Self-Critique & Refinement Loops
Getting Claude to catch and fix its own mistakes in a structured second pass
The single highest-leverage two-step trick
Of all the advanced techniques in this course, self-critique loops may have the best effort-to-improvement ratio. The idea is simple: instead of accepting Claude’s first draft, run a second prompt asking Claude to critique that draft against specific criteria, then a third asking it to revise based on the critique. This consistently produces better output than asking for perfection on the first try — because critiquing existing text is a fundamentally easier task than generating ideal text from nothing.
The three-step loop
STEP 1 — Draft:
"Write a 300-word LinkedIn post announcing our new pricing tier
for solo consultants."
STEP 2 — Critique (paste the draft back in):
"Critique this LinkedIn post against these criteria:
1. Does the hook (first line) earn a scroll-stop, or is it generic?
2. Is there a specific number or detail, or is it all abstraction?
3. Does it sound like a person or like corporate marketing copy?
4. Is the call to action clear and low-friction?
Be specific and harsh — name the weakest sentence."
STEP 3 — Revise (paste the critique back in):
"Revise the post to address every point in this critique. Don't
just soften the critique's language — actually fix each issue."The critique step is where the real value is created. A vague critique produces a vague revision; a critique that names the specific weak sentence and the specific reason produces a genuinely better second draft.
Why a rubric beats “make it better”
Asking Claude to simply “critique this and make it better” tends to produce shallow, agreeable feedback — mild positive spin with a few generic suggestions. A rubric with named, specific criteria forces the critique to actually evaluate against something, rather than free-associate.
Key principle: The quality of a self-critique loop is bottlenecked by the quality of the critique criteria, not by asking for “more” critique. Three sharp, specific criteria beat ten vague ones.
Avoiding self-congratulatory critique
Left unguided, Claude’s critique of its own work can lean toward being generically positive — noting a few small tweaks rather than identifying real weaknesses. Two things counter this reliably: explicitly asking it to “be harsh” or “find the weakest part,” and giving it an external standard to compare against (a competitor example, a past piece that performed well, an explicit rubric) rather than asking it to judge in a vacuum.
Compare this draft against the tone and specificity of this example
of a post that performed well for us: [paste past example]
Where does the draft fall short of that bar?Tip: For high-stakes work, run the critique step with a different framing than the one used to generate the draft — for example, critique “as a skeptical reader who sees fifty of these posts a day” rather than as the same helpful assistant persona that wrote it. A different vantage point surfaces different flaws.
Knowing when to stop looping
Self-critique loops have diminishing returns — usually one critique-and-revise cycle captures most of the available improvement, and a second or third cycle tends to produce marginal polish rather than substantive fixes. If a second loop’s critique starts nitpicking word choice instead of finding real issues, that’s your signal to stop.
Try it: Take a piece of writing Claude generated for you recently that was “fine but not great.” Run it through a critique step with 3 named criteria, then a revision step. Compare the revision to the original.