AI revision without the scar tissue

One of the most annoying failure modes of AI-generated writing (or code) is the scar tissue that builds up after several rounds of feedback/iteration. Let’s say you have a skill called concise-copy:

Keep responses concise and direct.

But then you notice that the agent is being a little too terse, so you tell it not to drop important context just for the sake brevity. You take a look at the updated SKILL.md file and see something along the lines of:

Keep responses concise and direct.

IMPORTANT: Do not be overly terse. Always include sufficient context.

CRITICAL: Concision must never come at the expense of completeness.

Technically, none of that is wrong. The problem is that the document has now started preserving the history of its own correction.

Raise your hand if you’ve experienced this before: you read a file and can more or less reconstruct the back-and-forth between the user and the LLM that led to the final output. 🙋🏻‍♂️

After enough iterations you have a jumble of warnings, exceptions, and counter-exceptions where the loudest instruction usually reflects the most recent mistake rather than the most important principle.

Ideally, the revision should just disappear into the design:

Be concise and direct while preserving the context necessary for a complete and useful answer.

Instead of “memorializing” my feedback, understand what it implies, reconsider the surrounding material, and produce the version I would have written had I understood the problem better from the beginning. 🙃

I like to call this integrative revision, and it should treat new feedback as evidence about what the artifact is ultimately trying to accomplish. The question isn’t where should I add this correction? but given what I know now, what should the whole thing say? Sometimes that means adding a sentence, but more often it means rewriting or deleting existing language. And sometimes a specific mistake reveals a general principle that makes several old instructions unnecessary.

Undue emphasis is another tell-tale sign of iterative AI revision. If you point out a mistake forcefully enough, the next revision will likely revolve around preventing that one single mistake, e.g. a minor edge case becomes a bold heading, a single bad output becomes an ALWAYS rule, a local correction reshapes the tone of an entire document.

Prominence starts tracking the recency or intensity of the feedback rather than the importance of the idea. The fix is what I like to call proportional salience, i.e. how prominent an instruction is should reflect what its worth is to the system as a whole, not how emphatically someone complained about it.

I include the following in any prompt or skill passed to agents that must maintain artifacts over time. The core idea is to incorporate feedback by redesigning the artifact, not by accumulating patches.

When incorporating feedback, corrections, criticism, or new constraints into an existing artifact, treat them as information to integrate into the artifact's underlying logic rather than as amendments to append or corrections to foreground.

Prefer synthesis over accretion.

Revise existing language, structure, and assumptions as necessary so that the resulting artifact remains coherent, internally consistent, and even in emphasis. Do not make a newly supplied correction unusually prominent merely because it is recent, emphatic, or arose from a previous mistake.

Avoid turning iterative feedback into an accumulating history of patches, warnings, exceptions, disclaimers, or special-case rules. When new guidance overlaps with, qualifies, supersedes, or conflicts with existing guidance, reconcile the relevant material and rewrite it into the smallest coherent set of principles that captures the intended behavior.

Preserve proportional salience. The emphasis, repetition, placement, and structural prominence of any instruction or idea should reflect its actual importance to the artifact as a whole, not the recency or intensity of the feedback that introduced it.

Do not "show your work" by conspicuously echoing the latest correction unless doing so is genuinely useful to the artifact. Successful incorporation of feedback is demonstrated by the quality and coherence of the revised result, not by visibly advertising that the feedback was incorporated.

Where possible, generalize from a specific correction to the underlying principle that would have prevented the issue, while avoiding overgeneralization. A local mistake should not automatically become a global prohibition.

After multiple rounds of revision, the artifact should still read as though it were designed coherently in its current form. A reader should generally not be able to reconstruct the sequence of corrections, disputes, or mistakes that produced it.

Apply the same principle to tone and framing: do not let the latest criticism disproportionately reshape the voice, priorities, or character of the whole artifact unless the feedback actually warrants such a change.

The goal is not to preserve the history of revision. The goal is to produce the best current version.