clearframe-direction
Skill1 starsChanged 23 days agoMITView on GitHub
Direct a ClearFrame film from a document, report, transcript, podcast or recording plus a little direction — ingest the material, find the story, choose a treatment and references, plan pictures per beat, critique, and review with a fresh reviewer. Use when someone hands over source material ("turn this into a video") rather than a finished script.
In this repository22 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- clearframe-blocksSkill
- clearframe-canvasSkill
- clearframe-datavizSkill
- clearframe-engineSkill
- clearframe-explainSkill
- clearframe-inspirationSkill
- clearframe-integritySkill
- clearframe-librarySkill
- clearframe-motionSkill
- clearframe-reviewSkill
- clearframe-sceneSkill
- clearframe-scriptSkill
- clearframeSkill
- clearframe-workplaceSkill
- gemini-imageSkill
- gemini-omniSkill
- gemini-ttsSkill
- lyria-musicSkill
- runway-videoSkill
- veo-videoSkill
The file
We split this skill into the rules it gives an agent
7 rules. About 0 are common, 1 is distinctive, and 6 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
6
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 18
Write these four lines at the top of DIRECTION.md:
- Very specific: line 29
If there is a reference video, run clearframe reference VIDEO.
- Very specific: line 34
Use clearframe directions research or directions podcast to compare starting combinations.
- Very specific: line 55
Write the storyboard; clearframe critique DIR flags deck-like runs, static holds, dense text, weak hooks and missing causal links.
- Very specific: line 62
Story: Read only the vo lines: is it an argument with a turn, or a list?
- Very specific: line 64
Film, not slides: Check the eight tells in skills/clearframe-motion/SKILL.md: a card per line, build-then-freeze, a heading on every scene, …
- Distinctive: line 67
Truth: Every figure sourced and dated.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
From material to film
The goal is not to summarise the document on slides. It is a short film with a question, a turn and a payoff, where every beat has its own picture. Read skills/clearframe-motion/SKILL.md for film grammar (the eight tells of a slideshow, shots, camera, continuity), docs/ideas.md for pictures, docs/canvas.md for drawing, and the clearframe-script skill for narration.
1. Ingest
- Ideas/documents with brand assets:
clearframe start DIR --idea "..." --document report.md --brand brand.json --playbook NAME. Read BRIEF.md, EVIDENCE.md and BRAND.md. It copies local assets and source hashes, preserves sample attribution and leaves the story to you. Seedocs/intake.md. Usepipeline DIR --draft --scale 0.5for a retained smaller first review, then check full-resolution output before delivery. - Research reports (markdown):
clearframe ingest DIR --markdown report.md. It writesBRIEF.md: every figure with its sentence and source, tensions, questions, quotes, chart-ready tables and the source list, plus a research-digest storyboard to rewrite. Convert PDF/Docs to markdown or text first (pdftotext -layout,textutil -convert txt). - Recordings (podcasts, interviews, talks):
clearframe ingest DIR --audio episode.wav [--script script.txt] [--from s --to s] [--vertical]. Beats replay the recording gaplessly with measured word timings: without--words, local Whisper measures them (free; seconds with whisper.cpp). A script (HOST: …lines or JSON turns) provides spelling and speakers.--words words.jsonimports timings you already have (Whisper JSON, or paid Gemini transcription). Thenclearframe paper DIR: the paper edit (chapters, timecodes, speakers, the picture each beat is meant to get, the transcript) is the plan the person reads. The recording is the narration: keep it as recorded unless they ask for cuts (--suggest-cutslists filler words, doubled words and long pauses as proposals; struck words come back withcut --paper). Imported beats start as captions; give each idea the guest explains a picture of its own (the paper edit suggests one per beat).examples/podcast-picturesis the worked clip to copy: itsbuild.mjsreplaces the props of each beat by id, never the words or the timing.
Long recordings (several minutes) are not long short films. The beat counts and per-minute quotas below are for 60–90 s films. For a long recording, choose a small number of deliberate visual ideas (a world for the central mechanism, the few numbers that matter, speaker plates), mark where pictures will go with placeholder, and leave most beats as their captions for the rough cut. The person's notes on that rough cut decide which other moments deserve a picture. Work chapter by chapter (preview DIR --chapter NAME) and say how much of the film is covered.
2. Find the story (before any visuals)
Write these four lines at the top of DIRECTION.md:
- Question. What does the viewer want to know?
- Misconception or tension. What do they believe now? Starting from the wrong belief teaches better than a clean exposition.
- Turn. The "but…" that changes the picture.
- Payoff. What they understand or do at the end, and the last image.
Then choose 3–5 claims from the brief that carry the answer. Everything else goes to the source card. Link beats with but or therefore, never "and then". Open a loop in the first 5 seconds and close it at the peak. A 60–90 s film has roughly 12–18 beats.
3. Choose the look
clearframe treatmentslists art directions:workplace(benefits, policy, training, client films),business(reports, architecture and PR walkthroughs),editorial,research-study,explanation,sketchbook,blueprint,audiogram,keynote,cutpaper. Start withnew DIR --playbook NAME --treatment ID(or apply one to ingested material), then adapt.- Some playbooks start from shots, not cards:
journey(one drawing explored stop by stop),product-reveal(a dark stage, a part in close-up, specs as inserts, the whole),title-sequence(cut-paper shapes on the beat, ending on the title). - If there is a reference video, run
clearframe reference VIDEO. ReadREFERENCE.md, opensheet.pngandopening.png, and write Keep (rhythm, camera, type roles, transitions) and Change (brand, copy, subject) inDIRECTION.md. The fewer creative decisions left to guesswork, the better the film. - The user's direction ("make it feel like a Vox explainer", "brisk for a phone", "calm and warm") maps to a treatment plus overrides (palette,
type, motion preset). Say which you chose and why. A request for a trailer or film look is answered with the story's own pictures, not a look: the library has no grades, letterbox or flash cuts.
Keep the creative space open
Use clearframe directions research or directions podcast to compare starting combinations. --direction ID works with new, start and ingest; explicit playbook, treatment and theme choices override it independently. Add custom JSON profiles with --library DIR when useful, or skip profiles entirely.
Before authoring, propose three materially different approaches from the actual source in DIRECTION.md: change the narrative entry, visual metaphor, camera/edit grammar and rhythm. At least one should depart from the nearest preset. Choose one for the audience and evidence; do not rotate palettes and call that variation. Invent native canvas scenes and combine arcs when the story needs them. Reuse reliable rendering primitives without repeating the same film. If the user requested a one-shot film, decide and record these choices without a new approval stop.
For product or feature launches, connect the audience's problem to a specific behavior, a visible before/after proof and a payoff. Keep literal interface evidence distinct from conceptual imagery. Use approved product captures and brand assets; a drawn workflow panel is not a screen recording. New native scenes remain open-ended; the playbook does not limit the concept.
Show numbers flat: a stat, kpis, delta or chart block, or a canvas chart that morphs across cuts. Keep the domain fixed, the source visible and final labels aligned with settled values. Prefer a short reveal or outline emphasis over perspective, bounce or repeated looping of data. See the clearframe-dataviz skill.
Recording imports apply a treatment only to visual settings. The profile's playbook is a reference, never a replacement for recorded words, source slices, speakers or measured timing. Keep those when replacing kinetic captions with source-specific drawings. See docs/source-playbooks.md.
4. Plan pictures
First decide whether the film is one world or a sequence of scenes. When the material describes a place, a process, a chain of causes or a system, draw one world (props.world) and move the camera through it: stations in story order (up for the sky, down for a plan view, onward for the next stop), the world changing state with the story (a behind sky lighting up, a roof filling with heat), and a final pull-back to the opening shot, changed. Interrupt it only for a turn (a kinetic card on a colour block). examples/night-city shows the pattern end to end.
Fill the beat plan table in DIRECTION.md: purpose, picture, shot scale and landing word. Never three identical scales in a row. Every cut needs a reason: a world move, a morph, a match cut, a push or panel at a chapter, or a cut between two moving cameras. For places and journeys, build three planes with z and let the camera dolly or truck; rack focus on the word that shifts attention. Mechanisms to start from: sketch route|pipeline|network|balance|versus, and the workplace pictures (match, vesting, deadline, limit, portal…). Use docs/ideas.md: hooks, scale, mechanism, change, tension, hidden-in-the-average, people, turns, endings. Per minute, aim for:
- at least one drawn mechanism (canvas);
- one colour-block punctuation;
- cuts within a sequence, carrying an object across by id; at one or two turns an
into(through an element of the last scene) or apush; - imagery (plates) where the world matters;
- kinetic type where the words are the picture.
5. Build, critique, look
- Write the storyboard;
clearframe critique DIRflags deck-like runs, static holds, dense text, weak hooks and missing causal links. Fix, then run it again. - Voice:
voice DIR --draft(the whole narration in one continuous take, one shortvoice.style), thenalign DIR --whisperfor measured words. The final voice is Gemini 3.8 TTS in the same single take (gemini-ttsskill). sheet DIR --draftand open it;still --gridto place art;render DIR --draft.
6. Fresh review
Run reviewer rounds on the fine cut, after the person has seen the rough cut and their notes are in (one-shot: after your rough-cut decision); polishing every beat before anyone has seen motion is the most expensive way to find out a structural note. Ask a separate reviewer (a fresh subagent, with no authorship bias) to judge the sheet and draft against DIRECTION.md using this rubric. Collect specific, actionable notes:
- Hook. Does the first frame make you want the second? Is the question clear by 5 s?
- Story. Read only the
volines: is it an argument with a turn, or a list? Does each beat follow from the last? - Pictures. Does each beat show a different idea? Are there three similar frames in a row? Is anything a heading over bullets?
- Film, not slides. Check the eight tells in
skills/clearframe-motion/SKILL.md: a card per line, build-then-freeze, a heading on every scene, a flat camera, a small subject, the same grammar, a screen-flat image, an edit set only by the voice. - Motion. Does something move in every held frame? Do reveals land on the stressed words? Are
pushandpanelcuts only at turns? - Craft. One hero per frame; type roles (sans claim, serif feeling word, mono facts); margins; readable at phone size.
- Truth. Every figure sourced and dated; estimates labelled; nothing implies data it does not have.
Give the reviewer the sheet,
review DIRframes (strip.png),qa DIR'stimeline.png,phone.pngand findings, the video and thevolines, and ask for a verdict, the five highest-impact changes, and visible faults with timestamps. Keep the same reviewer across rounds (continue it rather than starting fresh) so it can say which notes are resolved. Separate its notes into tool problems (fix in the engine, and every film benefits) and film choices (fix in the storyboard). Apply the notes, re-sheet, and render the final only when the reviewer has no structural notes left. Four rounds turned the night-city example from narrated slides into a designed film.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
Reviews can't be read right now.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

