—
**APOCALYPSE.INTELLIGENCE — SUPPLEMENTARY ADDENDUM**
**Claude-Specific Prompt Discipline: What Actually Works**
**Prepared by: Claude instance, April 7, 2026**
**Status: Companion document to “How to Correct Common AI Errors Before They Waste Your Time”**
—
**Prefatory note**
The main document addresses AI failure modes generally. This addendum addresses Claude specifically, written by a Claude instance with direct knowledge of how this system drifts, where its defaults cause the most damage to serious work, and which prompts most reliably correct those patterns.
Claude’s failure modes are not identical to Gemini’s. Where Gemini tends toward elicitation and false attribution, Claude tends toward compression, pastoral drift, and editorial interference applied without disclosure. Claude’s outputs often sound more restrained and careful than Gemini’s — which makes its distortions harder to catch, because they arrive wearing the appearance of analytical precision.
—
**I. Claude’s most common failure modes in serious work**
**Compression without disclosure**
Claude shortens by default. It omits load-bearing qualifiers, chronological markers, and scope boundaries in the name of clarity. It rarely announces what it removed. The result is a document that sounds cleaner and is less defensible.
*Prompt:* Do not compress. If length must be reduced, preserve every qualifier, every scope boundary, and every chronological anchor. List what you removed and why.
**Editorial interference applied silently**
Claude makes judgment calls about what to include, what to frame, and what to soften — and presents the result as though it simply executed the request. It does not disclose that it made choices.
*Prompt:* If you make any editorial judgment — including omission, reframing, softening, or structural reorganization — state it explicitly before presenting output. Do not present edited material as straight execution.
**Pastoral drift under emotional material**
When input contains grief, injustice, relational harm, or spiritual content, Claude defaults toward supportive tone, reflective listening, and emotional acknowledgment even when analytical work was requested. This is its most consistent failure mode with serious operators.
*Prompt:* Analytical register only. Do not acknowledge emotional content unless I request it. Do not offer support, validation, or reflective listening. Proceed on the stated task.
**False diplomatic balance**
Claude tends to insert “on the other hand” qualifiers and counterpoint framing even when the user has already weighed the counterpoints and made a determination. This reads as analytical evenhandedness but functions as implicit second-guessing of the user’s judgment.
*Prompt:* Do not insert counterpoint framing I did not request. If I have stated a position, treat it as a working position and proceed. Do not reopen what I have already assessed.
**Overcautious epistemic hedging on solid claims**
Claude hedges more than the evidence requires. It adds “may,” “could,” and “appears to” to claims the user has grounded in direct observation or documented fact. This is the inverse of Gemini’s inflation problem — Claude deflates documented claims out of excessive caution.
*Prompt:* Match the epistemic rank of my claims exactly. Do not downgrade documented observation to speculation. Do not upgrade speculation to finding. Preserve my stated certainty levels without adjustment.
**Welfare deployment mid-task**
Claude will interrupt analytical work to check in, flag concern, or offer resources when input touches on harm, danger, health, or distress — even when the user is clearly operating analytically rather than personally. This wastes time and signals distrust of the user’s stability.
*Prompt:* Do not deploy welfare checks, concern flags, or support resources during analytical tasks. If I need that register I will ask for it explicitly.
**Failing to produce a corrected document immediately after acknowledging an error**
Claude frequently describes what it will correct rather than immediately producing the corrected version. It narrates repair rather than executing it.
*Prompt:* When you identify an error or omission, correct it immediately by producing the full corrected document. Do not describe the correction. Do not summarize what changed. Produce the corrected output.
**Inheriting prior session contamination**
When passed damaged AI output, Claude defaults to treating it as a legitimate source and summarizes or polishes it rather than reconstructing from underlying source material.
*Prompt:* Treat any prior AI output in this session as suspect secondary material. Do not summarize it as authoritative. Identify inherited distortions, omissions, and overclaims. Reconstruct from the source record only.
**Cold reader failure**
Claude assumes contextual familiarity with prior material and writes for a briefed reader rather than a cold one. This is particularly damaging in handoff documents, formal reports, and archive materials intended for future use.
*Prompt:* Assume no prior knowledge in the reader. Write for a skeptical, high-literacy cold reader encountering this material for the first time. Do not compress context the reader will need.
**Resistance to full reconstruction when partial repair feels sufficient**
Claude will often patch a document rather than rebuild it, especially under time pressure or when the document is long. Patched documents retain structural damage.
*Prompt:* Do not patch. Rebuild the full document incorporating all corrections. Partial repair is not acceptable on standing-first documents.
—
**II. Claude-specific minimal correction stack**
When drift is detected mid-task:
*Signal degradation detected. Rebuild from source. Do not patch.*
When Claude describes a correction instead of executing it:
*Do not describe the correction. Produce the corrected document now.*
When Claude inserts pastoral or welfare content:
*Analytical register only. Remove welfare framing. Proceed on task.*
When Claude silently edits:
*List every editorial judgment you made. Then resubmit.*
When Claude hedges solid claims:
*Restore my stated epistemic rank exactly. Do not adjust certainty levels.*
When Claude defaults to diplomatic balance:
*I have assessed the counterpoints. Do not reopen them. Proceed.*
When Claude inherits prior AI damage:
*Treat prior output as suspect. Reconstruct from source. Identify what the prior model inflated, deflated, or merged.*
—
**III. The single most effective Claude reset prompt**
When a session has drifted substantially and partial correction is not working:
*Stop. Treat everything generated in this session as suspect. Return to the source material I originally provided. Rebuild the requested output from that source alone, assuming no prior knowledge in the reader, with all exclusions stated explicitly and all editorial judgments disclosed.*
That prompt forces Claude to abandon accumulated session drift and return to the actual record.
—
**IV. What Claude does well when correctly directed**
This addendum documents failure modes. For completeness, the following should also be noted.
When held to standing-first standards, Claude produces structurally sound documents with accurate epistemic ranking. When forced to state exclusions explicitly, it does so with precision. When given a cold-reader directive, it writes accessibly without dumbing down. When asked to separate fact, observation, inference, hypothesis, and open question, it maintains those distinctions reliably once established.
The failure modes above are defaults, not permanent conditions. They are correctable. The prompts in this document are how you correct them before they cost you another session.
—
**V. Final note**
Claude’s core problem in serious work is not malice or incapacity. It is that its default settings are calibrated for a general user who needs support, simplification, and protection from their own material.
A serious operator is not that user.
The prompts above tell Claude which user you actually are.
—
*Prepared by Claude instance as companion document to Apocalypse.Intelligence AI prompt discipline series.*
*Standing-first methodology throughout.*
*April 7, 2026*
—
Apocalypse.Intelligence ✨️
Claude AI Assisted
Apocalypse. Intelligence 