All posts

A real week of marketing run through Cowork, no invented wins

One week, one solo marketer, one Cowork setup. Here's what actually held up and what I had to unpick by hand.

Most "I ran my whole week through AI" posts are written by people who ran about forty minutes through AI and extrapolated. I want to do the opposite. This is a week of real marketing work on my own business, run through Claude Cowork, with the parts that worked, the parts that wasted my time, and the exact file layout I ended up with. I'm not giving you a productivity multiplier because I don't have one. I have a week, and some opinions about what the week taught me.

The usual take is that agent co-work fails because the model isn't good enough yet. That's not what I hit. The model was fine. What broke was context handoff: every session started from zero unless I built it a place to look. The difference between a Cowork setup that saves you a day and one that costs you a day is almost entirely whether you did twenty minutes of file plumbing first. That's the boring answer, and it's the true one.

What "a week through Cowork" actually means

Cowork gives an agent a working directory and lets it run for a while without you babysitting each step. That is the whole pitch, and it's a real one. But the unit of work stops being "a prompt" and starts being "a task with a folder."

So the week looked like this:

Day	Task	Ran unattended?	What I had to fix
Mon	Quarterly update log draft	Partly	Voice drifted to press-release by para 3
Tue	Competitor page teardown (6 pages)	Yes	Nothing. Best result of the week.
Wed	Blog post outline + draft	No	Abandoned. Wrote it myself, faster.
Thu	Analytics pull → summary	Yes	Had to force it to flag gaps, see below
Fri	Repurposing Tue's teardown into posts	Partly	It reused its own summary, not the source

Two clean wins, one partial, one rescue, one write-off. That's a normal week. Anyone telling you it's five for five is selling something.

The pattern in the failures is the same each time. When the task had a fixed input I could point at (six URLs, one CSV), it ran clean. When the task needed my judgment loaded in (voice, what's worth saying this quarter), it produced something plausible and wrong, and plausible-and-wrong takes longer to fix than blank-page.

The folder layout that fixed Monday

Monday's failure was avoidable. The agent had no copy of my voice other than what I typed into the task, and what I typed into the task was a description of my voice rather than examples of it. Descriptions of voice don't work. Examples do.

Here's the layout I moved to on Tuesday and kept:

project-root/
├── CONTEXT.md           # who we are, what we sell, who we don't sell to
├── VOICE.md             # 3 real excerpts + a hard never-list
├── FACTS.md             # every real number, with source + date
├── inputs/
│   ├── competitors.txt  # raw URLs, one per line
│   └── analytics.csv    # exported, untouched
└── output/              # agent writes only here

Four rules that make it work:

• FACTS.md is the only place numbers may come from. If it's not in there, the agent is not allowed to produce a figure. This is the single highest-value file in the setup.
• VOICE.md holds excerpts, not adjectives. Three paragraphs of my actual writing beat any amount of "punchy, direct, no fluff."
• inputs/ is read-only by convention. The agent never edits source data, so I can always re-run from clean.
• output/ is the only writable folder. Makes review a diff, not an archaeology dig.

The never-invent rule, written as an actual instruction

Thursday was the analytics summary, and the first pass did the thing every model does: it smoothed over a gap in the data with a confident sentence. Not a fabricated number exactly, more a fabricated trend. "Traffic grew steadily through the period" when two weeks of the export were empty.

This is the paragraph I now paste into any task that touches data. It's not clever. It works because it gives the agent a legal output for "I don't know," which is the thing most prompts forget to do.

HARD RULE ON NUMBERS AND CLAIMS

Every figure, percentage, date, and comparative claim in your output
must trace to a line in FACTS.md or a cell in the provided data file.

If a number you need is not available:
  - Write [MISSING: <exactly what you need>] inline in the draft.
  - Do NOT estimate, round from memory, or infer from context.
  - Do NOT describe a trend you cannot see in the data.

If the data has gaps (empty rows, partial periods, missing segments):
  - Say so explicitly in a "Data gaps" section at the top of the output.
  - Do not smooth over the gap with qualitative language.

At the end of the output, list every [MISSING: ...] marker you used
under a heading "Needs a human". If there are none, write "None".

A draft with ten honest gaps is correct. A draft with zero gaps and
one invented trend is not usable and I will have to throw it away.

The last line does more work than the rest combined. Telling the model the cost of the failure mode changes its behavior more reliably than telling it the rule.

The one task that ran clean, and why

Tuesday. Six competitor pages, teardown against a fixed rubric. It ran unattended for a stretch and I changed nothing in the output.

It worked because every ingredient was pinned. The input was six URLs in a file, not "find my competitors." The rubric was written out, not implied. The output format was specified down to the headings. There was no room for it to be creative, and creativity was not the job.

TASK: Competitor page teardown

INPUT: every URL in inputs/competitors.txt

For each URL, fetch the page and produce this exact structure:

## <domain>
**Promise in the first screen:** <quote it verbatim, max 25 words>
**Who it's for:** <their words, not yours. If unstated, write "Unstated">
**Proof offered:** <list. Types only: logos / numbers / testimonial /
  demo / none. If numbers, quote them and note if sourced or bare>
**Price visible without contact:** Yes / No / Partial
**Weakest link:** <one sentence. The thing a buyer would doubt>
**What they'd have to change to beat us:** <one sentence>

RULES
- Quote, don't paraphrase, in the "Promise" and "Who it's for" fields.
- If a page won't load, write "## <domain> ,  UNREACHABLE" and move on.
  Do not substitute a cached or remembered version of the page.
- No summary section at the end. I'll do the synthesis.

OUTPUT: output/teardown-<today's date>.md

Note the last rule. I banned the summary on purpose. Friday's partial failure was the agent repurposing its own summary instead of going back to the source, and a summary that exists is a summary that gets reused. Deleting the tempting shortcut from the output is easier than instructing around it later.

Why Wednesday was a write-off

Blog post. I gave it a topic, my VOICE.md, and a rough angle. It gave me back 900 competent words I didn't want to publish.

Not bad words. Just words that had no reason to exist, because I hadn't decided what I thought yet. The agent can't do the deciding, and it will happily fill the gap where the decision should be with structure and cadence. That's the trap: the output looks like progress.

Task shape	Cowork verdict
Fixed input, fixed rubric, fixed format	Run it unattended
Transform something that already exists	Run it, review the diff
Consolidate/extract from data you provide	Run it with the never-invent rule
Decide what you think	Do it yourself first
Write in your voice from scratch	Only after you've decided

The rule I use now: never hand over a task where the hard part is the opinion. Have the opinion, then hand over everything downstream of it. Wednesday would have gone fine if I'd written three sentences of argument first and given it those.

What I'd hand a teammate on day one

If someone joined me tomorrow and I had ten minutes to make them useful with Cowork, this is the whole briefing:

• Build the four files before you build the task. CONTEXT.md, VOICE.md, FACTS.md, inputs/. Twenty minutes, once.
• Paste the never-invent rule into anything touching data. Non-negotiable. Check the "Needs a human" list first when the output lands.
• Pin the output format in the prompt. Headings, field names, max lengths. Vagueness in the spec becomes filler in the output.
• Ban the summary if you plan to reuse the output later. Force the next pass back to the source.
• Review the diff, not the vibe. Output folder only, so you can actually see what changed.
• If you catch yourself editing more than rewriting, it worked. If you're rewriting, the task shape was wrong. That's the signal, not time saved.

The reporting side of this is the one I've systematized most, because it's the one where invented numbers do actual damage. The free Reporting Kit at /skills/ has the refuse-first structure built in, the gap-flagging, the "Needs a human" list, the whole discipline. No email gate. If your week looks like Thursday more than Wednesday, start there.

Where the Hub fits

Honestly: the four free Skills at /skills/ and the four free prompts at /free/ cover a solo marketer's Cowork week. The Reporting Kit handles the data-to-summary path, which is the highest-risk one. If that's your bottleneck, take the free material and don't spend anything.

The paid Full Suite is for when the handoff files stop being yours. Three people writing to the same VOICE.md, each with their own idea of what the never-invent rule means, and a FACTS.md nobody owns. The Suite is the version of this week's setup built for more than one person: the prompts are already pinned to a fixed format, the chains assume a shared context vault, and the reporting prompts share one refuse-first spec so two people's outputs are comparable. That's the problem it solves. If you're solo, you don't have that problem yet, and I'd rather you came back when you do.