Agentic tools primer — applying the distinction to UGC work

02-chat-vs-agentic.md teaches the distinction and gives you the two-question test. This document does not repeat it. Read that first; this one starts where it stops, and goes at the question you actually have: which of the things I do in a week is this for?

1What you need, and what you don't

Every job in this document can be done on a free Claude account. All of them work in a chat, one item at a time, and the judgement — which is the part clients pay for — is identical either way.

What a paid plan buys is doing these jobs unattended and at volume. That is a real difference, and it is worth money once you have the volume. It is not a different capability, and it is not required to finish this module.

Verified 2026-08-14 against support.claude.com: Cowork is available on Pro, Max, Team and Enterprise, across desktop, web and mobile, with Enterprise depending on admin settings. Plugins are "available to all paid plans."

If you are on free, you are not missing the module. You are doing the same work with your hands on it, which is how you should learn it regardless of plan — see the closing section, which says so about paid users too. Come back to the upgrade question when the volume starts to hurt, and not before.

Explain-on-first-use: Cowork is Anthropic's name for the agentic surface — the place where, per its own documentation, you "describe an outcome, step away, and come back to finished work." It is a different mode, not a different model.

2What it actually does, in terms that matter to this work

Verified behaviours worth knowing before you plan a week around it:

Its own documentation states plainly that "Cowork has unique risks due to its agentic nature and internet access." That is the vendor saying it, not us.

3The five UGC jobs worth handing off

Ordered by how reliably they pay off. Each maps to something already built in this module, and each says what it looks like on a free plan — because on free these are still jobs you do, just with your hands on them.

1. Running the guardrail check across a batch — the best one

You have twelve scripts for one client. claim-guard needs to run on every one, against the same BV-01 intake, checking the same off-limits list and the same experience file.

This is the ideal profile and it is worth understanding why, because the reason generalises: the standard is fixed and external, the work is identical per item, the output is checkable at a glance, and doing it by hand is tedious enough that you will skip it under deadline. The tasks you skip when tired are the best hand-off candidates, because a tool does not get tired at 11pm on a Thursday.

On free: paste the intake, the experience file and one script into a chat and run the check. Roughly ninety seconds per script, and the check is exactly as good — you are the part moving scripts in and out.

2. Generating a hook set per product across a catalogue

script-from-intake against one intake and fifteen products. Same brand voice, same constraints, different subject each time.

Run it on three before you run it on fifteen. §2's first habit exists because almost every expensive hand-off failure would have been caught on a sample that took two minutes.

On free: one product per chat, with the intake pasted once at the top and reused down the conversation. Slower per product, identical output.

3. Footage triage

"Go through this folder, transcribe everything, and tell me which clips mention the pan." Mechanical, checkable, and the thing that eats an hour of a shoot day.

On free: upload clips individually. This is the one job on the list where free is genuinely worse rather than just slower — the value was in not touching the folder, and on free you are still touching the folder. If you shoot a lot, this is the line item that justifies upgrading before any of the others.

4. Turning a week of notes into a filled experience file

You have voice memos, a notes app and a camera roll from four weeks of using a product. The experience file that script-from-intake requires does not exist yet, and that is why the skill keeps refusing.

This is a legitimate hand-off with one hard rule: it may only extract, never infer. If your notes do not say how long something took, the output must not contain a duration. That is exactly the failure demo run 02 recorded. Say so in the instruction, explicitly, every time.

On free: this one works properly. Paste the notes into a chat and ask for the extraction. No folder access needed — the notes are already in your hands, and this is the hand-off most likely to unblock a stalled script-from-intake run.

5. Competitor and reference sweeps

"Pull the last thirty posts from these three accounts and tell me what the openings have in common."

Treat the output as a hypothesis, not a finding. Per §2: you have almost no way to check whether the pattern is real or whether you were handed a confident-sounding summary of nothing. Go and read three actual posts against it before you build a pitch on it.

On free: gather the posts yourself and paste them in. More work to collect, same quality of read — and the verification step was always yours anyway.

4The three you should not hand off, and why

• Which take ships. No file contains the answer. This is the decision the client is paying for.
• Anything going out under a client's name that you haven't read. The temptation at volume is exactly proportional to the risk. This is how a brand ends up saying something it is contractually barred from saying.
• The final claim-guard read on a BLOCKED finding. Batch-checking is a hand-off. Deciding what to do about a barred claim is not.

5Writing an outcome instead of a task

The skill people lack is not prompting. It is specifying a finished state.

Three properties the right-hand column has and the left-hand column doesn't: where the output goes, what it's called, and what not to touch. That last one is the one beginners miss, and it is the one that prevents the expensive mistake.

6The four habits

Adapted from §2's three, with one added from the demo runs:

1. Sample first. Three items, not thirty.
2. Ask it to state its assumptions before it starts. One line — tell me how you're interpreting this before you do it. Catches the step-two error while it is still free.
3. Check the first and last item by hand. Not a random sample. Drift shows at the ends.
4. Never let a hand-off write to the file it is checking against. A batch job that can edit intake.md or experience.md can quietly resolve its own findings by rewriting the standard. Keep guardrail files read-only in any instruction you give it. This is the automated version of the failure skills/README.md warns about — fill the gap, never delete the constraint — and it is worse when a machine does it, because nobody watched.

7What this costs, honestly

Cowork "consumes more usage than standard chat." Documented by Anthropic, not inferred by us.

So the arithmetic is not free automation. It is: does the work this replaces cost more than the usage it consumes? For twelve guardrail checks it plainly does. For rewriting one hook it plainly does not — that is a chat, and it always was.

The honest beginner advice from §2 stands and gets stronger here: stay in the chat longer than feels efficient. You cannot supervise a process you have never performed. Do the twelve checks by hand once, badly and slowly, then hand them off knowing what a wrong one looks like.

This applies to paid users too, and it is why the free path is not a lesser version of this module. A reader on free does the work by hand because that is what is in front of them; a reader on Pro should do it by hand first anyway, and mostly won't. The free path teaches the thing the paid path lets you skip.

When to actually upgrade: when you can name the specific job costing you an hour a week. Not before. "Being on Pro" is not a workflow, and an upgrade bought ahead of the volume buys a subscription and nothing else.