|
-
existing captures if obtaining that safe state is not possible under the current permissions.
|
|
52
|
-
5. If sign-in, access, an unknown URL, blocked network permissions or capture capability prevents
|
|
53
|
-
this, ask for the smallest missing link or recording. The user signs in directly; never request
|
|
54
|
-
credentials, cookies or tokens. Do not fabricate screens, bypass a block, or claim code proves UI.
|
|
55
|
-
|
|
56
|
-
`agentcut_capture_frames` captures the edited composition, not an arbitrary website. The release-job
|
|
57
|
-
API still supports announcement only; do not submit a how-to job to it or pretend it captured a task.
|
|
58
|
-
After authentic material exists, add accurate captions or step labels without waiting for a separate
|
|
59
|
-
request, then build and review the editable tutorial. Preserve all earlier edits and source files.
|
|
60
|
-
|
|
61
|
-
## Plan for someone doing the task
|
|
62
|
-
|
|
63
|
-
Record in `notes.md`: the intended viewer, their starting state and prerequisites, the task they
|
|
64
|
-
will complete, and the visible evidence that confirms success. Show the direction in a few lines.
|
|
65
|
-
For example, "Create a project from a preset" needs the real starting screen, the preset selection,
|
|
66
|
-
the creation action and the newly opened project. A title that says "Easy presets" is not instruction.
|
|
67
|
-
|
|
68
|
-
Use a task title with a clear action. Start with the useful result or a brief orientation, then give
|
|
69
|
-
the minimum context needed to begin. Teach the actual procedure in its dependency order. Put any
|
|
70
|
-
necessary account, permission or prepared-file condition before the first action that needs it.
|
|
71
|
-
Do not turn the tutorial into a tour of every feature. Split distinct goals when the user requests
|
|
72
|
-
several tutorials; do not silently omit a required step to hit an arbitrary duration.
|
|
73
|
-
|
|
74
|
-
Build a flexible cue sheet. For each step note the source range or real captured view, exact control
|
|
75
|
-
label, action, visible reaction, focus region, text or narration cue and readable hold. Mark the
|
|
76
|
-
final state that proves success. Choose length from the requested duration and the time needed to
|
|
77
|
-
understand the steps. No duration slider or camera vocabulary is required from the user.
|
|
78
|
-
|
|
79
|
-
## Real actions and useful cues
|
|
80
|
-
|
|
81
|
-
- Use real screenshots and recordings from the current app or user material as the procedural
|
|
82
|
-
evidence. A screenshot can orient the viewer or show a result; it must not imply that an unseen
|
|
83
|
-
click was verified. Capture a missing transition or state clearly what remains unverified.
|
|
84
|
-
- Follow the main skill's website safety rules. Inspect before capture. Use safe sample content;
|
|
85
|
-
do not submit live messages, change settings or expose credentials to make a tutorial. If a write
|
|
86
|
-
action needs user preparation or authorization, explain the missing step. Do not invent it.
|
|
87
|
-
- Preserve actual labels, cursor targets, menu state, entry values and system feedback. Do not
|
|
88
|
-
replace procedural evidence with an imagined UI or pretend a decorative animation is a capture.
|
|
89
|
-
Any explanatory reconstruction must be clearly identified and must not be the only proof.
|
|
90
|
-
- Show where to act, the action, then its result. Give the viewer a stable pause at meaningful step
|
|
91
|
-
boundaries instead of relying on them to discover where to pause. Trim idle waits but retain
|
|
92
|
-
feedback that explains progress and completion. Speed up typing only when the entered value and
|
|
93
|
-
result remain clear. Keep important errors and their recovery when relevant to the task.
|
|
94
|
-
- Use a brief highlight, pointer or close-up on the relevant control as the instruction reaches it.
|
|
95
|
-
Do not cue several competing objects. Start wide enough for orientation, move to the detail,
|
|
96
|
-
then hold sharp and readable through the action and reaction. Follow cursor metadata when present
|
|
97
|
-
(`references/cursor.md`). Keep the current step and progress clear without excessive labels.
|
|
98
|
-
- Narration should explain the action as it appears. Use plain action words and exact UI labels.
|
|
99
|
-
Do not narrate a click several seconds before or after the visible action. For a silent tutorial,
|
|
100
|
-
place concise step instructions next to the demonstrated action and hold them long enough to read.
|
|
101
|
-
|
|
102
|
-
## Brand, library, motion and sound
|
|
103
|
-
|
|
104
|
-
Match the app's actual logos, fonts, colors and spacing. Discover useful existing project images,
|
|
105
|
-
B-roll, recordings, music and effects even when they are not chat attachments: list project `assets/`
|
|
106
|
-
with `agentcut_list_files`, then relevant shared library folders with `scope: "library"`. Follow
|
|
107
|
-
pagination and inspect relevant subfolders. Use `assets/...` and `library:...` media paths. Do not scan
|
|
108
|
-
other projects or every source directory to guess the task. Reuse known allowed source locations for
|
|
109
|
-
brand checks; if one is missing, ask instead of searching the home folder. Tell the user they can
|
|
110
|
-
drop a specific missing asset into the Library panel; a second chat attachment is not required.
|
|
111
|
-
|
|
112
|
-
Use B-roll only when it explains the task without hiding an essential step. Reuse existing components
|
|
113
|
-
and relevant assets; do not generate replacements by default. With website source available, inspect
|
|
114
|
-
the exact relevant components rather than reading the whole repository repeatedly.
|
|
115
|
-
|
|
116
|
-
Read `references/cinematic.md` when using camera depth or devices. Use existing `Device3D` hardware
|
|
117
|
-
for visible phones and computers. Smooth motion, backgrounds and restrained brand effects can frame
|
|
118
|
-
the tutorial, but essential UI must stay sharp, upright enough to read and clear of captions. Blur,
|
|
119
|
-
perspective and zoom should lead into a readable hold. Do not carry launch-film energy through every
|
|
120
|
-
instructional step. Keep optional backgrounds quiet during the action.
|
|
121
|
-
|
|
122
|
-
Read `references/sound.md`. Preserve a selected song; balance it below speech and keep busy music or
|
|
123
|
-
effects away from difficult steps. Use a small sound on a meaningful action, not every pointer move.
|
|
124
|
-
Caption narration accurately using the transcription workflow. Correct names, numbers and UI labels;
|
|
125
|
-
include meaningful non-speech audio when needed to understand the task. Captions must follow the
|
|
126
|
-
audio and avoid controls, results and the cursor target. Do not require voice for a silent brief.
|
|
127
|
-
|
|
128
|
-
## Success and review
|
|
129
|
-
|
|
130
|
-
Read `references/efficient-review.md`. Group independent file/guide reads and target changed shots
|
|
131
|
-
during revisions. Broaden after shared or timing changes; retain the complete final review below.
|
|
132
|
-
|
|
133
|
-
End at the real successful state. Tell the viewer what to check, for example the saved item appearing
|
|
134
|
-
in its list or an exported file opening correctly. An optional short "Try it with your own project"
|
|
135
|
-
prompt can encourage practice. Do not claim that an extra replay or recap guarantees learning.
|
|
136
|
-
|
|
137
|
-
Use the normal build and review loop: editable scene, text, footage and audio clips, status after
|
|
138
|
-
changes, source-frame checks, wide and mobile contact sheets, audio measurements and heavy-scene
|
|
139
|
-
playback checks. Check that a novice could follow every action without guessing a hidden step.
|
|
140
|
-
Check the order, cue timing, pointer visibility, label accuracy, reading time, final state and caption
|
|
141
|
-
placement. Retain both-format and sound review. Export only when requested.
|
|
142
|
-
|
|
143
|
-
Reuse current material and passed checks. Check changed regions first, then review the complete edit
|
|
144
|
-
in both formats. Run page-heavy checks in sequence as `website-howto.md` requires. If a tool stalls,
|
|
145
|
-
inspect active work before retrying. Record what passed and what remains blocked; do not claim a
|
|
146
|
-
build alone proves instructional accuracy.
|
|
147
|
-
|
|
148
|
-
## Evidence used for this guide
|
|
149
|
-
|
|
150
|
-
The application above is original Agentcut guidance, informed by these primary sources. These studies
|
|
151
|
-
do not establish a universal video length or guarantee an outcome for Agentcut users.
|
|
152
|
-
|
|
153
|
-
- [Software-training research (2013)](https://ris.utwente.nl/ws/portalfiles/portal/6827789/s4.pdf)
|
|
154
|
-
supports task-focused demonstrations, accurate interfaces and coordinated action and narration.
|
|
155
|
-
- [A controlled study of pause and navigation guidance (2022)](https://pmc.ncbi.nlm.nih.gov/articles/PMC9610327/)
|
|
156
|
-
supports explicit help for when to pause and review. Its experiment used educational material,
|
|
157
|
-
so applying it to software-step holds is a design inference.
|
|
158
|
-
- [A software-tutorial study of reviews (2018)](https://research.utwente.nl/en/publications/supporting-motivation-task-performance-and-retention-in-video-tut/)
|
|
159
|
-
found no extra benefit from its review or second-demonstration conditions. Keep recaps purposeful.
|
|
160
|
-
- [W3C caption guidance](https://www.w3.org/WAI/media/av/captions/) calls for synchronized, accurate
|
|
161
|
-
captions and correction of automatic output. This informs the caption review above.
|
|
1
|
+
# How-to
|
|
2
|
+
|
|
3
|
+
Help a viewer complete one real task and recognize success. Use this guide for the How-to action,
|
|
4
|
+
tutorial requests and software walkthrough videos. The user's brief, uploaded material, existing edit,
|
|
5
|
+
selected style and music take priority. This is not a fixed scene sequence or a promotional film.
|
|
6
|
+
|
|
7
|
+
## Establish the task before searching
|
|
8
|
+
|
|
9
|
+
The studio asks locally when the request is empty. If a bare request reaches the agent without a
|
|
10
|
+
task, brief, recording or relevant project material, ask immediately: "What should viewers learn to
|
|
11
|
+
do? Add a URL, recording, or a short description of the task." Stop before tool calls, source or
|
|
12
|
+
library searches, planning or rendering. A project name, generic earlier request, music choice or
|
|
13
|
+
unrelated shared stock does not establish a task. A known product alone may still lack a procedure.
|
|
14
|
+
|
|
15
|
+
This also applies when How-to is clicked while viewing a finished launch or another populated video.
|
|
16
|
+
The old edit, its incidental assets and prior launch conversation do not identify the new tutorial's
|
|
17
|
+
task. Ask locally before creating a destination or sending to the agent. An explicit current task,
|
|
18
|
+
genuinely established tutorial task, or material deliberately attached for the tutorial can proceed.
|
|
19
|
+
Keep project uploads available for inspection after the user answers; do not discard them.
|
|
20
|
+
|
|
21
|
+
With material, inspect the current project and attachments first. Read `references/format.md`,
|
|
22
|
+
`references/footage.md` for recordings, and `references/website-howto.md` for website execution and
|
|
23
|
+
tool recovery. Use media information, source frames and transcripts to find complete actions and
|
|
24
|
+
their results. Do not ask the user to repeat facts already in the brief or a visible recording.
|
|
25
|
+
If one crucial task detail remains missing, ask that one question and continue work that is clear.
|
|
26
|
+
When a user reply is required, state the exact missing input plainly, for example
|
|
27
|
+
"Question for you: Which voice should this use?" or "Still blocked: I need a recording of the task."
|
|
28
|
+
Do not bury the needed action beneath a long completion update or imply that generation continues.
|
|
29
|
+
|
|
30
|
+
## Obtain the real UI from a task-only brief
|
|
31
|
+
|
|
32
|
+
A request such as "how to add a style to Agentcut" supplies a task. Do not make the user also ask
|
|
33
|
+
for screenshots, captions or an edited video. Proactively obtain authentic evidence:
|
|
34
|
+
|
|
35
|
+
1. Check relevant captures or recordings already in the target project's assets and its supplied
|
|
36
|
+
source material. Reuse them only if they show the current labels, complete actions and success state.
|
|
37
|
+
2. Use the supplied or already-known product URL or an already-running local preview. Read the
|
|
38
|
+
relevant routes and control labels from configured source folders if that helps locate the task.
|
|
39
|
+
A source-code read locates controls; it is not a visual or proof that an action works. Do not
|
|
40
|
+
search the home folder for identity or start/restart the user's services as part of discovery.
|
|
41
|
+
3. Use an available permitted browser capture tool. Prefer a connected browser automation tool
|
|
42
|
+
when it supports actual screenshots/recording. Otherwise the helper includes `agent-browser`:
|
|
43
|
+
read `agent-browser skills get core` and its video-recording reference for the installed version,
|
|
44
|
+
use a named session, open the known URL, inspect the live controls, and take real screenshots or
|
|
45
|
+
a recording into the new video's `assets/`. Inspect before each action and re-inspect afterward.
|
|
46
|
+
The installed CLI supports `screenshot <path>` and `record start <path.webm> --cursor`, followed
|
|
47
|
+
by planned actions and `record stop`. Use only supported commands from its local documentation.
|
|
48
|
+
4. Demonstrate the requested task in prepared safe demo state, retaining the real pointer/action,
|
|
49
|
+
response and final success state. Do not modify the user's original video to make tutorial footage.
|
|
50
|
+
For Agentcut, use a disposable safe demo video rather than changing the completed launch. Reuse
|
|
51
|
+
existing captures if obtaining that safe state is not possible under the current permissions.
|
|
52
|
+
5. If sign-in, access, an unknown URL, blocked network permissions or capture capability prevents
|
|
53
|
+
this, ask for the smallest missing link or recording. The user signs in directly; never request
|
|
54
|
+
credentials, cookies or tokens. Do not fabricate screens, bypass a block, or claim code proves UI.
|
|
55
|
+
|
|
56
|
+
`agentcut_capture_frames` captures the edited composition, not an arbitrary website. The release-job
|
|
57
|
+
API still supports announcement only; do not submit a how-to job to it or pretend it captured a task.
|
|
58
|
+
After authentic material exists, add accurate captions or step labels without waiting for a separate
|
|
59
|
+
request, then build and review the editable tutorial. Preserve all earlier edits and source files.
|
|
60
|
+
|
|
61
|
+
## Plan for someone doing the task
|
|
62
|
+
|
|
63
|
+
Record in `notes.md`: the intended viewer, their starting state and prerequisites, the task they
|
|
64
|
+
will complete, and the visible evidence that confirms success. Show the direction in a few lines.
|
|
65
|
+
For example, "Create a project from a preset" needs the real starting screen, the preset selection,
|
|
66
|
+
the creation action and the newly opened project. A title that says "Easy presets" is not instruction.
|
|
67
|
+
|
|
68
|
+
Use a task title with a clear action. Start with the useful result or a brief orientation, then give
|
|
69
|
+
the minimum context needed to begin. Teach the actual procedure in its dependency order. Put any
|
|
70
|
+
necessary account, permission or prepared-file condition before the first action that needs it.
|
|
71
|
+
Do not turn the tutorial into a tour of every feature. Split distinct goals when the user requests
|
|
72
|
+
several tutorials; do not silently omit a required step to hit an arbitrary duration.
|
|
73
|
+
|
|
74
|
+
Build a flexible cue sheet. For each step note the source range or real captured view, exact control
|
|
75
|
+
label, action, visible reaction, focus region, text or narration cue and readable hold. Mark the
|
|
76
|
+
final state that proves success. Choose length from the requested duration and the time needed to
|
|
77
|
+
understand the steps. No duration slider or camera vocabulary is required from the user.
|
|
78
|
+
|
|
79
|
+
## Real actions and useful cues
|
|
80
|
+
|
|
81
|
+
- Use real screenshots and recordings from the current app or user material as the procedural
|
|
82
|
+
evidence. A screenshot can orient the viewer or show a result; it must not imply that an unseen
|
|
83
|
+
click was verified. Capture a missing transition or state clearly what remains unverified.
|
|
84
|
+
- Follow the main skill's website safety rules. Inspect before capture. Use safe sample content;
|
|
85
|
+
do not submit live messages, change settings or expose credentials to make a tutorial. If a write
|
|
86
|
+
action needs user preparation or authorization, explain the missing step. Do not invent it.
|
|
87
|
+
- Preserve actual labels, cursor targets, menu state, entry values and system feedback. Do not
|
|
88
|
+
replace procedural evidence with an imagined UI or pretend a decorative animation is a capture.
|
|
89
|
+
Any explanatory reconstruction must be clearly identified and must not be the only proof.
|
|
90
|
+
- Show where to act, the action, then its result. Give the viewer a stable pause at meaningful step
|
|
91
|
+
boundaries instead of relying on them to discover where to pause. Trim idle waits but retain
|
|
92
|
+
feedback that explains progress and completion. Speed up typing only when the entered value and
|
|
93
|
+
result remain clear. Keep important errors and their recovery when relevant to the task.
|
|
94
|
+
- Use a brief highlight, pointer or close-up on the relevant control as the instruction reaches it.
|
|
95
|
+
Do not cue several competing objects. Start wide enough for orientation, move to the detail,
|
|
96
|
+
then hold sharp and readable through the action and reaction. Follow cursor metadata when present
|
|
97
|
+
(`references/cursor.md`). Keep the current step and progress clear without excessive labels.
|
|
98
|
+
- Narration should explain the action as it appears. Use plain action words and exact UI labels.
|
|
99
|
+
Do not narrate a click several seconds before or after the visible action. For a silent tutorial,
|
|
100
|
+
place concise step instructions next to the demonstrated action and hold them long enough to read.
|
|
101
|
+
|
|
102
|
+
## Brand, library, motion and sound
|
|
103
|
+
|
|
104
|
+
Match the app's actual logos, fonts, colors and spacing. Discover useful existing project images,
|
|
105
|
+
B-roll, recordings, music and effects even when they are not chat attachments: list project `assets/`
|
|
106
|
+
with `agentcut_list_files`, then relevant shared library folders with `scope: "library"`. Follow
|
|
107
|
+
pagination and inspect relevant subfolders. Use `assets/...` and `library:...` media paths. Do not scan
|
|
108
|
+
other projects or every source directory to guess the task. Reuse known allowed source locations for
|
|
109
|
+
brand checks; if one is missing, ask instead of searching the home folder. Tell the user they can
|
|
110
|
+
drop a specific missing asset into the Library panel; a second chat attachment is not required.
|
|
111
|
+
|
|
112
|
+
Use B-roll only when it explains the task without hiding an essential step. Reuse existing components
|
|
113
|
+
and relevant assets; do not generate replacements by default. With website source available, inspect
|
|
114
|
+
the exact relevant components rather than reading the whole repository repeatedly.
|
|
115
|
+
|
|
116
|
+
Read `references/cinematic.md` when using camera depth or devices. Use existing `Device3D` hardware
|
|
117
|
+
for visible phones and computers. Smooth motion, backgrounds and restrained brand effects can frame
|
|
118
|
+
the tutorial, but essential UI must stay sharp, upright enough to read and clear of captions. Blur,
|
|
119
|
+
perspective and zoom should lead into a readable hold. Do not carry launch-film energy through every
|
|
120
|
+
instructional step. Keep optional backgrounds quiet during the action.
|
|
121
|
+
|
|
122
|
+
Read `references/sound.md`. Preserve a selected song; balance it below speech and keep busy music or
|
|
123
|
+
effects away from difficult steps. Use a small sound on a meaningful action, not every pointer move.
|
|
124
|
+
Caption narration accurately using the transcription workflow. Correct names, numbers and UI labels;
|
|
125
|
+
include meaningful non-speech audio when needed to understand the task. Captions must follow the
|
|
126
|
+
audio and avoid controls, results and the cursor target. Do not require voice for a silent brief.
|
|
127
|
+
|
|
128
|
+
## Success and review
|
|
129
|
+
|
|
130
|
+
Read `references/efficient-review.md`. Group independent file/guide reads and target changed shots
|
|
131
|
+
during revisions. Broaden after shared or timing changes; retain the complete final review below.
|
|
132
|
+
|
|
133
|
+
End at the real successful state. Tell the viewer what to check, for example the saved item appearing
|
|
134
|
+
in its list or an exported file opening correctly. An optional short "Try it with your own project"
|
|
135
|
+
prompt can encourage practice. Do not claim that an extra replay or recap guarantees learning.
|
|
136
|
+
|
|
137
|
+
Use the normal build and review loop: editable scene, text, footage and audio clips, status after
|
|
138
|
+
changes, source-frame checks, wide and mobile contact sheets, audio measurements and heavy-scene
|
|
139
|
+
playback checks. Check that a novice could follow every action without guessing a hidden step.
|
|
140
|
+
Check the order, cue timing, pointer visibility, label accuracy, reading time, final state and caption
|
|
141
|
+
placement. Retain both-format and sound review. Export only when requested.
|
|
142
|
+
|
|
143
|
+
Reuse current material and passed checks. Check changed regions first, then review the complete edit
|
|
144
|
+
in both formats. Run page-heavy checks in sequence as `website-howto.md` requires. If a tool stalls,
|
|
145
|
+
inspect active work before retrying. Record what passed and what remains blocked; do not claim a
|
|
146
|
+
build alone proves instructional accuracy.
|
|
@@ -0,0 +1,89 @@
|
|
|
1
|
+
# Launch direction: from brief to checked shots
|
|
2
|
+
|
|
3
|
+
Use this with `product-launch.md` or `feature-launch.md` for a new film or a major revision.
|
|
4
|
+
For a small edit, keep the existing direction and check the changed shots. The user's script,
|
|
5
|
+
reference, selected style and format take priority. Keep decisions in the project's `notes.md`.
|
|
6
|
+
This guide uses Agentcut tools and frame timing; it does not require another video engine.
|
|
7
|
+
|
|
8
|
+
## Choose a story that the images can tell
|
|
9
|
+
|
|
10
|
+
Compare two possible stories using the available evidence, not just two opening titles. For example,
|
|
11
|
+
show a finished result and then how it was made, or follow one real input through an action to its
|
|
12
|
+
result. Pick the clearest story and record why. Continue without an approval stop unless the user
|
|
13
|
+
asked to choose a direction. Do not repeat this exercise for an approved script or storyboard.
|
|
14
|
+
|
|
15
|
+
Keep the same recognizable file, query, project or other subject through the demonstration when
|
|
16
|
+
the source supports it. A viewer should see what changed. Unrelated screens with benefit labels
|
|
17
|
+
do not explain a workflow. Do not fabricate a failure in the old workflow to create tension.
|
|
18
|
+
|
|
19
|
+
## Turn the direction into build decisions
|
|
20
|
+
|
|
21
|
+
Before building scenes, add a compact direction and shot table to `notes.md`:
|
|
22
|
+
|
|
23
|
+
- **Look:** actual background and text colors, accent use, font files/families, weights, text sizes,
|
|
24
|
+
margins and product framing for each output format. Use selected style tokens where available.
|
|
25
|
+
Replace vague words such as "premium" with these choices. Keep meaningful product status colors.
|
|
26
|
+
- **Motion:** entrance, camera and exit durations in frames, easing or spring settings, and the
|
|
27
|
+
settled reading window. Choose motion by purpose; do not give every element the same entrance.
|
|
28
|
+
- **Avoid for this film:** identify likely distractions, such as repeated feature cards, tiny UI
|
|
29
|
+
inside a large empty stage, or decoration that competes with proof. This is specific to the brief;
|
|
30
|
+
do not ban a color, material or effect that the user's style needs.
|
|
31
|
+
|
|
32
|
+
| Frames [start, end) | What the viewer learns | Source and visible change | Framing and main move | Exact words and settled read frames | Sound cue | Link to next shot |
|
|
33
|
+
| --- | --- | --- | --- | --- | --- | --- |
|
|
34
|
+
|
|
35
|
+
Fill one row per shot. Use real asset paths and source ranges, not placeholder evidence. Keep the
|
|
36
|
+
table aligned with the timeline after timing changes. Plan the final readable frame before `end`.
|
|
37
|
+
Do not add claims, voice or captions just to fill the table.
|
|
38
|
+
|
|
39
|
+
For each demonstration, establish where the task happens, bring the relevant action into clear
|
|
40
|
+
view, then show its result in context. These can be parts of one shot. A pan across a still image
|
|
41
|
+
can explain a layout; it cannot prove an interaction. Keep real footage for the action.
|
|
42
|
+
|
|
43
|
+
## Prove the visual direction early
|
|
44
|
+
|
|
45
|
+
Build the opening and the main proof shot first, in story order when possible. Capture their
|
|
46
|
+
settled frames in each output format with `agentcut_capture_frames`. Inspect actual words, assets,
|
|
47
|
+
framing and contrast before extending that treatment to the rest of the film. A full storyboard
|
|
48
|
+
approval step is optional, only when requested; these are working checks in the live project.
|
|
49
|
+
|
|
50
|
+
The first frame should give the viewer useful visual context unless the brief calls for darkness.
|
|
51
|
+
Give the product enough space for the relevant action to read. At small display size, crop closer
|
|
52
|
+
or reduce the words before reducing text size. A full app view establishes context; its incidental
|
|
53
|
+
labels do not all need to be readable. Essential labels and results do.
|
|
54
|
+
|
|
55
|
+
## Make motion carry the story
|
|
56
|
+
|
|
57
|
+
- Give each shot one main movement. Delay or reduce secondary motion so it does not compete.
|
|
58
|
+
Let essential titles and UI settle while the viewer reads. Idle floating is not a demonstration.
|
|
59
|
+
- Contrast short transitions with longer action and result holds. As a starting estimate, allow
|
|
60
|
+
about `0.5 + 0.3 * wordCount` seconds for a short line after it settles. Validate with the actual
|
|
61
|
+
layout and audience; dense UI needs more time. Cut words or decoration to meet a fixed duration.
|
|
62
|
+
- Connect shots through a visible action, a continuing subject, matching framing or a deliberate
|
|
63
|
+
cut. Name the connection in the shot table. Do not require a transition quota or a new effect at
|
|
64
|
+
every cut. A clear hard cut is useful when a morph would distort the evidence.
|
|
65
|
+
- Keep important labels sharp during the reading window. Motion blur belongs to movement, not
|
|
66
|
+
the proof hold. Use the existing component catalog and `cinematic.md` tools where suitable.
|
|
67
|
+
Follow `browser-export.md`; a browser-only effect is not proof that an export works.
|
|
68
|
+
- Use source action times, narration timings or analyzed music cues. For a regular beat grid,
|
|
69
|
+
calculate each frame from its beat index and round once; repeated rounded intervals cause drift.
|
|
70
|
+
|
|
71
|
+
## Review the cuts and the meaning
|
|
72
|
+
|
|
73
|
+
Keep the main skill's full picture, sound and performance review. Add these checks to it:
|
|
74
|
+
|
|
75
|
+
1. Inspect frame 0, the settled proof and the final readable frame in each format. Read the shot
|
|
76
|
+
sequence as a new viewer. Without the headline, is the promised change still visible?
|
|
77
|
+
2. For each cut at frame `c`, sample `c - 1` and `c`, plus nearby frames across the transition.
|
|
78
|
+
Clamp samples to the video bounds. For a longer move, sample intermediate frames too. Look for
|
|
79
|
+
blank gaps, duplicated subjects, jumps in labels, obscured results and conflicting motion.
|
|
80
|
+
3. Inspect essential words and UI at a small display size, about 360 px wide where available.
|
|
81
|
+
If a tool cannot show that size, record this limit; do not claim a small-screen check passed.
|
|
82
|
+
4. Record concrete issues as `frame / variant / issue / change / recheck result` in `notes.md`.
|
|
83
|
+
Fix wrong claims, unreadable evidence and broken transitions before decorative details. Recheck
|
|
84
|
+
changed shots using `efficient-review.md`; timing or shared changes require broader checks.
|
|
85
|
+
|
|
86
|
+
At hand-off, state what was checked and any remaining weak shot or missing evidence. A build pass
|
|
87
|
+
only proves the project compiles. Do not describe a film as visually verified without viewing its
|
|
88
|
+
captures, or say sound was heard when it was only measured.
|
|
89
|
+
|
|
@@ -0,0 +1,19 @@
|
|
|
1
|
+
# Working through MCP from outside the workspace
|
|
2
|
+
|
|
3
|
+
Use the same creative and review workflow whether you work inside Agentcut or through MCP.
|
|
4
|
+
Start with `agentcut_capabilities`, `SKILL.md` through `agentcut_guide`, then `agentcut_status`.
|
|
5
|
+
Supporting `references/*.md` files are available through `agentcut_guide` and MCP guide resources.
|
|
6
|
+
If you cannot read the helper's local disk, use `agentcut_list_files`, `agentcut_read_file`,
|
|
7
|
+
`agentcut_write_file` and `agentcut_copy_file`. These operate on the connected helper's workspace,
|
|
8
|
+
not your own current folder. Read before writing and pass the returned hash to preserve hand edits.
|
|
9
|
+
Use `baseHash: null` only for a new file. Call status after scene or timeline edits to check the build.
|
|
10
|
+
|
|
11
|
+
File scopes are `project`, `library` and read-only `toolkit`. Workspace `library/presets/...` is
|
|
12
|
+
scope `library`, path `presets/...`; `.agentcut/components/...` is scope `toolkit`, path
|
|
13
|
+
`components/...`; `.agentcut/runtime/...` is scope `toolkit`, path `runtime/...`. List preset files,
|
|
14
|
+
read their recipes and copy their scenes/assets into the project. Music, SFX and b-roll catalogs
|
|
15
|
+
use their existing search/add tools. Capture and contact-sheet calls return MCP images as well
|
|
16
|
+
as local paths, so visual review does not require a shared disk. Keep all creative and review steps.
|
|
17
|
+
For local app code, call `agentcut_sources`. Scope `source` reads only folders already allowed
|
|
18
|
+
with the helper's `--add-dir` option; pass the returned ID as `source` for list/read, or
|
|
19
|
+
`sourceRoot` for copy. Source folders stay read-only. MCP cannot grant itself new folder access.
|