[Learn from Others]
Vague initial description
Impact: Causes 5–10 correction rounds. Every vague detail becomes a coin flip.
Skipping the ICP Skills prompt
Impact: Without ICP Skills, agents invent token interfaces. Expect broken ledger calls and wrong decimal handling.
Not thinking through architecture upfront
Impact: Retrofitting architecture mid-project is expensive — always decide your data model, canister structure, and feature scope before writing the first prompt.
Mixing unrelated changes in one prompt
Impact: Agents partially implement some changes and miss others. More rounds, more credits wasted.
Never running an audit
Impact: Real users find the issues instead. Security gaps, broken flows, and data loss in production.
Ignoring a blocked message
Impact: Rewriting with technical jargon compounds the block. Simple, intent-focused language almost always resolves it immediately.
> Assuming file storage has hard limits
Impact: Planning your app around limits that don't exist — or being caught off guard if limits are introduced in the future.
[File Storage — Current Status (2025)]
- →As of 2025, Caffeine has no hard limits on individual file size or total storage per project.
- →Files are stored on Caffeine's managed infrastructure (not the decentralized network directly) and identified by cryptographic hash for automatic deduplication.
- →If storage limits are introduced in the future, Caffeine will communicate this in advance.
- →Best practice: Still use efficient image formats (WebP, compressed JPEG) for performance \u2014 smaller files load faster for your users regardless of storage limits.
> Ignoring GitHub Import and ZIP Upload
Impact: Manual file uploads instead of direct code import — slower, more error-prone, more credits wasted.
[Code Import — New Options Since April 2026]
- →GitHub Import: Import code from GitHub repositories directly into Caffeine \u2014 including version history and branch selection.
- →ZIP Upload: Upload ZIP files containing complete projects (paid plans only).
- →Both save time and reduce errors from manually copying files.
> Writing prompts without checking character limits
Impact: Prompt gets truncated or rejected — important details lost.
[Character Limit — Since June 12, 2026]
- →Messages have a character limit \u2014 exceeding it will reject the prompt.
- →Attach long prompts as .txt files instead of pasting into the text field.
- →Break prompts into smaller, specific steps — better for the AI and for your budget.
> Not knowing project deletion rules
Impact: Unintentional data loss or unexpected restrictions when deleting.
[Deletion Rules — Since May 25/26, 2026]
- →Deleted projects cannot be restored \u2014 all data, builds, and settings are permanently lost.
- →Free Plan: Maximum 3 projects at once — delete old ones to create new ones.
- →Paid Plans: Higher limits, but still be careful when deleting.
> Not understanding credit hold behavior
Impact: Confusion about why credits aren't immediately available after a cancelled build.
[Credit Hold — Since April 16, 2026]
- →Credits are reserved (held) when a build starts \u2014 not immediately deducted.
- →On successful build: hold becomes actual deduction.
- →On cancellation or failure: hold is released, but not instantly — can take minutes.
- →Balance display since May 22, 2026: shows available + reserved credits separately.
[Blocked Messages — What You Need to Know]
- →Blocked messages never cost credits. The filter evaluates text, not intent. Rephrase using simpler, more direct language and describe what the feature does for the user rather than the technical mechanism.
- →A blocked message is automatically removed from the AI's conversation history \u2014 it does not contaminate future prompts. You can rephrase and try again immediately.
- →Accidentally triggering the filter has no account consequences. Account-level actions are only taken for clear, repeated, deliberate policy violations.
- →If you're unsure why something was blocked, try breaking the request into smaller, more specific steps \u2014 this almost always resolves it.
// Pre-Build Mistake-Prevention Checklist
- Written a full app description (goal, audience, features, data model)?
- Added the ICP Skills one-liner?
- Specified canister count and architecture?
- Listed authentication and storage requirements?
- Planned which features are Phase 1 vs later?
- Committed to auditing after the first build?
- If a message is blocked — rephrase, don't panic. No credits lost.
- If using file uploads — use efficient formats (WebP, compressed JPEG) for best performance.
- Consider GitHub Import or ZIP Upload for existing projects?
- Attach long prompts as .txt files instead of pasting into the text field?
- Understood project deletion rules (no recovery possible)?
- Understood credit hold behavior (credits reserved at build start)?