Smart Capture

Smart Capture is the AI mode of Quick Add. When you toggle the AI checkbox in the Quick Add modal header, your input gets parsed by an AI model instead of (or in addition to) Pensum’s local natural-language parser. The model has context the local parser doesn’t: your full list of projects, your recently-used tags, the file you currently have open, and today’s date.

Smart Capture is the smallest, fastest AI feature in Pensum — it runs in under a second on a small model. It’s designed to disappear into your capture flow rather than feel like an “AI feature.”

What it does that local parsing doesn’t

The local parser is good at the patterns it knows: p1, tomorrow, #tag, @projectname. Smart Capture handles things the local parser can’t:

Input	Local parser	Smart Capture
Send pricing to Acme by next Fri	”Send pricing to Acme by next Fri” (no date)	Parses “next Fri” → next Friday’s date
Follow up on the comp plan (you’re in Projects/Acme.md)	No project assigned	Assigns #project/acme from context
Talk to Jamie about timeline @al...	No project assignment	Auto-completes @alpha from your project list
Buy birthday gift for mom	Just description	Suggests recent #personal tag if you’ve used it
Wrap deck before EOD	”Wrap deck before EOD”	Parses “EOD” → today’s date
Set up the thing Sam mentioned	Generic	Cleans up wording, adds context if the file you’re in has clues

The result still appears in the Quick Add preview, just with an AI badge so you know it’s enhanced.

How to use it

1. Open Quick Add (Cmd+Shift+A)
2. Toggle the AI checkbox at the top of the modal (or set it as default — see below)
3. Type your task
4. Watch the preview update with AI-parsed values (priority / date / project / tags)
5. Enter to save

If you don’t toggle AI, you get the local parser, which is also great (and instant). AI is an opt-in for cases where context-aware parsing helps.

Making AI the default

Quick Add remembers the AI toggle state per session. To make AI on by default at plugin load, you currently need to toggle it manually each session — a persistent default setting is a small future addition.

What context the AI sees

Smart Capture sends the following with each call:

• Your input text — exactly what you typed
• Today’s date — so relative phrases (tomorrow, next Fri, EOD) resolve correctly
• Your project list — names only, not file contents. From every file with type: project in frontmatter.
• Your recent tags — pulled from the last 100 tasks you’ve interacted with. Names only.
• Current file path — the file you have open when you triggered Quick Add. Helps the AI weight nearby projects.

What it does not send:

• The contents of your current file
• The contents of any other files
• Your task history beyond tag names
• Anything about meetings, transcripts, or wiki entries

What Smart Capture does NOT do

Smart Capture is intentionally narrow:

• It doesn’t guess priorities. If you don’t say p1 or !!!, the task has no priority. The AI won’t invent urgency.
• It doesn’t invent dates. If you don’t mention a date (even an indirect one like “Friday”), the task is undated.
• It doesn’t guess at projects when there’s no signal. If your input doesn’t reference a project (in the description, via @name, or via context from the current file), no project is assigned. Better to leave it for Smart Triage later than guess wrong.
• It doesn’t add tags you didn’t imply. Suggestions come from your own recent usage, and only when the description clearly relates.
• It doesn’t change the meaning of what you typed. Cleanup is for clarity, not interpretation.

Smart Capture is a parser, not an author. The resulting task is marked as user-authored in provenance — it’s still your task, just parsed better. See provenance.

What happens when AI fails

If the call to the AI provider fails for any reason (network down, rate limit hit, key invalid, model returning malformed JSON), Quick Add silently falls back to the local parser. You don’t get an error notice — capture should never break, even if AI is unavailable. The task lands using local-parser results, no AI badge.

This is the only AI feature in Pensum with silent fallback. Every other AI feature surfaces a clear notice with action links when something fails — they all have a known good “manual” alternative that’s worth surfacing. For Smart Capture, the local parser is a true fallback that produces a working task; making you click through an error notice for every transient hiccup would punish you for using the AI mode in the first place.

The Activity view’s AI tab still logs the failure, so you can see what happened. See Activity.

Choosing a model

Smart Capture defaults to a lightweight model (Claude Haiku 4.5 in the catalog) because:

• Latency matters. Capture flow breaks if there’s a noticeable pause between Enter and “task saved.”
• The task is simple. Parsing one line into description + priority + date + project + tags doesn’t need a flagship model. Mini-class models do it perfectly.
• Cost matters. You’re going to capture dozens of tasks per day. Pennies add up.

Change which model under Settings → AI → Smart Capture. The dropdown shows what’s available given your plan and configured providers; Custom... lets you type any model ID (BYO power users with non-catalog models).

For Pro All-in-One users, Smart Capture runs through Pensum’s managed proxy — no key required. For Pro BYO, you need a key for the selected model’s provider (Anthropic for Claude, OpenAI for GPT, etc.).

Cost

On a lightweight model like Claude Haiku 4.5, Smart Capture costs roughly $0.0001 per call — about a penny per 100 captures. For a heavy capture day (50 captures), that’s half a cent in provider fees on BYO. Pro All-in-One has this included with no per-call charge.

When to leave AI off

• You captured the same way for years without AI and you have the local syntax memorised
• You’re on a noisy network and the latency is annoying
• You’re capturing in a private/sensitive context and prefer not to send the description out at all
• You’re rate-limited on your AI provider for the day

The local parser is always available. Toggle and forget.