← Back to Blog

Claude Chat, Cowork or Code? How to Match Every GTM Task to the Right Mode

By Carol-Lyn Jardine, Partner at Clarity & Motion Collective ·

Most GTM teams still run everything through the chat window, typing turn by turn even when the job could run on its own. Carol-Lyn Jardine recently handed four briefs to Claude, walked away for about three and a half hours, and came back to a week's worth of work: meetings booked, follow-up emails sent, decks built, program charters written. In this session she laid out the simple model she gives clients for picking between chat, work modes like Claude Cowork and ChatGPT Work, scheduled tasks, and code modes like Claude Code and Codex. Then she built a weekly pipeline risk report live, from brief to audit to a scheduled Monday run. This recap covers her four-rung ladder, the three questions that place a task on it, her six-part brief, the audit checklist, and the token habits that keep handoffs cheap.


What changed that makes handing off real work possible?

Two things, in Carol-Lyn's view. The newest models from Anthropic and OpenAI can do far more than most people have tried to give them, and with plugins and connectors they can now act on your behalf, not just answer. What worried her was that she wasn't thinking big enough about what to hand over.

So she tested it. After two weeks of travel she had about a week of client work backed up. She spent an hour to an hour and a half writing an outline of everything that needed to happen, had Fable turn it into a plan, and ended up with four briefs: three for Claude Cowork and one for Claude Code. Then she left the computer.

What came back after about 3.5 hours Where it ran
Meetings scheduled on her calendar Cowork
Follow-up emails sent Cowork
Slack messages drafted and ready to send Cowork
Decks and training materials built Cowork
Charters and project plans for five transformation workstreams, plus a consolidated plan and a tracking dashboard Claude Code

The Claude Code job was the one she wouldn't have trusted to AI a few weeks earlier. It was for a client with a 300-person marketing team working through five major AI transformation initiatives for the coming year. Each workstream needed its own charter and project plan, and the plans had to be aware of each other. Claude drafted all five charters, a plan for each workstream, a consolidated plan, and a dashboard artifact to track the projects.

Instead of a full week of back-and-forth sessions, she spent about a day reviewing everything. Two things made it work:

  • Context. Her knowledge base already held the context files the agents needed. Without them, she says, none of this would have worked.
  • Knowing what good looks like. On review she could spot where Claude had made a call she would have made differently. Hand off the work, not your judgment.

What are the four rungs of the AI work ladder?

The core of the session is a ladder for sorting work. Having Cowork and Code doesn't make chat obsolete, and Carol-Lyn still does about 30% of her own work there. When she audited her week she was surprised the number was that high.

Rung Mode Use it when Claude ChatGPT
1. Chat Turn-by-turn conversation You don't know the shape of the output yet and you're exploring or brainstorming Claude chat ChatGPT chat
2. Assign Work mode You know the assignment and what you want produced Claude Cowork ChatGPT Work
3. Install Scheduled task The job repeats on a cadence Scheduled tasks ChatGPT's scheduled tasks
4. Build Code mode You need a tool you'll reuse over and over Claude Code Codex

Why start in chat when you're unsure? Both Claude and ChatGPT have become very proactive about creating things, and when you don't yet know what you want, that eagerness burns tokens on outputs you aren't ready for. Chat is better at staying in the conversation. If chat still keeps spinning up documents, tell it plainly: "evaluate this, don't create a document, give me the output here." You can also put that preference into memory or settings so it sticks.

On the "Install" rung, she noted that ChatGPT has rolled out its version of scheduled tasks under a new consumer-friendly name. It's the same idea Claude has had for a while, now with some new capabilities on the Claude side. Either way, the work now repeats without you.

How do you decide which rung a task belongs on?

Carol-Lyn gives clients three questions:

  1. Can it run without you? If not, stay in chat. If you can't say "here's the thing I want" and walk away, you're still exploring.
  2. Does it end or recur? If it ends, it's an Assign job for Cowork or Work. If it recurs, install it as a scheduled task.
  3. Is the output a deliverable or a tool? A deck or document is a Work job. A mini app or website is a Build job.

Some GTM examples she mapped to the ladder:

GTM job Rung
Debate your messaging with a messaging skill Chat
Run a persona simulation Chat
Build an account research pack Assign
Produce a weekly pipeline risk report every Monday Assign first, then Install
Turn the logic into a reusable lead routing calculator Build

For really complex tasks that need sub-agents, she goes to Claude Code or Codex. She asks it to first produce a plan she can understand, approves the plan, and lets the tool spin up the sub-agents. You can do this in Cowork, but she finds Code and Codex easier and more efficient for it.

What goes into a good handoff brief?

Two years ago everyone was learning prompt engineering and getting LLMs to write long, elaborate prompts. Carol-Lyn's point is that today's models do better with plain, direct language plus good context. The brief doesn't need to be complicated. It needs to be clear.

Her template has six parts:

Part What it does
Goal What you want done
Finish line What the end state looks like
Context Which files or sources to use
Limits Guardrails, so it doesn't go off and do things you didn't ask for
Flag, don't decide Ask you when it's unsure instead of choosing for you
Proof Show what it used, what it assumed, and how it reached its decisions

How did the pipeline risk report demo work?

The demo used simulated data for a fictitious company. Three context files sat in one folder:

  • Call notes covering the last 30 days of sales calls
  • A pipeline report with everything currently in the pipeline
  • A win/loss review document with the rules and percentages for win rates

Step 1: Assign the report in Cowork

She connected the folder in Claude Cowork rather than attaching each file. Because the brief named the documents by name, the folder connection was enough. She ran it on Opus 5.5 at medium effort. This wasn't a Fable-level task, and in her words, Opus 5.5 is very good.

The brief ended with an instruction to ask questions before starting, and Claude came back with four:

Claude's question Carol-Lyn's answer
Which quarter counts as "this quarter"? Q4, October through December
How should dollars at risk be calculated: full deal amount, only the documented gap, or weighted by win rates? Weighted by win rates, using the win/loss data
Should the revised commit be a range (CRM vs. call notes) or a single number with at-risk deals removed? A range, CRM vs. call notes
Should a soft signal count as risk? Flag it as a watch item, to keep noise out of the report

The report landed back in the folder. Claude confirmed the three source files were untouched and said it had checked the math. The output had an analysis table, watch items, a "flagged but not resolved" section for places where the CRM and call notes disagreed, and revised commit totals.

Step 2: Audit it with a second model

Next she audited the report in ChatGPT, using plain chat mode and a high reasoning setting. She likes making the models work against each other, because sometimes one catches what the other missed. If you only have one LLM, use a more capable model as the auditor.

The audit prompt was short and plain. Paraphrased: you didn't write this report. Read it and the three source files. List every claim or number that isn't directly supported by a source, or that rests on an assumption the report didn't flag. Quote the claim, name the source, and say what's wrong. Don't rewrite anything.

It caught two things worth fixing:

  • The quarter definition. "Q4 is October through December" came from her answer in the session, not from any source file. Before scheduling, she'd put the date range into the context files so the task doesn't depend on a live answer.
  • Overstated risk language. For one deal, the source said low risk but not risk-free. The report called it no risk, which is a stronger conclusion than the source supports.

Why bother? She has set up scheduled tasks in the past without auditing the output first and ended up with a task that kept running on flawed logic every week. Audit once before you install, take the findings back to Cowork, fix the approach, then schedule it.

Step 3: Install it as a scheduled task

Back in Cowork, she requested a scheduled task in plain language: run every Monday at 7, follow the brief, use the three source files, run the audit prompt on the report, put the audit flags at the top, and don't edit the source files.

Claude produced a "weekly pipeline risk report" task set to run unattended. It used her session answers in place of the questions the brief would normally ask, and laid out four steps: write the report, run an independent audit, put the audit flags on top, and write a summary. It also made a few calls of its own that she reviewed:

Claude's decision Reason
Scheduled for 6:50, not 7:00 Tasks set exactly on the hour can start late when many runs are scheduled at once
Today's answers built in So the task can run without asking questions
Report replaced each week The brief allowed only one output file, so no older versions are kept
Same files, same report Unless the context files are refreshed, the report won't change week to week

The last one is the reminder that a scheduled task is only as fresh as its inputs. Once a task is set up, she always triggers a first run right away to check that it behaves the way she wants.

How do you check AI work before trusting it?

Carol-Lyn's audit runs through seven checks:

  1. Does it answer the goal in the brief?
  2. Trace the three biggest decisions it made.
  3. Spot-check every number against the sources.
  4. Have it list its assumptions.
  5. Confirm what it touched and what it left untouched.
  6. Use a second model to audit anything that leaves your desk.
  7. Could somebody else pick this up and run it?

The last check is about handoff beyond yourself. If you went on vacation for two months with scheduled tasks running, could a colleague step in and run the same workflows? She designs every setup with that in mind.

She doesn't audit everything. Brainstorming in chat goes unchecked. She does audit anything complex that's going to a client, anything with numbers she can't afford to get wrong, and anything written in her voice. Audits can also be built into the workflow itself, as audit skills or audit scheduled tasks, like the one in the demo. And if a skill produces the work, the audit findings can be fed back into that skill so it gets better over time. You won't get to error-free, but you get closer to output you'd send to production without as much human intervention.

How do you keep AI handoffs from burning through tokens?

Handoff projects can eat tokens fast. Carol-Lyn's habits for keeping them cheaper:

  • End every brief with "ask me questions." Tuning the output up front costs less than generating everything and correcting it afterward.
  • Ask for the brief in plain text first. Edit it there, and request the file only as the last step.
  • Load connectors only when you need them. Every connected tool adds token overhead because the model keeps checking which ones to use. Disconnect the extras if you're hitting limits.
  • Redo a section, not the whole thing. Regenerating everything wastes time and tokens.
  • Pause scheduled tasks you never look at. She found several of her own running daily that she had set up when the feature launched and no longer used, so she disabled them.

Why is change management the hard part?

The biggest challenge in her client work is getting people to change how they work. What they build is the easier part. Two years ago the debate was whether AI could do the job at all. Now the models are capable enough that the limit is mostly imagination, and the harder part is getting people to collaborate differently with each other and with AI.

She expected to spend most of the past year building. She does build a lot, but she spends just as much time coaching: why teams need shared context, why they should audit each other's work, and which work is AI-ready and which should stay human. Large marketing teams tend to be siloed, and agentic workflows don't work well until those silos come down. The smaller teams she works with are moving very fast. The bigger ones are struggling.

That's why she leans on simple frameworks. A recent manufacturing client, outside her usual tech-company work, had a team that felt out of their depth. If you're already AI native, her challenge is to take the ladder and the brief template to the people around you who aren't there yet.

Key Takeaways for GTM Teams

  • Match the task to the mode. Chat to explore, Cowork or Work to assign, scheduled tasks to install, Code or Codex to build.
  • Use the three questions. Can it run without you? Does it end or recur? Is it a deliverable or a tool?
  • Write a brief, not a fancy prompt. Goal, finish line, context, limits, flag-don't-decide, and proof, in plain language.
  • Context and judgment are what make handoffs work. Keep your context files ready, and review the output against what you know good looks like.
  • Audit before you schedule. A recurring task built on flawed logic repeats the mistake every week. Use a second model as the auditor.
  • Start small. Get one folder, a few source files, and one scheduled task working before you take on complex automations.
  • Watch the token bill. End briefs with "ask me questions," keep connectors lean, and pause tasks nobody reads.

Want to Attend the Next Session?

GTM Clawllective is a community where go-to-market operators share how they're actually using AI in their work, live and unscripted. The live discussion stays off the record, and every recap includes the speaker's demo video. Come see the next build for yourself and join the community.


About the Speaker

Carol-Lyn Jardine is a Partner at Clarity & Motion Collective, where she helps companies with AI adoption and AI transformation strategy. She spent twenty years in marketing, most recently as a head of marketing, before moving into full-time consulting about fourteen months before this session. Most of her clients are tech companies, and her work ranges from building agentic workflows to coaching large marketing teams through the change.

Her approach is to break AI work into simple frameworks that explain the why, so teams can keep up as models and features change without feeling lost.

Connect with Carol-Lyn on LinkedIn.