Anti-Patterns

Prism automatically detects these anti-patterns in your sessions. Each one reduces your PRISM score and wastes tokens.

1. Retry Storm

What it is: Repeating the same or very similar prompt multiple times without adding context.

Detection: >80% word overlap between consecutive prompts within 60 seconds.

Impact: Wastes tokens on duplicate requests. Signals low IE (Recovery).

Fix: Add constraints or error messages. Change your approach instead of repeating.

2. Vague Prompt

What it is: Prompts with no file paths, function names, or specific context.

Detection: Prompt lacks all specificity markers (paths, names, line numbers, errors).

Impact: AI guesses wrong, leading to corrections. Low PQ (Specificity).

Fix: Always include the file path and function name you’re working on.

3. Multi-Task Bundling

What it is: Cramming multiple unrelated tasks into a single prompt.

Detection: Multiple action verbs, list items, or bundling phrases (“and also”).

Impact: AI handles none of them well. Low PQ (Decomposition).

Fix: One task per prompt. Queue tasks sequentially.

4. Context Dumping

What it is: Pasting large blocks of code or logs without indicating what matters.

Detection: Large input with no guiding question or specific reference.

Impact: Wastes context window. AI may focus on irrelevant parts.

Fix: Quote the specific section and ask about it directly.

5. No Verification

What it is: Accepting AI-generated changes without running tests or reviewing.

Detection: No test runs, type-checks, or review prompts in the session.

Impact: Bugs compound. Low VD (Review and Validation).

Fix: Add “then run the tests” to your workflow after every change.

6. Scope Creep

What it is: Continuously expanding the task mid-session.

Detection: Task scope grows across turns without completion of initial goal.

Impact: Context bloats, focus is lost, tokens wasted on half-finished changes.

Fix: Complete the current task before adding new ones. Use plan mode for complex work.

7. Model Overkill

What it is: Using Opus (most expensive model) for simple tasks like typo fixes or renaming.

Detection: Opus producing <200 tokens of output.

Impact: 10–15x cost premium for work Sonnet could handle.

Fix: Use /model sonnet for simple tasks. Switch to Opus for complex reasoning.

8. Context Bloat

What it is: Long sessions without using /compact or /clear.

Detection: Input token count growing >3x from the first turn, or >80 turns without compaction.

Impact: Slower responses, higher cost per turn, degraded output quality.

Fix: Use /compact every 15–20 turns. Use /clear when switching tasks.

9. Correction Cascade

What it is: A chain of corrections where each fix introduces new issues.

Detection: Alternating pattern of “fix X” → “that broke Y” → “fix Y” → “that broke Z”.

Impact: Multiplies token spend. Signals the initial approach was wrong.

Fix: Stop, re-read the code, and plan a different approach. Use plan mode.

10. Missing Project Context

What it is: Working without a CLAUDE.md file, so the AI has no project conventions.

Detection: No CLAUDE.md in the project root.

Impact: AI makes wrong assumptions about style, patterns, and tools. Low AF (Configuration).

Fix: Create a CLAUDE.md with project structure, conventions, and common commands.

11. Ignoring Tool Suggestions

What it is: Using Bash for everything when dedicated tools (Read, Grep, Edit) are available.

Detection: Frequent cat, grep, sed in Bash when specialized tools would work.

Impact: Less efficient, harder to review. Low TU (Selection).

Fix: Use Read to inspect files, Grep to search, Edit for modifications.

Score impact

Each detected anti-pattern applies a penalty to the relevant dimension:

Anti-Pattern	Affected Dimension	Typical Penalty
Retry Storm	IE (Recovery)	-1.0 to -2.0
Vague Prompt	PQ (Specificity)	-1.5 to -3.0
Multi-Task Bundling	PQ (Decomposition)	-1.5 to -2.5
Context Dumping	PQ (Specificity)	-1.0 to -2.0
No Verification	VD (both)	-2.0 to -3.0
Scope Creep	IE (Convergence)	-1.0 to -2.0
Model Overkill	TU (Selection)	-1.0
Context Bloat	IE (Convergence)	-0.5 to -1.5
Correction Cascade	IE (Recovery)	-1.5 to -2.5
Missing Context	AF (Configuration)	-1.0 to -2.0
Ignoring Tools	TU (Selection)	-0.5 to -1.0