The post (its title translates roughly as “How I got greedy with a single AI reviewer in Claude Code, failed, and split it into three agents”) is a practitioner’s account of using AI review on seminar slides. Its starting complaint is that asking one AI to “review this” returns inoffensive impressions — “well organised overall, and the flow is natural” — that leave nothing to fix. The author’s diagnosis is that one reviewer asked to look at everything mixes perspectives whose pass lines point in opposite directions, so they cancel out into “probably fine on the whole”.
The remedy is three Claude Code sub-agents, each defined in its own file
under .claude/agents/. The author prepares slides in three stages — an outline that fixes audience
and takeaway, a text-only storyboard that fixes what each slide says and in what order, and the
finished slide deck — and the agents are handed different inputs from them, because what an agent
can judge depends on what it is given: the harsh reviewer gets the slides, the logic reviewer gets the
storyboard together with the relevant range of slides, and the audience-reaction agent gets only the
slides of the chapter in question. The centre of the post is a failure: running the harshest of the three
repeatedly on an outline made the outline longer and its score lower, which is what led to the
second agent (see Review Loop Non-Convergence). The post publishes all three agent
definitions in full as an appendix.
Key Points
- A single reviewer told to check everything gives bland, unactionable feedback, because checks with opposite pass lines — logic versus the sharpness of the wording — cancel each other out.
- harsh-review-agent scores from 0 (“the bin”) and looks for evidence of why the material deserves 0: praise is forbidden, AI-typical “dead words” such as “optimisation”, “seamlessly” and “DX promotion” are listed and condemned, and a raw score somewhere in the 10s to 30s is treated as normal for AI-generated material. The author uses it only once the slides exist.
- Running the harsh reviewer on an outline for three fix-and-resubmit cycles grew the outline from 685 to 772 lines while its score fell from 47 to 44: because the agent is built to always find evidence, each addition made to answer a finding drew new findings of its own, such as “too dense”.
- When the author sorted the harsh reviewer’s findings into logical breakdowns and matters of taste, the majority turned out to be taste; the author concluded it was a fastidiousness filter, not a detector of logical failure.
- logic-reviewer scores from 100 and may report only five kinds of logical breakdown — a failed link from premise to conclusion, a term whose meaning shifts, a promise made by a title or opening that the body never fulfils, a missing dependency, and internal contradiction — and must report 100 when it finds none, without “just in case” remarks. It can be used from the outline stage onward.
- audience-reaction-agent plays one persona in the audience, reads only the scope it is given and nothing after it, ignores the speaker’s intent notes in the storyboard, gives no score and does not condemn; the author’s point is that designing what the agent is not given is what makes its reaction resemble a real listener’s.
- All three produce read-only, one-way reports saved by date and type; the agents do not edit the material, and the author decides what to adopt: earlier decisions come first, clear logic errors are fixed at once, and where agents disagree the author takes the one closer to their own intent or proposes a new option.
- From his own trials, the author reports that an agent reviewing slides for visual crowding caught 1 of 7 cases, and that piling up reviewers that all push to cut things erodes the argument itself.
Context
The post is one developer’s account of their own setup, published on their employer’s engineering blog, and its evidence is the author’s own runs and scores rather than any measured comparison. The author sums the approach up as splitting by perspective, inverting the starting point, and narrowing the input, and says it is not specific to slides: the author also uses the harsh and logic reviewers on blog posts, and the logic reviewer on outlines, proposals and specs. The post leaves open whether defending the author’s own claims against the reviewers could also be delegated to an agent, saying the right balance has not been found yet. Its account of review findings being weighed by a person, rather than applied wholesale, sits close to Review Finding Triage, and its use of several narrowly scoped agents is an instance of Sub-agent architecture.