Choose a test batch
Give it five photos, the product codes, and a destination. Ask to see the proposed matches first.
Every Friday, new product photos land in the same folder. Ask the agent to match filenames to your product codes, copy matches into the right folders, and leave anything it can’t identify alone. Check the first batch before scheduling another.
Windows 10 and 11 now. Mac release planned.
Source
Supply context
Run
Use available tools
Activity
Keep work visible
Review
Check the result
A sorted batch of photos, a supplier page open on the right item, or a project folder ready for a new client: pick something you can check at a glance.
Create the standard folders and copy in the kickoff material. Leave the originals alone.
Give it three part numbers. Ask it to open the matching listings and flag missing stock or lead times.
Read order numbers from receipt images, match them to purchase orders, and set uncertain matches aside for review.
Record a short walkthrough, then use your notes to draft and test a checklist.
Why these make good first workflows
For product photos, name the incoming folder, supply the product codes, and specify “copy, don’t move.” Ask it to list the files it couldn’t match.
Example task brief
“Match these five photos to the product codes. Create any missing product folders and copy each match there. Leave the originals alone and show me the filenames you couldn’t match.”
Input
What should it work from?
Output
What should it produce?
Limits
What should stay untouched?
Check
What does “done” look like?
Success criteria are optional. Add them when a measurable finish condition helps. They help define “done”; they do not replace reviewing the source.
Give it five photos, the product codes, and a destination. Ask to see the proposed matches first.
Think-Act can use the capabilities exposed to the run, subject to permissions and policy. A missing connector should be reported, not simulated.
Compare the output with the original material and inspect relevant activity. Correct the brief or method when something important is missed.
Preserve the reviewed method as a local Skill, including prerequisites, required tools, boundaries, and the expected result.
Review the tool activity and check the destination folders. A Skill saves the instructions; a scheduled task tells Think-Act when to run them.
Extract supplier, due date, and total. Ask the agent to flag unreadable fields rather than fill them in.
Ask for specific names, dates, amounts, or line items. Flag unclear source content so it remains visible during review.
Organize the extracted content into a table, CSV text, summary, or checklist with a structure you can inspect.
Use the reviewed information to prepare questions, a reply, or next steps. Sending or submitting it requires an appropriate tool and permission.
OCR and AI interpretation can miss detail or context. Check important names, dates, totals, clauses, amounts, and obligations against the source.
Source
Original document
Extract
Requested fields
Prepare
Useful structure
Review
Verify source
Record a short walkthrough. Use your notes to write a checklist and keep the clip as a reference for exceptions.
Capture the relevant steps while you perform them.
Prepare an SOP or checklist from the recording.
Correct omissions and add prerequisites or warnings.
Keep the tested method as local workflow guidance.
Check that the next task still meets its prerequisites.
Turn a walkthrough into an SOP, checklist, bug report, or tutorial draft. Review it for missing decisions, setup steps, and exceptions before sharing or reusing it.
Skills are reusable local instructions or playbooks. Capture the proven process, its prerequisites, and the expected result.
A Skill does not supply a missing capability. Any required tool must still be available, trusted, enabled, and permitted for the run.
Test first. Save second. Reuse after the behavior is understood.
The agent already has file and browser tools. Use an MCP connector if the task needs a compatible external service.
An enabled, healthy tool can expose specific capabilities to the run. Think-Act may use them according to configured permissions and policy.
For example
Read a permitted file through an available filesystem capability.
File changes, message sending, and external writes need appropriate tool access. Configured policy may require approval before the action continues.
For example
Preparing a reply is different from sending it through an enabled messaging capability.
Think-Act should identify the missing tool and explain what the task needs. It can still prepare useful content, but it should not claim the external action happened.
For example
Draft an issue from a bug report while making clear that it has not been submitted.
Categories can include filesystem, browser, database, GitHub, app APIs, and local MCP. These are capability categories, not a promise that every connector is installed or available.
Third-party local MCP packages may execute code on your device. PixLab does not control arbitrary packages. Review their sources before installation and enable only software you trust.
For a rename job, request a table of old and new names first. Set tool permissions to match what you want the agent to do.
Action policy
Read what will change, which file or system is involved, and whether it matches the task's boundaries.
Approve a gated action only when the proposal is understood. Deny it if the target, scope, or result is wrong.
Review supported activity and the actual output. A completed run still needs to be checked against the request.
Use available recovery mechanisms when the workflow supports them. Do not assume an external action can be undone.
Preserve the checks, prerequisites, and approval expectations that made the method reliable.
Practical limits, approvals, and how to reuse instructions safely.
Not every app. Think-Act can use supported desktop and browser controls; locked screens, missing tools, and permissions can prevent an action.
Yes. The security policy decides when to ask. You can also request a proposed change first and review it before asking the agent to apply it.
Start with a task you can check, like comparing two files. Leave payments, deletions, and publishing out of a first run.
The agent can't use a missing connector. Add a trusted one if the job needs it, or ask for a local draft instead.
Save it when you have tried the steps, checked the result, and expect to repeat the job with different inputs.
No. Use Writing Assist for a rewrite or OCR for a scan. Choose the agent when the result takes several steps.
Yes. The free allowance lets you try a task with your own files. See current plan options.
Start with five files. Preview where they belong, ask the agent to copy them, and open the destination to see whether the routine worked.