Step types

A workflow is a numbered list of steps. Each step has a step type that decides what kind of work it does. Some step types run an AI agent; others are plain, predictable processing that gives the same result every time for the same input (this guide calls those non-AI steps).

There are thirteen step types. Their exact names (used by the assistant and in saved workflows) and the labels shown in the workflow builder are:

Step type	Builder label	Uses AI?	In one sentence
Agent	Agent	Yes	Runs one AI agent.
ReviewLoop	Review Loop	Yes	Scores earlier work and sends it back for another try if the score is too low.
Conditional	Conditional Branch	No (optional AI mode)	Jumps to one step or another depending on a condition.
ForEach	For Each	Yes	Runs one agent once per item of a list.
Parallel	Parallel	Yes	Runs several agents at the same time.
Extract	Extract	No	Trims or reshapes an earlier step's result so later steps read less.
VideoAnalyze	Analyze Video	No (optional speech and vision AI)	Studies a video: shots, pauses, speech, colour, green screens.
VideoCompile	Compile Video	No	Cuts the video and adds graphics, inserts, music, effects, grade, voiceover.
EditRoom	Edit Room	Yes	A panel of AI editors decides what to keep.
GraphicsRoom	Graphics Room	Yes	A panel of AI artists plans graphics and screen inserts.
ColorGradeRoom	Color Grade Room	Yes	A panel of AI colourists picks one colour look.
VideoGenerate	Generate Video	No (calls a paid video service)	Pays a video-generation service to make new clips.
Voiceover	Voiceover	No (calls a speech service)	Turns written lines into spoken narration audio.

There is no "render Remotion" step type. Remotion animations (the code-based animated videos ReelBolt produces) are written and rendered by certain AI agents inside Agent steps (and inside the GraphicsRoom step's final director call). The AuthorAgent renders a whole promo video; the MotionGraphicsPlanner and MotionGraphicsDirector render small graphics and the scenes shown inside tracked screen inserts. See Agents.

Editing steps in the workflow builder

On a computer the workflow page shows the steps as cards on a canvas, one under the other with arrows between them. Each card says what the step does in plain words ("Watches your footage", "Picks the best clips", "Writes the narration") and a few chips summarising its main choices (how many clips, the output shape, whether music or narration is on). A card never grows into a form:

• Click a card, or its Settings button, to open that step's settings in a panel on the right. The everyday choices come first; technical settings (codecs, limits, thresholds, decision-model options, the exact data format a step hands on) sit under Advanced, closed by default. It opens by itself when the assistant proposes a change there or a save was rejected because of it.
• Scroll with the mouse wheel or trackpad to move around the canvas. To zoom, hold Ctrl (Cmd on a Mac) while scrolling, pinch on a trackpad, or use the + and − buttons in the corner. The canvas fills the window, and the whole workflow is framed when it opens, so the last steps of a long workflow are easy to reach. Auto Layout puts dragged cards back in a neat column.
• A step card warns you about a missing choice, for example a music step in a project that has no audio files, or a footage step with no video chosen.

On a phone the steps are a simple list instead, with the same settings.

Settings every step has

Every step, whatever its type, has these settings (the names in backticks are how the assistant and saved workflows spell them):

• stepOrder — the step's number, starting at 1. Other steps refer to it by this number.
• label — a human-readable name shown in the builder and on run pages.
• stepType — one of the thirteen types above. Defaults to Agent.
• An agent. Agent, ReviewLoop and ForEach steps run the agent you pick. Non-AI steps and the three room steps still need an agent slot filled, and use a built-in placeholder for it: ExtractTransform for Extract, and VideoTransform for VideoAnalyze, VideoCompile, VideoGenerate, Voiceover, EditRoom, GraphicsRoom and ColorGradeRoom. The placeholder is never run as an AI; a room's real AI members come from the room's own settings.
• cacheMode — whether a later run may reuse this step's earlier result instead of redoing the work: Default (ReelBolt decides), Always, or Never. See Execution and outputs.
• A settings block for the types that need one, stored as JSON text in a field named after the type: videoAnalyzeConfigJson, videoCompileConfigJson, extractConfigJson, editRoomConfigJson, graphicsRoomConfigJson, colorGradeRoomConfigJson, videoGenerateConfigJson, voiceoverConfigJson, conditionalConfigJson, reviewLoopConfigJson.

Inside every settings block, names are written in camelCase (detectInsertRegions, not DetectInsertRegions) and choices are written as words in quotes ("mode": "Reencode"). Anything you leave out takes its default value.

How steps refer to each other

ReelBolt steps pass work forward. Most settings blocks point at an earlier step in one of these shapes.

Pointing at an earlier step's result (used by VideoCompile, the rooms, VideoGenerate and Extract):

{"from": "Previous"}

means "the step that ran just before this one", and

{"from": "Step", "stepOrder": 2}

means "step number 2". Prefer the numbered form whenever there is more than one step in between. The footage templates always use it, because "Previous" from a compile step would usually mean the graphics, music or effects step, not the editing decision. (Extract steps additionally accept {"from": "Accumulated"} and {"from": "ProjectFiles"}.)

Pointing at a video (used by VideoAnalyze's source):

{"kind": "ProjectFile", "projectFileId": "PUT-THE-FILE-ID-HERE"}

uses a video you uploaded;

{"kind": "StepOutput", "stepOrder": 3}

uses the video produced by step 3 (for example a render or a generated clip); and

{"kind": "PreviousStepOutput"}

uses the most recent video any earlier step in this run produced.

What an AI agent sees. An Agent step's agentInputContextMode decides which earlier results the agent reads:

• FullWorkflow — all earlier steps' results (older ones may be shortened to keep the prompt a sensible size).
• PreviousStepOnly — only the step just before.
• SelectedPriorSteps — only the steps you list (for example steps 1, 4 and 5).
• CustomMappedSubset — a custom mapping you write.

An agent that needs two earlier results (for example a graphics planner that needs both the video analysis and the editing decision) must use FullWorkflow or SelectedPriorSteps.

Your brief. The text you type in Your brief when you click Run is passed to every AI agent as extra instructions. It is optional unless the workflow has Require a brief before each run on.

Agent step (Agent)

An Agent step runs one built-in AI agent (such as VideoStoryEditor or AuthorAgent). The agent reads earlier results (see agentInputContextMode above), may use its tools (reading project files, writing code, rendering), and returns its answer. Built-in agents return a fixed answer shape that later steps rely on.

A custom agent can be put in an Agent step, but in this version the step is marked Skipped when the run reaches it, because custom agents cannot run inside workflows yet (see Agents).

• Uses AI: yes.
• Retries: if the agent fails or returns an answer of the wrong shape, ReelBolt tries again, telling the agent what went wrong, up to 3 attempts by default. An agent that runs out of time is not retried.
• Reuse on later runs: agents that only read (analysis agents, VideoStoryEditor, MusicSupervisor, SoundDesigner, ShotDirector, Colorist, VideoReviewAgent) may be reused from an identical earlier run unless the step is set to "cacheMode": "Never" (the templates set every creative step to Never). Agents that write code, write files or render (RemotionComponentTranslator, AuthorAgent, MotionGraphicsPlanner) are never reused.

Minimal step (as the assistant would write it):

{"stepOrder": 2, "label": "Decide which spans to keep", "stepType": "Agent",
 "agentDefinitionId": "ID-OF-THE-VideoStoryEditor-AGENT",
 "agentInputContextMode": "PreviousStepOnly", "cacheMode": "Never"}

Review loop step (ReviewLoop)

A ReviewLoop step runs a reviewing agent (usually ReviewAgent for promos or VideoReviewAgent for edited footage), reads its score out of 10, and decides whether to go back and redo work.

• minScore — the passing score. Below it, the run jumps back. Default 9 when not set (the templates use 8 or 9).
• loopTargetStepOrder — the step to jump back to (for footage templates, the editing decision, step 2; for promos, the AuthorAgent step).
• maxIterations — how many review rounds at most (templates use 2 or 3). When the last round still scores too low, the run continues and finishes with what it has; it does not fail.

When the run jumps back, the reviewer's comments are handed to the steps being redone, so they can fix what was criticised.

An optional reviewLoopConfigJson ({"enabled": true, ...}) lets a fast "decision model" skip the reviewer when it is very confident; it is off by default and needs an admin-configured decision provider.

• Uses AI: yes. Reuse on later runs: never.

Conditional step (Conditional)

A Conditional step looks at the previous step's result and jumps to one of two steps.

• conditionExpression — a test written against the previous result's fields, for example $.overallScore >= 8 after a ReviewAgent step. Field names must match the earlier result exactly as it appears on the run page. You can combine tests with && and ||.
• trueBranchStepOrder / falseBranchStepOrder — the step number to go to in each case. If the branch is not set, the run simply continues with the next step.
• An expression that cannot be worked out counts as false.

An optional conditionalConfigJson with "mode": "Decision" asks a decision model a yes/no question instead; the default is "mode": "Expression".

• Uses AI: no (unless Decision mode). Reuse on later runs: never.

For-each step (ForEach)

A ForEach step takes a list from the previous step's result and runs the chosen agent once for each item, several at a time.

• loopSourceExpression — where the list is, for example $.components (or $ when the whole result is a list).
• maxIterations — the most items to process; 0 means all.

The step's result is {"results": [...], "sourceCount": N}. If the list is empty the step completes without running the agent.

• Uses AI: yes. Reuse on later runs: never.

Parallel step (Parallel)

A Parallel step runs several agents at the same time on the same input and collects all their answers into one list, each tagged with the agent's name. List the agents in parallelAgentIdsJson (a JSON list of agent ids). The five code-analysis agents are good candidates, since they do not depend on each other.

A Conditional step after a Parallel step can test one agent's answer with $.AgentName.field, for example $.Review.overallScore > 8.

• Uses AI: yes. Reuse on later runs: never.

Extract step (Extract)

An Extract step is a non-AI step that shrinks an earlier result so later AI steps read less (and cost less). It can pick out a list and keep only certain fields ("operation": "Project"), look up chosen ids in an earlier list ("operation": "Resolve"), or list project files ("operation": "Files").

Example from the lean-context-promo template (keep at most 40 components, three fields each):

{"version": 1, "operation": "Project",
 "inputs": {"source": {"from": "Step", "stepOrder": 3}},
 "path": "$.components", "fields": ["name", "filePath", "responsibility"],
 "take": 40, "maxOutputChars": 8000, "expect": {"minItems": 1}}

• Uses AI: no. Retries: none (it would fail the same way again). Reuse on later runs: yes.

Analyze video step (VideoAnalyze)

A VideoAnalyze step studies a video and produces a compact description that AI agents can read, plus a full detailed record that the compile step uses later. It finds:

• shots (camera cuts), given ids s0, s1, ...;
• pauses / silences, ids g0, g1, ...;
• speech as text, ids t0, t1, ... (needs a transcription provider);
• per-shot look and sound facts (movement, brightness, colour, loudness, repeated takes);
• optionally, good places for graphics (ids p0, ...), music choices (m0, ...), sound-effect choices (x0, ...) and tracked green-screen regions (r0, ...).

It never cuts anything. It is non-AI, except for the optional speech-to-text and the optional picture descriptions, which call providers an admin configured.

Important settings (defaults in brackets):

Setting	Default	What it does
source	required	Which video to analyse (see "Pointing at a video" above).
sources	none	A list of several videos (takes, angles) analysed together into one set of ids. When set, source is ignored.
transcription	"Optional"	Speech-to-text: "Off", "Optional" (use it if a provider exists), "Required" (fail without it).
language	auto	Language hint for speech-to-text, for example "en".
vision	"Off"	AI descriptions of what is in each shot: "Off", "Optional", "Required". Costs one AI call per described shot, up to maxCaptionedShots (50).
locateSubjects	false	Finds where each shot's main subject is, so a vertical or square compile with outputFit: "Subject" keeps it in the frame. Uses the image AI (two calls per shot, up to maxSubjectShots, 40). subjectLabel names what to look for, e.g. "the phone"; blank finds the main subject. Shots with a tracked green screen don't need it.
emitOverlayPlacements	false	Find good spots for titles and lower-thirds (p ids). Needed for graphics.
offerMusicTracks	false	Offer every uploaded audio file as a music choice (m ids).
offerSfxClips	false	Offer every uploaded audio file as a sound-effect choice (x ids).
detectInsertRegions	false	Find and follow coloured screens (green screen) in the footage (r ids). Needed for tracked screen inserts.
insertRegionColor	"green"	Screen colour to look for: "green", "blue" or "magenta".
maxInsertPlatesPerFrame	1	How many coloured screens can be followed at the same time (1 to 16). Raise it for two phones in one shot.
measureInsertCurvature	true	Measure corners and curvature of each screen more precisely.
solveInsertChromaKey	true	Work out the exact screen colour so the insert can be cut to the screen's true shape (rounded corners, notch, a hand in front).
silenceThresholdDb / minSilenceMs	-34 / 350	How quiet and how long a pause must be to count.
sceneThreshold	0.30	How different two frames must be to count as a cut.
maxOutputChars	24000	Size limit of the description AI agents read.
maxDurationSeconds	1800	Videos longer than 30 minutes are refused.
derushPrePass	"Off"	Optional decision-model hint per shot; "Fast" currently behaves exactly like "Prior".

Minimal example (the template default; you must add the file id):

{"version": 1, "source": {"kind": "ProjectFile", "projectFileId": "PUT-THE-FILE-ID-HERE"}}

Photos work as clips. A PNG, JPEG, WebP, BMP, TIFF or AVIF image picked as a source is turned into a short moving shot: a slow push in, a pull back out and a pan, each in two short beats. The editor picks which move fits and whether it holds for one beat or two, so photos from a website or an Instagram account can be cut into a reel alongside video.

Choosing clips in the builder. In the step's settings, under Your footage, a first step offers your uploaded videos directly (there is no earlier step to take a video from). To cut several clips into one video, switch on Use several clips and pick them all at once in Choose clips…; they are used in the order listed, which you can adjust row by row, and Add another clip adds a row for one more uploaded file. Switching on Offer your music tracks or Offer your sound-effect clips in a project with no audio files shows a reminder to upload some first.

If the description grows past maxOutputChars, ReelBolt first makes each shot's details shorter, then leaves out whole optional lists (music choices, effect choices, green-screen regions), and only then leaves out shots. A left-out list cannot be used by later steps in that run, so for green-screen work check that the run summary shows the regions were offered (see the tracked screen insert recipe).

• Retries: none (re-running gives the same result; speech-to-text has its own small retry).
• Reuse on later runs: yes, an identical analysis is reused.
• Common errors: SOURCE_UNRESOLVED (no file selected, or the step you pointed at made no video), DURATION_EXCEEDED, INPUT_TOO_LARGE (over 2 GB), TRANSCRIPTION_UNAVAILABLE or TRANSCRIPTION_FAILED (only with "Required"), VISION_UNAVAILABLE / VISION_FAILED (only with "Required").

Compile video step (VideoCompile)

A VideoCompile step produces the finished edited video file. It takes the editing decision (which ids to keep), looks up their exact times in the analysis, cuts the video frame-accurately and, in the same pass, adds whatever extras you switched on. It is non-AI: every number comes from ReelBolt, never from an AI.

It always needs:

• decision — the step whose answer says what to keep: a VideoStoryEditor agent step or an EditRoom step, for example {"from": "Step", "stepOrder": 2}.
• analysisStepOrder — the number of the VideoAnalyze step, for example 1.

Main settings (defaults in brackets):

Setting	Default	What it does
mode	"Reencode"	"Reencode" cuts exactly and allows every extra. "StreamCopy" is fast but cuts only on keyframes, needs "allowKeyframeSnapping": true, and allows no extras.
outputFileName	"edited.mp4"	Name of the result. Left at "edited.mp4", the file added to your project is named after the workflow and the run instead, for example Spring teaser - run 3.mp4.
outputFormat	"Source"	Shape of the video: "Source" (the size of your largest clip), "Vertical" (1080x1920, for Reels, Shorts and TikTok), "Square" (1080x1080) or "Landscape" (1920x1080). Needs "Reencode".
outputFit	"Crop"	When a clip has a different shape from outputFormat: "Crop" fills the frame and trims the sides (a wide shot keeps its middle), "Subject" fills the frame but keeps the subject in shot (a tracked screen, or the subject the analysis step found with locateSubjects) — a clip whose subject is too wide for any crop, like an end card's headline, is shown whole instead, "Pad" shows the whole shot with bars, filled per canvasFill ("Black" or "Blur").
targetDurationSec	0	The length you want, in seconds (0 = no target). A longer cut is shortened to it, trimming the ends of the longest shots and never the last one or the narration. ReelBolt does not invent footage; if the video comes out more than 20% shorter or longer, the run tells you.
loudnessTarget	"Off"	How loud the finished video is: "Social" matches Instagram, TikTok and YouTube, "Web" websites, "Broadcast" TV. The balance between voice and music stays the same. Under Loudness in the builder.
registerProjectFile	true	Also add the result to the project's files, so it can be edited again.
minSegmentMs	250	Drop kept pieces shorter than this.
transitionPolicy	"Off"	Joins between cuts: "Off" (plain cuts), "AudioOnly", "Auto", "Expressive" (visible dissolves and dips chosen automatically), or "Editor": the editing AI picks each join from 56 transitions (slides, wipes, zooms, flashes, dips, iris and more) and how fast it plays.
programFadeInMs / programFadeOutMs	0	Fade the picture in at the start / out at the end.
enableGraphics + graphicsPlan	off / none	Add titles and lower-thirds from a MotionGraphicsPlanner or GraphicsRoom step.
enableInserts (+ graphicsPlan)	off	Put rendered scenes into tracked green screens. Reads the same graphicsPlan.
enableMusic + musicPlan or musicTrackProjectFileId	off	Add a music bed chosen by MusicSupervisor, or a specific file you name.
enableSfx + sfxPlan	off	Add sound effects planned by SoundDesigner.
enableColorGrade + colorGradePlan	off	Apply a colour look from Colorist or a ColorGradeRoom step.
enableGeneratedClips + generatedClips	off	Place AI-generated clips from a VideoGenerate step.
enableVoiceover + voiceoverStepOrder	off / none	Mix in narration from a Voiceover step (a step number, not a from reference).
voiceoverMode	OverMusic	What happens to the original speech under each narration line: leave it (OverMusic), turn it down (DuckDialogue, by voiceoverDuckDb, default -12) or remove it (ReplaceDialogue, for dubs).
enableCaptions + captionStyle	off / Subtitle	Burn the narration into the picture as it is spoken: Subtitle (a few words at a time), Karaoke (spoken words highlighted), Punch (one word at a time) or Bold (big words two or three at a time, the Reels/TikTok look). Karaoke and Punch need a transcription provider for word timing.
enablePickups + pickupVoiceoverStepOrder	off / none	Replace flubbed spoken phrases with re-recorded lines from a Voiceover step whose source is PickupPlan. The picture is not changed (no lip sync).
expect.minRetainedRatio	0.15	Refuse an edit that keeps less than 15% of the footage.

In the builder these appear under Format and length near the top of the step's settings: Shape (Original size, Vertical 9:16 — Reels/TikTok/Shorts, Square 1:1 — feed, Landscape 16:9 — YouTube), When a clip has a different shape (Fill the frame (crop the middle), Fill the frame, keep the subject in shot or Fit whole clip (bars), shown only for a fixed shape) and Target length (seconds). The extras read Add background music, Add sound effects, Add titles and graphics and so on; switching music or sound effects on in a project without audio files shows a reminder to upload some. Codec, preset, quality (CRF), cut padding, transition timing and the before-encoding checks are under Advanced.

Every "enable" switch defaults to off, and when off the result is exactly what it would be without that feature. Every extra requires "mode": "Reencode". When an extra cannot be applied (a missing plan, an unknown id, a bad file), it is skipped and the reason is written in the step's result; the edit itself still completes.

When the finished video is not quite what you asked for, the step's result lists warnings in plain words, and the run page shows them: some or most of the narration could not be placed, music or sound effects are switched on but the project has no audio files to choose from (upload some to the project's files), or the video came out much shorter or longer than targetDurationSec. If your clips have no sound of their own, narration, music and sound effects are still heard: ReelBolt adds a silent sound track for them to play on.

Minimal example (from video-derush-edit):

{"version": 1, "decision": {"from": "Previous"}, "analysisStepOrder": 1}

• Retries: none. Reuse on later runs: yes, when everything it reads is identical.
• Common errors: DECISION_UNRESOLVED (the decision step is wrong or missing), ANALYSIS_NOT_FOUND (wrong analysisStepOrder), EMPTY_KEEP (the editor kept nothing), UNKNOWN_ID, EXPECT_FAILED (too little kept), GRAPHICS_REQUIRE_REENCODE, INSERTS_REQUIRE_REENCODE, MUSIC_REQUIRES_REENCODE, SFX_REQUIRES_REENCODE, COLOR_GRADE_REQUIRES_REENCODE, GENERATED_CLIPS_REQUIRE_REENCODE, TRANSITIONS_REQUIRE_REENCODE, OUTPUT_FORMAT_REQUIRES_REENCODE (an extra or a fixed outputFormat was switched on with "StreamCopy"), CODEC_NOT_ALLOWED, ENCODE_FAILED.

Edit room step (EditRoom)

An EditRoom step replaces a single VideoStoryEditor with a panel of AI editors. By default three seats — PacingEditor, StoryEditor and CraftEditor — take turns discussing the analysed footage, a VideoEditDirector moderates, and the director then writes one final "what to keep" decision in exactly the same shape a single editor would, so the compile step treats it the same.

Setting	Default	What it does
view	none	The analysis to discuss, for example {"from": "Previous"} or {"from": "Step", "stepOrder": 1}.
seats	the three above	Your own seats: a list of {"name": ..., "persona": ...}.
rounds	2	Full passes round the table before the director speaks.
maxTurns	8	Hard turn limit (2 to 20).
fallbackToSoloEditor	true	If the panel fails, let one VideoStoryEditor decide instead.
persistTranscript	true	Keep the discussion so you can read it (for information only).

{"version": 1, "view": {"from": "Previous"}}

• Uses AI: yes, many calls. Retries: not retried as a whole (the director's final write-up has its own retries). Reuse on later runs: possible by default; the template sets Never.

Graphics room step (GraphicsRoom)

A GraphicsRoom step replaces a single MotionGraphicsPlanner with a panel of AI artists. By default four seats — LayoutArtist, TimingArtist, CopyArtist and InsertArtist — discuss the offered graphics spots and green-screen regions; a MotionGraphicsDirector moderates and then, in one final call, writes the graphics plan and renders the graphics and insert scenes with Remotion. During the discussion itself nobody can render; only the director's final call can. Its plan has the same shape as the single planner's, so the compile step's graphicsPlan can point at it.

Settings are the same as the edit room (view, seats, rounds, maxTurns default 10, ...), with fallbackToSoloPlanner (default true) instead of fallbackToSoloEditor. Point view at the analysis step, not at the editing decision:

{"version": 1, "view": {"from": "Step", "stepOrder": 1}}

An empty plan (no graphics) is a valid result. Reuse on later runs: never (it may render). Retries: not retried as a whole.

Color grade room step (ColorGradeRoom)

A ColorGradeRoom step has a panel of AI colourists — ToneArtist, PaletteArtist and ContinuityArtist by default — discuss the measured colour of each shot; a ColorGradeDirector then picks one look for the whole video as words only: look (None, Warm, Cool, Filmic, Vibrant, Muted, Mono), strength (Subtle, Normal, Strong), shadowTone and highlightTone. ReelBolt turns the words into the actual colour change at compile time. A single Colorist agent step produces the same shape and can be used instead.

{"version": 1, "view": {"from": "Step", "stepOrder": 1}}

fallbackToSoloColorist defaults to true. Retries: up to the normal agent retry count. Reuse on later runs: possible by default; the template sets Never.

Generate video step (VideoGenerate)

A VideoGenerate step pays a video-generation service (MiniMax, configured by an admin) to create new video clips. It is non-AI in the sense that it does not think; it sends prompts and collects clips. It works in two ways:

• "planSource": "Inline" (default) — you write one prompt in inlinePrompt (up to 7000 characters) and it makes takes clips from it. The first clip becomes the step's output video, so a later VideoAnalyze step can use it with {"kind": "PreviousStepOutput"}.
• "planSource": "StepRef" — it makes one clip per shot planned by a ShotDirector step (plan), for the compile step to place (generatedClips).

Setting	Default	What it does
maxSpendUsd	0 (must be set)	Hard spending cap for this step in US dollars. Must be above 0.
model	"MiniMax-H3"	Or "MiniMax-H3-Max".
resolution	"768P"	MiniMax-H3: 768P or 2K. MiniMax-H3-Max: 480P or 768P.
durationSeconds	6	Clip length for inline prompts.
ratio	"16:9"	One of 21:9, 16:9, 4:3, 1:1, 3:4, 9:16.
takes / maxClips	1 / 1	Clips per prompt or shot / most clips in total (both 1 to 4).
allowSourceMediaEgress	false	Permission to send frames of your own footage to the service.
loopBackBehavior	"Reuse"	On review-loop retries, reuse clips ("Reuse") or pay for new ones ("Regenerate").

{"planSource": "Inline", "inlinePrompt": "Slow aerial shot over a misty pine forest at sunrise",
 "durationSeconds": 6, "takes": 1, "maxClips": 1, "maxSpendUsd": 1.0}

Before anything is bought, ReelBolt checks this cap plus daily caps per project and across the whole platform; if any would be exceeded, nothing is bought and the step fails with BUDGET_EXCEEDED. Other errors: CONFIG_INVALID, INVALID_PROMPT, EGRESS_REFUSED, POLL_TIMEOUT. Identical earlier clips are reused for free unless you choose to regenerate. See the generated b-roll recipe.

Voiceover step (Voiceover)

A Voiceover step turns written lines into spoken audio with a speech service (Fish Audio, or a local OpenAI-compatible speech server, configured by an admin). It does not decide what to say.

• source — one of four kinds:
  • {"kind": "Inline", "lines": [...]} with lines you write, each with text, startSec (seconds on the original video's clock) and an optional voiceId;
  • {"kind": "ScriptScenes", "scriptwriterStepOrder": 8} to speak the narration a ScriptwriterAgent step wrote, scene by scene;
  • {"kind": "NarrationPlan", "narrationWriterStepOrder": 3} to speak the lines a NarrationWriter (or NarrationTranslator) step planned against the edit;
  • {"kind": "PickupPlan", "pickupPlannerStepOrder": 3} to re-record the sentences a PickupPlanner step named.
• defaultVoiceId — the provider's voice to use when a line has none.
• projectVoiceId — a voice from the project's Voices tab instead. A cloned voice is usable only while its recorded consent is in place; otherwise the step fails rather than speaking in another voice.
• voiceCasting — Off (default) or Decision: let the decision model pick one of the project's usable voices from their descriptions.
• pronunciationDictionary — Off (default) or Auto: tell Fish Audio how to say the product and library names the code analysis found (from a fixed list ReelBolt knows how to spell out).
• language — a note recorded on the result (the speech service hears the language from the text).
• providerId, model — optional overrides.

{"source": {"kind": "Inline", "lines": [{"text": "Welcome to the tour.", "startSec": 0}]},
 "defaultVoiceId": "YOUR-FISH-AUDIO-VOICE-ID"}

Warnings in the builder. The step's settings check the speech service a run would use right now. If it failed its last connection test, a red box says Narration will probably fail; if a working backup will be used instead, a yellow box says Using a backup voice service. Fixing the service is an admin task (see Inference providers). Under Voice, pick one of the project's voices, or follow the link to the project's Voices tab to add one; the model, voice ID, size limit, pronunciation help and automatic casting are under Advanced.

To hear it in an edited video, switch on enableVoiceover and set voiceoverStepOrder on the VideoCompile step; add enableCaptions to see it too. In a promo, the AuthorAgent picks the audio up by itself and sizes each narrated scene to the measured length of its line. See the voiceover recipe.

• Retries: up to the normal retry count. Reuse on later runs: yes.