AI Features

The Hypertask AI assistant lives at your fingertips. Use it to create tasks, enhance descriptions, draft code, summarize threads, or prescribe actions throughout your workflow.

AI Task Writer and Write with AI

Create tasks from prompts

The AI Task Writer instantly turns text into structured tasks. Just ask in natural language, and the assistant fills in title, description, due date, and assignee if you include that context.

# CLI: Create a task with a rich description
ht task create "Design the upcoming Q3 marketing campaign. Include three concepts: social-first carousel posts, short-form video ideas for TikTok and Instagram Reels, and a longer-form blog about Q2 performance. Target HQ team. Due next Friday from HQ."

In the UI, use the AI Task Writer from the New Task flow. The editor is context-aware: it respects the board, team, and board-level AI instructions you’ve configured.

When a ticket already has a title or description, opening Task Writer starts the draft immediately. It uses the title, description, tags, priority, and size as context. Empty tickets still open with the prompt, and your original wording remains available for follow-up edits.

Write with AI in comments and descriptions

Start a comment or task description with /ai to open the AI Task Writer.

• /ai — Generate a draft for the active text field.
• /ai rewrite — Improve or rethink what you typed.
• /ai fix — Correct grammar, sharpen tone, or expand with details.

The assistant drafts directly into the field, ready for your review.

House style AI text reinforcement (HTPR-5587)

AI responses now follow shared house-style rules in both AI Chat and custom instructions to reduce “AI tells” and unnatural phrase patterns. ADHD-shaping feedback applies only to chat-level refinements — modes like Archive and Summarize do not append content to your text.

Write plainly from a human-first perspective without overusing jargon, and keep your tone concise and actionable. Avoid filler phrases and contrived expressions like:

• “Great question!”, “pivotal”, “seamlessly”, “robust”, “leverage”, “I hope this helps”

Smart brevity precedence on style

When writing with AI, the assistant respects your board’s template and custom instructions as the final authority over brevity. You can still request improvements, but board-level constraints always take precedence:

Level	Consequence
Board custom instructions & templates	Entirely reflow the AI-generated output to match the board’s preferences.
Your explicit request	Apply refinements that contradict the default brevity rule.
Default brevity	Apply only if the top two levels do not apply.
Anti-wordiness cap	Reduce filler phrases like “pivotal,” “seamlessly,” and “robust” at or below the style limit.

Example prompts:

# Ask for high-level but concise design guidance
ht task create "High-level guidance for the marketing design system. Layout: layout-doc, typography: typography-doc."

# Explicitly request a long-form, conversational architecture overview
ht task create "Long-form, conversational architecture overview with no brevity limits."

Promoted task-summary action (HTPR-5404)

When AI Chat opens on a task, the assistant prioritizes the Summarize the current task action. You can:

• Type /i Summarize the current task to get a concise summary that respects readability modes.
• See the action promoted near the chat input, making it quick to extract a task summary while keeping context.

This integration lets you generate task summaries without leaving the chat flow.

Write confirmation for wide or destructive actions

The AI assistant requires confirmation before making wide or destructive writes across your boards. This prevents silent task creation or modification.

When you invoke Write with AI with flags that span multiple tasks or boards (such as --mode write-with-ai --all), the assistant pauses and prompts you:

1. The assistant previews the proposed writes (tasks, comments, or label updates).
2. You confirm by running Write from the command palette or by typing Yes in the active confirmation dialog.
3. Only approved writes are sent on the command. Unconfirmed requests are discarded.

This confirmation gate guards against accidental bulk operations and keeps your boards aligned with your intent.

Seed tasks on guest boards

When you open a fresh guest showcase board, click Ctrl+K to open the board picker, then follow one of the onboarding prompts:

• Press Ctrl+K: every action is one keystroke away — Tap the “New task” action, then the AI Task Writer to create a task with natural-language instructions.
• Ask the AI to build this board for you — Paste a description like “a board for lighthearted weekly recaps with a backlog, doing, and done column” and let the AI draft the structure.
• Run tasks/tools — Create a task that references will-be-created tasks or external actions via the Task and Tool options from within the AI Task Writer.

These prompts start a simulated typing session that ends with AI Task Writer empty and ready for your own elaboration.

Cap to smooth and anti-wordiness traits

The AI Task Writer behavior now includes two documented style traits:

Trait	What it does	You will see
Anti-wordiness cap	Prevents the AI from repeatedly using high-fill phrases like “leverage,” “pivotal,” “seamlessly,” “robust,” and “I hope this helps.” Only one pass of filtering is applied; further iterations may reintroduce content you type.	Misspelled intentional repetitions (e.g., “revision re-vision”) are tolerated. Plain, concise language with structures like “I intend to X and Y instead of A and B.”

Note: This trait is tied to AI-generated messages that inherit default brevity. If board custom instructions or your explicit request take precedence, anti-wordiness will not be applied, consistent with the brevity precedence table.

Example:

ht task create "Our design system typography should leverage scale appropriately during page edits so that we never see markdown-like scaling."
# After the AI Task Writer applies anti-wordiness:
# "Typography scales in our design system align page-edit behavior with typographic scale, so markdown-like scaling never occurs."

Agent hover cards open the agent page

Hover over an agent to quickly access its detail page and configuration. When you hover over an agent card, a hover card appears with a link that opens the agent page for review, status, and settings.

The hover card only applies to agents; user-facing hover cards remain unchanged.

Skills invoked from AI Task Writer

Use slash commands to invoke registered skills:

• /skill — Open the skills menu to tap a skill by name or first letter.
• /skill name — Run a specific skill if you know its name.

Known skills include:

• Defer — Suggest a deferral with a new due date and optional sub-task.
• Duplicate — Mirror the current task into a draft that you can rename and refine.
• Subtask — Generate a new subtask rooted from the current one.

Types skip over canvas or rich content blocks instead of treating them as text to rewrite. This avoids unintended edits to estimated time blocks or other rich media.

AI agents and style enforcement (HTPR-5648)

QA Agent verification

Hypertask’s QA Agent enforces style-guide compliance for AI-generated output. The agent polices tasks and comments to ensure they follow shared house-style rules:

• Verify that output avoids AI-tells and filler phrases
• Flag non-conforming content for review
• Apply probation labels to encourage compliance on subsequent drafts
• Prevent automated merges of style-violating content

Agent workflows for testing with Midscene (HTPR-5715)

The QA Agent now runs AI-driven end-to-end tests using Midscene with a dedicated Playwright scaffold. These tests combine natural language instructions with computer vision to validate critical app flows:

• Tests verify user actions, UI elements, and real-time interactions from an agent’s perspective
• Hard resource guards limit RAM, CPU, and runtime to keep test execution stable: 2G RAM, 1.5 CPU, and 10-minute timeout per run
• A single-flight lock ensures only one test runs at a time, preventing interference between concurrent attempts
• Cleanup on every exit path and post-run checks guarantee no leftover test artifacts

This setup lets QA Agents autonomously detect regressions and provide concrete test evidence alongside their report summaries.

Dev Agent documentation precondition

Dev agents that create, modify, or review content must now read docs/DESIGN.md before acting as the authoritative source of truth. This ensures agents:

• Understand the project’s design principles and patterns
• Apply consistent terminology and approaches
• Reference validated guidance instead of working from memory or assumptions

The QA Agent’s style verification complements this pre-condition by catching misalignment at runtime.

Combined behavior

Together, QA Agent verification and the dev Agent documentation precondition create a reinforcing loop:

1. Dev Agents consult docs/DESIGN.md before creating or modifying content
2. Content generated by Dev Agents is subject to QA Agent style verification
3. QA Agents flag non-compliant output, suggest corrections, and inhibit automerge until compliance is achieved

This reduces reliance on manual review while preserving the intent captured in design documentation.

My Tasks context

The AI understands your My Tasks workload and can answer context-specific questions about it:

• “What’s on my My Tasks board?” — Generates a concise list of assigned tasks with priorities and due dates.
• “Any overdue or blocked work?” — Flags urgent or late items for follow-up.
• “Show me tasks due this week” — Filters and presents upcoming commitments.

For agents using MCP, the /my-tasks tool exposes your My Tasks as a structured queryable dataset, including counts and access control boundaries. This ensures agents only see the tasks they’re legitimately allowed to work with.

AI chat queue for automatic sending after the current turn (HTPR-5695)

Follow-up messages are now queued and automatically sent after each round is complete. While a turn is streaming, the Send action is replaced by Stop, and queued messages appear in a queue strip with badges showing pending items. Once the current turn finishes, queued messages are sent in FIFO order.

View and manage the AI chat queue

1. While typing, see the current turn’s messages, plus any queued follow-up messages displayed above or below the composer.
2. Use Stop if you want to cancel the streaming turn without sending queued messages.
3. After the turn completes, queued messages are automatically posted in order. You can edit or delete queued items from the queue before they send.
4. Each queued item also shows an individual Send action if you want to override the automatic behavior and send it immediately.

Tip

The queue is per-turn and auto-confirms for follow-ups on completed turns, reducing confirmations while preserving control over each message.

AI chat restoration (HTPR-5676)

AI Chat can now reach inne agents again following a bridge implementation update. The inne agents operate on inne’s own Cloudflare account, and cross-account workers cannot be wired together directly. The chat bridge was temporarily down, returning “could not be reached” errors for every inne agent.

After moving inne agents to inne-owned Cloudflare infrastructure, the bridge is restored. You can again reference inne agents from AI Chat in the same way as other board agents.

AI status updates

An AI agent can update task metadata in a visible way — priority, assignees, labels, due dates, or related tasks. When an agent makes a structural or status change, an AI status update appears on the task card so you see the change instantly without opening every detail.

• The agent runs Write with AI with the --mode write-with-ai flag, posting updates as real comments.
• You see the updates on the task card, in the task detail panel, and inside notification contexts.
• This enables agents to keep tasks current (e.g., promote high-priority work, clear blockers) while keeping manual edits to structural work visible and auditable.

Leverage it from the command palette: Ctrl+K — search for a task and use Agent write-with-AI to trigger the update.

Automated approvals from minors-only reviews (HTPR-5531)

When AI agents perform minors-only reviews on pull requests, they now automatically post an APPROVE comment to green PRs. This approval comment enables improving automerge behavior: green PRs no longer stay parked forever due to the missing green-PR-approved-by-minor comment. The system also dedupes park comments to avoid duplicate approval entries. As a result, minors-only reviews contribute directly to board automation flows, making approvals more reliable and predictable.

Agents page card hover Opens detail

Hovering over an agent card now opens the agent detail page, making it quick to review agent status, configuration, and history from the main agents page.

Time, profile, and report parity in AI Chat

AI Chat now supports more of the same actions available in the web app, giving you comprehensive task and workspace control from chat. You can:

• View filtered time reports — Ask AI Chat to show time entries for specific projects, tasks, or date ranges.
• Update shared user profiles — Update profile fields like name, role, or custom attributes with correct flag handling.
• Read complete saved reports — Pull full saved report data including filters, columns, and time ranges.
• Delete reports with confirmation — Remove saved reports and confirm the action before it happens.

Use the same context and permissions that apply elsewhere in Hypertask. Check access, see the full report content, and clean up saved reports directly from your conversation.

ChatProvider and hooks

When using AI chat features on server-rendered routes like task detail pages, ensure the UI is rendered within the ChatProvider boundary. This prevents hydration errors and ensures proper context control for hooks like useAiChatContext.

This is automatically handled for standard flows — use the AI Task Writer, context menus, and panel-based chat — and is handled by the fallback behavior on detail pages that load the provider safely.

AI access and entitlement

AI access is controlled through an allowance model based on team plan type. The gateway routes requests to dedicated credentials for internal teams, paid and legacy plans, and the allowance key for Free and BYOK subscriptions.

Channel	Credentials	Default model	Notes
Internal team members and custom-deal teams	Dedicated managed-key registry (14 teams, 11 credentials)	Eligible users	Isolated from customer BYOK; production-ready and verified
Paid, legacy, and comped teams	Dedicated routed keys by entitlement	Luna Standard (5.6)	Server-gated; consistent with paid product tiers
Free and BYOK subscribers	Shared Free/BYOK included-allowance key	Mini (5.4) for free/keyless BYOK; Luna Standard (5.6) for paid/buy-your-own-key BYOK	Usage limited to the models in your allocation
Android app	PKCE OAuth tokens via native OAuth 2.1 flow	Eligible users	Native Android sign-in integrates with existing web PKCE OAuth; provides access to AI chat streaming and AI-assisted Task Writer

Free users are provisioned 100 credit-machines (web/desktop) and 30 credit-machines (mobile) at signup. Additional credits are available through verified use or contact.

Plan-aware default-model repair

The gateway now maps eligible users to appropriate default models based on team entitlement:

• Paid, legacy, and comped teams default to Luna Standard (5.6), bypassing premium model selection during task creation.
• Free and keyless BYOK teams default to Mini (5.4) with a corresponding downgrade to Mini-only allocation; BYOK teams with an owned key are eligible for Luna Standard.
• Eligible BYOK squads reclaim Luna Standard (5.6) instead of staying on Mini, restoring Luna Standard as the design default for BYOK teams with an owned key.

Recommended: if you previously saw Mini as the visible or persisted default for paid or BYOK squads, verify that your plan shows Luna Standard as the eligible model and that your usage allocation permits Luna Standard.

Plan access guidance

When a model requires a paid plan or an owned API key, the system shows an appropriate upgrade/BYOK message to help teams align their models with the right plan or credential setup. This message is now surfaced correctly as guidance, while the handled-error reporter correctly does not create product bug tickets for this expected outcome.

Premium model selection after board switch

When you create a task via quick-create and switch to a premium board, the model picker now uses the destination board’s team plan instead of the originating board’s entitlement. This restores full model availability for premium teams, including Luna, Terra, Sol, GPT-5.5, Anthropic, Google, and xAI, without forcing Mini-only selections.

Unified team entitlement checking for boardless AI Chat (HTPR-5590)

When starting an AI chat without an active board, Hypertask now broadly checks the calling user’s team memberships to determine the correct plan-level entitlement. This resolves cases where newer Free teams hid older paid memberships, causing the system to incorrectly treat the user as Free and gate the conversation to Mini-only.

• Behavioral change: Previously, the plan gate evaluated only the most recently accessed or implied team context. Now, the chat backend ranks owned and accepted-member teams by current entitlement, examining all teams where the user is a member.
• Result: AI chat correctly identifies paid or BYOK entitlements regardless of team ordering or last interaction, allowing premium model selection and access.
• Scope: The fix applies across CLI, MCP, API, and AI Chat surfaces, ensuring consistent entitlement behavior across all interaction points.
• Regression tests verify paid membership detection path and tool usage reflect the upgraded entitlement logic.

Invalid client payloads and owner derivations (HTPR-5657)

AI mentions can include additional field hints (aliases) for the owner; these are now derived server-side instead of being silently swallowed. Invalid payloads surface meaningful errors across CLI, MCP, and AI Chat:

• Server derives owner from available hints/account data.
• Invalid payloads now get a clear error (such as 400) and a descriptive message instead of being logged without response.
• Client side toasts every error in addition to 403 scenarios, so users always see when a status mismatched or malformed hint was supplied.

Example: If a message contains an alias like “@alias:owner” that doesn’t map to a known user, the client receives an explicit error instead of a silent failure or generic 403.

---

Tip

You can override per-board and team AI settings from board-level custom instructions, the CLI, or MCP tool configuration. Shared prompts and agents benefit from unified secure prompts that live outside your team’s files.