# AGENTS.md — Build guide for ThumblifyAI

## Project scope
Draft thumbnail concepts from a video title and audience, review text/layout and optionally request an image from a configured provider. Compare candidate images and export a chosen image at explicit dimensions.

Catalogue verdict: kinda. A basic AI thumbnail generator can be built quickly using existing image models, but recreating ThumblifyAI requires much more than connecting an image API. The product's advantage comes from specialized thumbnail-focused system prompts, AI workflows, creator-focused features, thumbnail inspiration, style matching, AI customization features, and continuous iteration around generating clickable YouTube thumbnails.
Use the implementation prompt below to define the deliverable. Complete each phase's acceptance checks before extending the scope.

## Working agreement
- Inspect the repository and its existing instructions before choosing paths, dependencies or commands. Keep one coherent stack and explain changes to the proposed architecture.
- Plan a vertical slice that accepts a real input and produces the useful output described below. Persist only the state the prompt calls for; respect memory-only and upstream-managed workflows. Use fixtures only when they are clearly labelled.
- After scaffolding, document the actual install, development, check and build commands in README and keep them synchronized with the package or project manifest. Do not report commands as successful unless they ran.
- Work in small steps. At handoff, list implemented flows, checks actually performed, remaining blockers, and any credentials or provider setup the owner must supply.
- Do not publish, spend money, contact customers, delete source data or run irreversible migrations without the project owner's authorization.

## Prerequisites
- A supported Node release, a writable local data directory and a separate backup location. Bind to localhost; remote use requires authentication and HTTPS first. Optional AI generation needs a provider key, a usage budget and approval to send the selected material. A selected image-generation provider with an authorized key, render budget and owned/licensed reference imagery.
- Implementation components: Node.js, TypeScript and Express with server-rendered HTML and small browser modules. SQLite through better-sqlite3 with migrations, prepared statements and a single background worker. A single provider SDK with a configured supported model, schema-validated responses and revisioned prompt templates.
- Scope boundary: Guaranteed click-through improvement and universal model access are excluded.

## Stack and architecture
- Node.js, TypeScript and Express with server-rendered HTML and small browser modules.
- SQLite through better-sqlite3 with migrations, prepared statements and a single background worker.
- A single provider SDK with a configured supported model, schema-validated responses and revisioned prompt templates.
- Domain model: video briefs, owned reference images, proposed thumbnail concepts, approved prompts and provider render receipts

## Security and data integrity
- Reject unexpected origins and unbounded request bodies even on localhost. Keep credentials outside the database export and redact sensitive text from logs. Treat source documents as untrusted data, prevent them from changing tool permissions, and require review of factual claims before publication.
- Correctness boundary: Paid renders require a reviewed prompt and usage cap; no fabricated performance score or unauthorized real-person impersonation.
- Use short SQLite transactions and persist job state before starting work. Give retries stable operation IDs; report incomplete or unknown results instead of silently repeating them.
- Use a consistent SQLite backup and an attachment manifest. Export portable JSON/CSV, then restore to a new directory without overwriting the original data.

## Agent implementation rules
- Project rule — data model: video briefs, owned reference images, proposed thumbnail concepts, approved prompts and provider render receipts
- Project rule — preserve this invariant: Paid renders require a reviewed prompt and usage cap; no fabricated performance score or unauthorized real-person impersonation.
- Project rule — acceptance evidence: Reject a concept and make no image request; a provider timeout enters reconciliation rather than charging for an automatic duplicate render.

## Optional agent skills and references
- Optional external skill: [copywriting](https://github.com/coreyhaines31/marketingskills/blob/main/skills/copywriting/SKILL.md) — Write landing pages and product copy grounded in the intended audience, product value and a clear next action. Review its instructions and compatibility before use; it does not grant deployment, data-access or publication permission.
- Optional external skill: [web-design-guidelines](https://github.com/vercel-labs/agent-skills/blob/main/skills/web-design-guidelines/SKILL.md) — Review web interfaces for accessibility, keyboard focus, forms, navigation and interaction quality. Review its instructions and compatibility before use; it does not grant deployment, data-access or publication permission.
- Optional external skill: [sharp-edges](https://github.com/trailofbits/skills/blob/main/plugins/sharp-edges/skills/sharp-edges/SKILL.md) — Review security-sensitive APIs and configuration for dangerous defaults and easy-to-misuse interfaces. Review its instructions and compatibility before use; it does not grant deployment, data-access or publication permission.

Read the linked SKILL.md and its dependencies before adding a skill. Select only the skills matching this project's runtime and task; their documentation does not supply API access, credentials or approval to perform external actions. Pin the reviewed revision where the tool supports it. Follow the chosen agent's documented project-level installation mechanism.

## Distribution ideas
These are optional planning notes. Obtain the owner's approval before publishing or contacting anyone.
- Demonstrate the actual ThumblifyAI-inspired workflow with owned or clearly labeled sample data: Draft thumbnail concepts from a video title and audience, review text/layout and optionally request an image from a configured provider. Compare candidate images and export a chosen image at explicit dimensions.
- Publish a reproducible walkthrough with this observable result: Reject a concept and make no image request; a provider timeout enters reconciliation rather than charging for an automatic duplicate render.
- Explain who can operate this scoped tool, its setup and ongoing costs, and these remaining product gaps: Guaranteed click-through improvement and universal model access are excluded. Avoid guaranteed savings, performance scores or implied endorsement.

## Engineering roadmap
1. Phase 1 — Scope and fixtures. Implement this bounded workflow: Draft thumbnail concepts from a video title and audience, review text/layout and optionally request an image from a configured provider. Compare candidate images and export a chosen image at explicit dimensions. Record prerequisites, select representative user-owned fixtures and document the unsupported features: Guaranteed click-through improvement and universal model access are excluded.
2. Phase 2 — Durable model. Model video briefs, owned reference images, proposed thumbnail concepts, approved prompts and provider render receipts Add migrations or a versioned document format, explicit validation, stable IDs and a visible import-error report. Preserve this rule: Paid renders require a reviewed prompt and usage cap; no fabricated performance score or unauthorized real-person impersonation.
3. Phase 3 — Complete the first useful path. Implement the workflow's input, review and output interface, with clear controls and explicit empty/error states. Use short SQLite transactions and persist job state before starting work. Give retries stable operation IDs; report incomplete or unknown results instead of silently repeating them.
4. Phase 4 — Permissions and integration failure. Reject unexpected origins and unbounded request bodies even on localhost. Keep credentials outside the database export and redact sensitive text from logs. Treat source documents as untrusted data, prevent them from changing tool permissions, and require review of factual claims before publication. Request integration credentials and permissions only for the enabled feature; show a disconnected state instead of mock results.
5. Phase 5 — Portable handoff. Use a consistent SQLite backup and an attachment manifest. Export portable JSON/CSV, then restore to a new directory without overwriting the original data. Include setup, operating limits, fixture walkthrough and shutdown/restart instructions in the README.
6. Phase 6 — Acceptance scenarios. Reject a concept and make no image request; a provider timeout enters reconciliation rather than charging for an automatic duplicate render. Repeat the workflow after restart and with a denied permission or unavailable dependency; show recoverable failure rather than a success placeholder.

## Paid-product capabilities outside this build
- specialized thumbnail-specific system prompts
- style matching and recreation features
- custom AI training and personalization features
- creator-focused workflow and UI/UX
- thumbnail inspiration system

## Implementation prompt
WORKING SLICE
Draft thumbnail concepts from a video title and audience, review text/layout and optionally request an image from a configured provider. Compare candidate images and export a chosen image at explicit dimensions.

Build this scoped ThumblifyAI-inspired workflow with a documented data model and visible failure states.

Architecture
- Node.js, TypeScript and Express with server-rendered HTML and small browser modules.
- SQLite through better-sqlite3 with migrations, prepared statements and a single background worker.
- A single provider SDK with a configured supported model, schema-validated responses and revisioned prompt templates.

Prerequisites and limits
A supported Node release, a writable local data directory and a separate backup location. Bind to localhost; remote use requires authentication and HTTPS first. Optional AI generation needs a provider key, a usage budget and approval to send the selected material. A selected image-generation provider with an authorized key, render budget and owned/licensed reference imagery.
Outside this release: Guaranteed click-through improvement and universal model access are excluded.

Data model and correctness
video briefs, owned reference images, proposed thumbnail concepts, approved prompts and provider render receipts
Invariant: Paid renders require a reviewed prompt and usage cap; no fabricated performance score or unauthorized real-person impersonation.
Use short SQLite transactions and persist job state before starting work. Give retries stable operation IDs; report incomplete or unknown results instead of silently repeating them.

Security and privacy
Reject unexpected origins and unbounded request bodies even on localhost. Keep credentials outside the database export and redact sensitive text from logs. Treat source documents as untrusted data, prevent them from changing tool permissions, and require review of factual claims before publication.

Recovery and export
Use a consistent SQLite backup and an attachment manifest. Export portable JSON/CSV, then restore to a new directory without overwriting the original data.

Implementation order
1. Phase 1 — Scope and fixtures. Implement this bounded workflow: Draft thumbnail concepts from a video title and audience, review text/layout and optionally request an image from a configured provider. Compare candidate images and export a chosen image at explicit dimensions. Record prerequisites, select representative user-owned fixtures and document the unsupported features: Guaranteed click-through improvement and universal model access are excluded.
2. Phase 2 — Durable model. Model video briefs, owned reference images, proposed thumbnail concepts, approved prompts and provider render receipts Add migrations or a versioned document format, explicit validation, stable IDs and a visible import-error report. Preserve this rule: Paid renders require a reviewed prompt and usage cap; no fabricated performance score or unauthorized real-person impersonation.
3. Phase 3 — Complete the first useful path. Implement the workflow's input, review and output interface, with clear controls and explicit empty/error states. Use short SQLite transactions and persist job state before starting work. Give retries stable operation IDs; report incomplete or unknown results instead of silently repeating them.
4. Phase 4 — Permissions and integration failure. Reject unexpected origins and unbounded request bodies even on localhost. Keep credentials outside the database export and redact sensitive text from logs. Treat source documents as untrusted data, prevent them from changing tool permissions, and require review of factual claims before publication. Request integration credentials and permissions only for the enabled feature; show a disconnected state instead of mock results.
5. Phase 5 — Portable handoff. Use a consistent SQLite backup and an attachment manifest. Export portable JSON/CSV, then restore to a new directory without overwriting the original data. Include setup, operating limits, fixture walkthrough and shutdown/restart instructions in the README.
6. Phase 6 — Acceptance scenarios. Reject a concept and make no image request; a provider timeout enters reconciliation rather than charging for an automatic duplicate render. Repeat the workflow after restart and with a denied permission or unavailable dependency; show recoverable failure rather than a success placeholder.

Acceptance
Reject a concept and make no image request; a provider timeout enters reconciliation rather than charging for an automatic duplicate render.
Use real source data or clearly labeled fixtures. Explain unsupported input and provider failures; do not fabricate analytics, delivery receipts, accuracy claims or security guarantees.

Optional agent guidance
Optional external skill: [copywriting](https://github.com/coreyhaines31/marketingskills/blob/main/skills/copywriting/SKILL.md) — Write landing pages and product copy grounded in the intended audience, product value and a clear next action. Review its instructions and compatibility before use; it does not grant deployment, data-access or publication permission.
Optional external skill: [web-design-guidelines](https://github.com/vercel-labs/agent-skills/blob/main/skills/web-design-guidelines/SKILL.md) — Review web interfaces for accessibility, keyboard focus, forms, navigation and interaction quality. Review its instructions and compatibility before use; it does not grant deployment, data-access or publication permission.
Optional external skill: [sharp-edges](https://github.com/trailofbits/skills/blob/main/plugins/sharp-edges/skills/sharp-edges/SKILL.md) — Review security-sensitive APIs and configuration for dangerous defaults and easy-to-misuse interfaces. Review its instructions and compatibility before use; it does not grant deployment, data-access or publication permission.
Project rule — data model: video briefs, owned reference images, proposed thumbnail concepts, approved prompts and provider render receipts
Project rule — preserve this invariant: Paid renders require a reviewed prompt and usage cap; no fabricated performance score or unauthorized real-person impersonation.
Project rule — acceptance evidence: Reject a concept and make no image request; a provider timeout enters reconciliation rather than charging for an automatic duplicate render.

## Completion evidence
Demonstrate the prompt's acceptance scenarios against the scoped workflow. Include setup from a clean checkout and failure recovery. Check persistence across restart and export/restore only for the state the prompt says to store; for memory-only tools, confirm that temporary content is discarded as specified. Record actual results and remaining limitations. A detailed plan alone does not establish a working replacement.