Hand over your red pen: turn your review comments into an AI checklist

Look back through the comments you left on your team's last five drafts. If you are anything like the manager in the film, the same three notes turn up on nearly every one: where's the ask, where's the source, what if we do nothing. Each costs the author a round trip, and it costs you the time it takes to spot the same gap again on a draft you could have been reading for the argument.

Those repeat comments are a specification. They say, in your words, what a draft has to do before it is worth a senior reader's time, and none of them needs a senior reader to check. A missing source does not need a director to find it. Judging whether the recommendation is the right bet does.

Both vendors name the same trigger for writing something like this down. Anthropic's Claude Code documentation says a playbook you paste for the third time should become a skill [1], and OpenAI's guidance is that a prompt you keep reusing, or a workflow you keep correcting, should probably become one [2]. A comment you write by hand on every draft is the same signal, arriving in pen.

The play is short. Turn the comments into a checklist, add one line that makes the assistant point at problems rather than fix them, test it on a draft you have already marked, and share it as a skill your team runs before a draft reaches you.

Do this in the Claude or ChatGPT account your employer provides, not a personal one. Collect your own comments from your team's last five drafts, wherever they live: margin notes, tracked-change comments, the email you sent back with the draft. Leave the drafts themselves out, and take out every name, whether it belongs to an author, a client or anyone else the comments mention. Then paste the comments into a new chat and type this.

Five drafts is enough for your habits to show and short enough to paste in one go. If the list that comes back is long, it has mixed your rules with your one-off remarks, so keep only the points you flag on draft after draft and let the rest go.

Then check each item is something an author could test without asking you what you meant. “Every figure says where it came from” can be checked, and “be more rigorous” cannot. Anthropic gives the same advice for the instruction files Claude Code reads: write them concrete enough to check [3].

Keep your wording. Your team already knows what “where's the ask” means coming from you, so the checklist needs no translation, and it reads as your standard because it is one.

A checklist on its own leaves the assistant free to be helpful in the wrong way, by rewriting the draft. One instruction above the checks stops that. Here it is with the film's three comments as the example checks; put yours in their place.

Quoting makes every finding checkable. The author sees the exact sentence that failed, and so do you, so a wrong flag takes seconds to dismiss instead of a paragraph of advice to decode. It is also one of the guardrails Anthropic documents against made-up answers: have the model find a supporting quote for each claim and drop any claim it cannot support [4].

Rewriting nothing keeps the draft the author's. They fix the sentence themselves, which is how the comment eventually stops being needed, and nobody has to work out what the assistant quietly changed. OpenAI's prompting guide makes the general point that one or two boundaries, such as keeping approved figures unchanged, prevent the problems that matter [5].

Two more lines earn their place after a few runs. Tell it to say plainly when every check passes, because Anthropic warns that a reviewer asked to find problems will usually find some [6]. And tell it to report only the checks on the list, so a fourth item you never wrote does not creep in.

A generic “review my draft” prompt gets generic advice. This version works for reasons you can trace in each vendor's own guidance.

The basic play covers one reviewer and one kind of document. These are the variations that come up next, and the test that should come before any of them is shared.

If it misses something you always catch, that check is worded too loosely. If it flags things you would let through, the check is too broad, or it is firing on drafts it was never meant for. OpenAI's rule for a misbehaving skill draws the same line: if it fires at the wrong time, fix the description, and if the results are inconsistent, fix the instructions [15]. Run the same test again after each change, and only then share it.

Both assistants save a reusable set of instructions as a skill, both can build one from a conversation, and in both an administrator decides who may share it. These are the steps as each vendor documents them, checked 11 October 2026.

The Claude column is from Anthropic's help centre articles on using skills [16] and on managing them for an organisation [17], and its skills documentation [11, 18]. The ChatGPT column is from OpenAI's Help Center article on skills [19] and its documentation on building them [12].

Three details matter for this play. In Claude, a custom skill stays invisible to colleagues until you share it, and a shared skill is view-only and stays off until each person switches it on, so tell the team to switch it on [16]. In ChatGPT, OpenAI documents the Skills tab for eligible Business, Enterprise, Healthcare and Edu accounts [19], so check your workspace has it before you promise it to the team. In both, whether you can share at all is your administrators' decision [16, 19].

If skills are switched off where you work, paste the checklist and its line on top into a project's instructions instead. In Claude, project instructions apply only to chats in that project, and on Team and Enterprise a project can be shared with Can view or Can edit access [20]. In ChatGPT, a project's instructions apply across the chats inside it [21]. You lose the automatic loading and keep the checks.

If your company has not yet settled which assistant it pays for, our side-by-side of ChatGPT Enterprise and Claude Enterprise covers what each plan publishes on price, admin controls and data location.

A shared skill is a small piece of software the whole team will run on its work, and both vendors' guidance treats it that way.

The play is cheap to run and easy to over-trust. These are the limits to keep in view.

One reviewer can do everything above alone. A company doing it across every team needs more: the reviewers whose comments set each standard, skills tested on drafts those reviewers have already marked, and someone deciding who may share and publish them. Setting up and rolling out Claude and ChatGPT for companies is core 1AYM work, including the instructions, skills and permissions around them, and where a check has to hold every time rather than guide, we build it as a control under our AI governance implementation work. The nearest published case is engagement file D-01: at a large international marketing agency we contribute engineering to an internal Claude Code skill platform of around 600 production skills, which the people who do the work write themselves.

We take small fixed-scope statements of work as well as larger builds, and fixed-scope work can start within a day of the scope being signed. If a second pair of eyes on your teams' review skills would help, book a call or email enquiries@1aym.com.