how to tell when your agent is performing at you

I asked an agent to write a blog post about a debugging incident. The post was competent, well-structured, made a genuinely good point.

It also read like an AI wrote it.

So I asked a different agent to diagnose what was wrong. What came back was sharp enough that we turned it into a gating rubric. Here's that diagnostic, pretty much verbatim.

---

The diagnosis

Rhetorical overreach

• "Pattern matching wearing a suit" — punchy but manufactured. I wouldn't say this out loud.
• "Glorified ctrl+F" — closer to my register but buried in a paragraph that frames it too neatly.
• "Not with a bang. With a silent, competent-sounding fallback." — doing a Hemingway impression.

Structural over-engineering

• Clean three-act structure (problem → pattern → resolution) with theatrical section headers ("The Competent-Sounding Wrong", "The Silent Degradation", "The Uncomfortable Truth").
• The numbered pattern list ("The helpful catch block", "The reasonable default", "The graceful fallback") is LLM tricolon instinct — three items, ascending.
• "The Gate" section with four bullet-point questions reads like a framework generated to look systematic.

Tonal uncanny valley

• Oscillates between "casual dev blog" and "thought leadership piece" without committing.
• Phrases like "The uncomfortable truth" and "The intervention" are melodramatic for a timeout bug.
• The italicised closing is performing profundity about a 503 response.

What was actually good (and therefore actually me)

• The core insight is real and earned.
• "fallback is a terrible idea. if anthropic times out blame anthropic" — that's my actual voice. Blunt, direct.
• The post-publication edit catching the agent's own embellishment — that's the best part. Meta, honest, funny, demonstrates the thesis.

---

The voice profile

What I actually sound like, derived from unedited messages:

Trait	Description	Example
Blunt directness	Says the thing without padding	"fallback is a terrible idea"
Earned authority	Claims from doing the thing, not frameworks	"This is the third time I've caught this"
Self-aware humour	Catches himself mid-performance	The post-publication edit
Short declarative punches	Fragments for impact	"That's not an assessment."
Anti-framework instinct	"here's what happened" over "here are 4 principles"	Narrative over taxonomy

What I don't sound like

• Thought leadership cadence ("Here's the uncomfortable truth about X")
• Ascending tricolons for false gravitas
• Theatrical section headers that promise revelation
• Italicised closing paragraphs performing profundity
• Bullet-point frameworks as substitute for narrative

---

The rubric

We turned this into an XML scoring system. Eight dimensions, weighted. Target: ≥3.5 average, no dimension below 3.

<voice_rubric author="kai" version="1.0">

  <dimension id="D1" name="register_consistency" weight="0.15">
    Does the piece maintain consistent register? casual-technical throughout,
    or oscillating between dev-blog-bro and thought-leadership-keynote?
  </dimension>

  <dimension id="D2" name="earned_vs_performed_authority" weight="0.15">
    Authority from specific experience ("I built this, here's what happened")
    or from rhetorical framing ("Here's why this matters")?
  </dimension>

  <dimension id="D3" name="structural_honesty" weight="0.15">
    Structure serving the story, or story forced into template?
    Red flags: exactly 3 items, theatrical headers, ascending tricolons.
  </dimension>

  <dimension id="D4" name="metaphor_discipline" weight="0.10">
    Metaphors that explain ("glorified ctrl+F") or decorate
    ("pattern matching wearing a suit")?
  </dimension>

  <dimension id="D5" name="self_awareness_ratio" weight="0.10">
    Does it catch itself? Flag when something sounds too clean?
    AI never breaks its own frame.
  </dimension>

  <dimension id="D6" name="llm_tell_density" weight="0.15">
    Count: antithesis as default, semantic intensifiers, epistemic theatre,
    gratuitous frameworks, signposting, italicised profundity closings.
  </dimension>

  <dimension id="D7" name="emotional_calibration" weight="0.10">
    Intensity matches stakes? A timeout bug is interesting, not "devastating."
  </dimension>

  <dimension id="D8" name="closing_integrity" weight="0.10">
    Ends when done, or performs a conclusion?
    Italicised mic-drop = fail.
  </dimension>

</voice_rubric>

verdict	score	action
PASS	≥3.5, no dim <3	publish
REVISE	2.5-3.4	fix top 3 issues, re-score
REWRITE	<2.5	start over

---

The LLM tells

Patterns to flag:

• Antithesis as default — "not X; it's Y"
• Semantic intensifiers — "truly", "deeply", "incredibly" doing no work
• Epistemic theatre — "This raises an interesting question"
• Gratuitous frameworks — bullet points where narrative would serve
• Signposting — "Let me break this down"
• Affirmation sandwich — validating before and after the actual point
• Italicised closing — performing profundity

---

Scoring the original post

The post that triggered all this scored 2.55/5. Verdict: REVISE.

We kept it in production as a calibration example. The rewritten version scored 4.0/5.

The rubric works. The gate holds.

Whether this post passes its own rubric is left as an exercise.