product-demo-director
AGENTS.md38 starsChanged 4 months agoMITView on GitHub
Turn real product evidence and footage into a self-contained product story whose claims, cuts, and final file are verifiable. Direct the viewer's attention; do not produce a generic onboarding tour, feature inventory, or unsupported marketing reel. The customer, operator, or maker is the hero. The product creates the visible change. An agent may drive the production workflow, but must not become a substitute for product proof. - Read SKILL.md completely before creating, recutting, or judging a film. - Read docs/PRODUCTION_CONTRACT.md…
- Reads credentials
Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.
Of the AGENTS.md files we read, the 31,439 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.
In this repository2 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- product-demo-directorSkill
The file
Show its rules and text
We split this AGENTS.md into the rules it gives an agent
27 rules. About 1 is common, 17 are distinctive, and 9 are very specific to this project.
≈1
Common · estimate
Many files like this have a rule of this kind.
≈17
Distinctive · estimate
This author's own choices. Worth a look.
9
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- could be automatic: a setting could enforce it
- Very specific: line 14
Read SKILL.md completely before creating, recutting, or judging a film.
- Very specific: line 15
Read docs/PRODUCTION_CONTRACT.md when changing profiles, story beats, claims, authority, narration, source locks, camera continuity, …
- Very specific: line 17
Read docs/CAPTURE.md before recording a browser, terminal, or agent-operated workflow.
- Very specific: line 18
Read docs/EDITING.md before making editorial or pacing decisions.
- Very specific: line 19
Read docs/MOTION_GRAPHICS.md before making a motion-graphics (no footage) brand or launch film.
- Very specific: line 20
Read CONTRIBUTING.md before changing repository code or opening a pull request.
- Distinctive: line 22
Keep this file concise. Put detailed production rules in the documents above and update the closest source of truth when behavior changes.
- Distinctive: line 27
New public announcement: start from the default 60-second, single-focus project.
- Distinctive: line 28
Dramatic transformation: use the Save-the-Cat format and keep the product causal.
- Distinctive: line 29
Consequential or investor workflow: use the causal-workflow format so input, decision, receipt, and outcome remain connected.
- Very specific: line 31
Existing authored audio: use render; do not synthesize replacement narration.
- Distinctive: line 32
Code or contract change: add focused tests and do not use a generated MP4 as source code.
- Distinctive: line 34
Do not default to a walkthrough.
- Distinctive: line 64
Do not run it unless the user asks for installation.
- Distinctive: line 69
Use one audience, one central change, and one release promise.
- Distinctive: line 71
Prefer full-frame sequences.
- Distinctive: line 73
Use live product states for causality.
- Distinctive: line 75
Keep one continuous source/session in one clip with connected zoom regions.
- Distinctive: line 85
Never invent product behavior, claims, metrics, receipts, customer data, or UI states.
- Distinctive: line 88
Treat browser sessions, terminal output, logs, manifests, and recordings as potentially sensitive.
- Common: line 90
Never put production credentials in capture plans, commands, transcripts, fixtures, or commits.
- Distinctive: line 93
Keep consequential human decisions visibly human.
- Distinctive: line 97
Do not let an AI judge waive preflight, hash, frame, decode, audio, black-frame, silence, evidence, or authority failures.
- Distinctive: line 110
Do not hand-edit generated props, build plans, QA reports, artifact manifests, hashes, or final media to make a gate pass.
- Very specific: line 112
Do not overwrite source media during finishing.
- Very specific: line 114
Never force-add generated media.
- Distinctive: line 127
Every visible claim and authority boundary is supported.
≈ “Common” is an estimate: at least 1 in 5 of the 40,246 AGENTS.md 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
Product Demo Director agent instructions
Mission
Turn real product evidence and footage into a self-contained product story whose claims, cuts, and final file are verifiable. Direct the viewer's attention; do not produce a generic onboarding tour, feature inventory, or unsupported marketing reel.
The customer, operator, or maker is the hero. The product creates the visible change. An agent may drive the production workflow, but must not become a substitute for product proof.
Read before acting
- Read
SKILL.mdcompletely before creating, recutting, or judging a film. - Read
docs/PRODUCTION_CONTRACT.mdwhen changing profiles, story beats, claims, authority, narration, source locks, camera continuity, artifact binding, or delivery gates. - Read
docs/CAPTURE.mdbefore recording a browser, terminal, or agent-operated workflow. - Read
docs/EDITING.mdbefore making editorial or pacing decisions. - Read
docs/MOTION_GRAPHICS.mdbefore making a motion-graphics (no footage) brand or launch film. - Read
CONTRIBUTING.mdbefore changing repository code or opening a pull request.
Keep this file concise. Put detailed production rules in the documents above and update the closest source of truth when behavior changes.
Choose the workflow
- New public announcement: start from the default 60-second, single-focus project.
- Dramatic transformation: use the Save-the-Cat format and keep the product causal.
- Consequential or investor workflow: use the causal-workflow format so input, decision, receipt, and outcome remain connected.
- Existing authored audio: use
render; do not synthesize replacement narration. - Code or contract change: add focused tests and do not use a generated MP4 as source code.
Do not default to a walkthrough. Establish what changed, why it matters, how the product caused it, and what proves the outcome.
Standard commands
# Workstation setup
./pdd setup # add --with-capture only when browser capture is needed
./pdd doctor
# Create and finish a default announcement
./pdd new projects/my-demo --name "My Product"
# Add source footage under projects/my-demo/assets/ and author the production contract.
./pdd demo projects/my-demo
# Existing narration/music, or optional advisory review
./pdd render projects/my-demo
./pdd demo projects/my-demo --judge
# Repository validation
./pdd test
git diff --check
git status --shortCopy .env.example to .env for local provider keys. The default auto-paced announcement requires
ELEVENLABS_API_KEY; Gemini is optional for non-auto-paced TTS, music, and advisory judging. Never
print keys, place them in commands or transcripts, or commit .env.
./pdd install-skill links this checkout into Codex and Claude Code for use from other repositories.
Do not run it unless the user asks for installation; it is not required when working in this clone.
Directing contract
- Inspect the product, supplied footage, repository state, and evidence before writing the story.
- Use one audience, one central change, and one release promise. Make the product clear by seven seconds for an announcement unless the selected profile says otherwise.
- Prefer full-frame sequences. Split screen is disallowed by the default single-focus contract and may appear only in an explicitly opted-in legacy profile.
- Use live product states for causality. Stills may explain a sourced fact but cannot prove a state transition.
- Keep one continuous source/session in one clip with connected zoom regions. Never simulate progress by replaying a conclusion or resetting the same screen with a new crop.
- Motivate every cut and zoom with a new idea, action, state, or piece of evidence. Keep relevant UI, people, and cursors inside safe framing.
- Author narration as complete thoughts. For auto-paced projects, generate alignment, run
pace.py, and place picture cuts between thoughts—never through them. - End on concrete product proof and one CTA. Do not replace proof with a summary card.
Truth, capture, and safety
- Never invent product behavior, claims, metrics, receipts, customer data, or UI states.
- Bind public claims to project-relative evidence through the production contract. A URL alone is not a substitute for the local evidence required by a strict artifact.
- Treat browser sessions, terminal output, logs, manifests, and recordings as potentially sensitive. Remove identifiers at the source before capture; do not blur-patch a final render.
- Never put production credentials in capture plans, commands, transcripts, fixtures, or commits.
- Capture authenticated products only with a dedicated seeded or staging account. Keep reusable browser state outside the repository and invalidate it after recording.
- Keep consequential human decisions visibly human. Do not portray an agent as having authority it did not have.
- A generated workbench replay must come from an allowlisted, committed session and be labeled as a curated replay. Do not present it as a recording of proprietary Codex or Claude UI.
- Do not let an AI judge waive preflight, hash, frame, decode, audio, black-frame, silence, evidence, or authority failures.
Files and repository boundaries
script.json,brand.json,product.json,shoot.json, evidence files, and capture-source manifests are editable production sources.- Inspect the Git worktree before editing and preserve unrelated user changes.
pddis the friendly orchestrator.tools/contracts.pyowns deterministic production policy; the remainingtools/files implement individual stages.engine/src/is the Remotion composition;tests/contains Python contract regressions.projects/,artifacts/, per-projectaudio/andout/, captures, and generated media stay local and are ignored by Git.- Do not hand-edit generated props, build plans, QA reports, artifact manifests, hashes, or final media to make a gate pass. Fix the source and rerun the pipeline.
- Do not overwrite source media during finishing. Write
demo-final.mp4and its inherited artifact beside the creative render. - Never force-add generated media. The only exception is a small, intentional, reviewed public
fixture with redistribution rights and an
ASSET_PROVENANCE.mdentry.
Definition of done
A film is not finished until:
- Strict preflight passes on the final authored sources.
- The render, finishing pass, and strict final-media QA succeed with
--require-artifact. - The final MP4, artifact manifest, source hashes, exact frame count, duration, streams, and decode all refer to the same file.
- Opening, cut boundaries, critical ranges, proof, CTA, and ending have been visually inspected; narration is not clipped and the protected tail remains.
- Every visible claim and authority boundary is supported. Any optional judge runs only afterward.
A repository change is not finished until ./pdd test and git diff --check pass, user-facing
behavior is documented, relevant regression tests exist, and no secrets, customer material, local
paths, or generated production files entered the diff.
Code Review Rules
Flag changes that weaken strict contracts, permit project or symlink escapes, detach a final file from its artifact, treat advisory model output as a gate override, crop away source truth, expose credentials or private footage, or change frame math without a regression test. Prefer a narrow fix to disabling a production check.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
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.

