Build a Tool with AI, Then Let AI Use It

One Tuesday evening, around 8 p.m., I received the last video testimonials needed for a project I had accepted a few weeks earlier. A first complete version was exported that night. It ran five minutes and thirty-one seconds.

In between, I had examined three editing proposals, refined my criteria, chosen a narrative structure, worked on the sound, reframed images and added visual elements. During that same production, I was also developing the software I was using to make it.

The experience centred on Raccord!, a macOS editing application I developed with Codex. A few days later, I asked an AI agent to produce two promotional films for VitaScriptor with the same tool, providing the direction and feedback on its proposals.

These two uses let us examine a question that extends beyond editing: how is work organized when we can build our own tools with AI, then delegate much of their use to it?

Preparing the Means Before All the Inputs Arrive

I had accepted the project in September. Four initial videos were available; the remaining testimonials, including several important contributions, would arrive after my vacation. My original plan was to use Premiere. I eventually decided to experiment with developing a personal tool suited to this production.

The first media files helped me develop the essential functions: importing, navigating, selecting passages, working on sound, adding titles and producing a film. Transcription became central to this workflow.

My editing experience informed some choices. I wanted navigation inspired by older jog/shuttle controls: finding a cut precisely, moving backwards or forwards at variable speed, and hearing the sound follow the movement. The Mac’s trackpad could serve as the control without additional equipment.

The development records we found begin on October 4, 2026. They show a progression through stages, trials and corrections. Some of this preparation took place in the gaps between my other activities, while the material needed for the film was still incomplete.¹

By the time the last testimonials arrived, I already had a working capability built around the first files. Looking back, I estimate that I personally invested the equivalent of about one working week in the project, spread across small, intermittent work sessions over several weeks. This is my own rough estimate, without a precise breakdown between development, testing and production.²

From Spoken Words to a Proposed Narrative

A large part of editing testimonials consists of discovering the material: listening, finding interesting passages, understanding their relationships, avoiding repetition and building a progression. This was the editorial dimension in which I wanted AI to participate.

Raccord! can export transcripts with references to the corresponding videos and passages. I could give an AI my intention, criteria and available content, then ask it for an editing proposal.

To make the proposal usable, we added a structured exchange format. The AI identifies which sentences to keep and in what order. Raccord! retrieves their boundaries in the sources, presents a preview and creates the clips when the proposal is applied.

A discussion about the narrative can thus lead to an assembly that can be watched and revised. Precise passage references make it easier to move from analysing speech to arranging it audiovisually.

I tried three edits while refining my criteria with the AI. The third list was usable as it stood for the structure of the spoken excerpts. I then added shots without dialogue, intertitles and graphics, and continued the technical adjustments. The assessment of narrative coherence is my own; it has not been independently evaluated.

AI’s contribution therefore also lies in exploring several possible organizations of the material. My role was to express the intention, assess the proposals and specify what needed to change.

Difficulties in Use Become Design Questions

The software continued to evolve during production and finishing. Several changes can be traced to specific difficulties.

A treatment called “Clear Voice” had been added, but I found its effect barely noticeable on one video. In clarifying the problem, I explained that surrounding conversations made it hard to hear the interviewee. This observation led to a voice-separation trial, then a choice between two isolated voices. On the passage tested, one was indeed the voice I wanted to retain.

Listening had made the need clearer. A technically functional setting was still doing little for the problem at hand.

Transcription prompted other adjustments. Some testimonials alternated between French and English. A bilingual mode was added, then corrected using a real file. Building an assembly from transcript sentences also revealed words being cut off at their ends; we introduced adjustable margins around the selected passages.

I requested bilingual intertitles and a way to show a different image while keeping the speaker’s voice. Later, the audio delay after an insertion prompted a change to how previously processed sound was reused.

Subtitle review revealed an interface need. To correct a sentence while watching, I had to return to the source file, find the language and locate the passage. A command was added to open the subtitle currently displayed. An export to a spreadsheet, with a column for corrections, then made human review easier.

These changes form a loop: use the tool, observe a difficulty, clarify the need, modify the tool and resume the task. Editing provides a testing ground for the software that makes it possible.

A Result Delivered in Stages

The first retained export is dated October 28 at 3:38 a.m. It runs 5 min 31 sec and has no displayed subtitles. Two exports from October 29 keep that duration and add subtitles in the opposite language: English when the speaker uses French, and French when the speaker uses English.³

The first complete version was therefore followed by finishing work. Taking the approximate arrival of the final media at 8 p.m. as the starting point, the retained export comes about seven hours and forty minutes later. That interval includes exchanges, processing and waiting; it does not measure my active working time and comes on top of the earlier work.

The reports still contained warnings about the presentation of some subtitles and audio limitations. Producing a completed file and evaluating its quality remain two stages of the work.

Then Entrusting Production to AI

The second experience addressed another need: introducing VitaScriptor in a short video about memory and passing stories on. I had provided the direction in a conversation and requested a narration. After several voice trials, I suggested that the agent produce a first edit with Raccord!, using real or generated images.

This time, the agent handled almost all the audiovisual execution. It reused the text developed in the conversation, generated the narration, created eight illustrations for the first edit, prepared the shots and their movements, added the closing card and exported the film. At my request, it then composed piano music and mixed it with the voice.

My contribution was mainly direction and evaluation. A voice lacked naturalness; a pronunciation differed between two occurrences; a correction introduced an unconvincing pause. I described what I heard, and the agent revised the production.

The same process applied to images. The first shots evoked memories but showed little of the service itself. We added real screenshots of VitaScriptor and a published story. I requested framing that gave more space to the text; the agent captured the screens and changed the edit.

I supplied voice references and a visual reference, provided the necessary access and approved versions. Producing the elements and carrying out the changes were entrusted to the agent.

To make the film, it combined automation of Raccord!’s engine with use of its interface. The result remains available as an editable project with embedded media. The personal software was becoming a set of capabilities the agent could call on.

I then requested an English version. The agent adapted the text, prepared screenshots in English, generated a narration and reworked the edit. My feedback concerned the accent, voice variations and pacing. After several trials, I chose a version whose pace I found even more effective in English.

The deliverables include a 60-second French film and a 44-second English film, each accompanied by a Raccord! project and subtitles. Both films appear below. Preparing thumbnails and descriptions was also delegated. Distribution involved some manual actions on my part, including an initial publication and account access.

I could ask for a more measured voice or a more representative shot. The agent looked for a solution, produced a new version and presented it to me. Direction and evaluation remained with me; much of the production work was delegated.

What These Experiences Open Up

The first change concerns preparation. Some activity can move ahead while inputs are still missing. The first media helped build the means to handle those that would arrive later. This experience extends Bifurcations’ reflection on software before the decision.

The second concerns tools. A difficulty can become a request for a modification immediately relevant to the work in progress. That ability requires trials, corrections and maintenance. It gives concrete ground to the proposal to standardize systems, not people.

The third concerns delegation. In the first case, AI proposes an assembly from transcripts. In the second, it uses the software and other tools to produce the elements, build the films and carry out corrections. A capability created for one task becomes reusable for another production.

My judgement remains necessary: setting criteria, recognizing the desired voice, assessing a cut or asking for different framing. In the VitaScriptor production, my attention is more concentrated on direction and evaluation. It puts into practice the distinction between conception, execution and judgement examined in “A Candide Without Voltaire?”.

This study describes two situated experiences. It measures neither a comparative gain over Premiere, nor the full cost of the software, nor a percentage of autonomy. For Bifurcations, it opens an applied research question: under what conditions does the ability to build and adapt our tools, and delegate their use, expand the ways we can act?

Find Raccord! and the other tools in Projects.

Method and attribution — This retrospective study cross-checks working conversations against the projects, export reports and retained files. Assessments of use are the author’s own. Stéphanie Savard contributed the experience, direction, decisions and feedback. The AI assistant contributed to reconstructing the records, analysis, structure, writing, revisions and English translation; its intervention extends beyond stylistic presentation. The sharing cards were generated with ImageGen.