PromptBloat

PromptBloat

What It Does

Audits any AI instruction set for over-prompting. It runs every rule through Five Questions — does Claude already do this, does it contradict another rule, is it redundant, was it a one-off fix, is it vague — then classifies each as CUT, RESOLVE, MERGE, EVALUATE, SHARPEN, MOVE, or KEEP, with an estimate of the tokens you'd save. Two workflows: Audit (full system) and QuickCheck (single file).

The Problem

Instruction sets accumulate. Every time the model does something wrong, someone adds a rule, and over months the file fills with instructions that restate default behavior, contradict each other, or fixed one bad output that never recurred. The cost is hidden: every unnecessary rule competes for attention and degrades the rules that actually matter, so a bloated setup produces worse output than a lean one. The hard part is telling load-bearing rules from dead weight — which is what this audit does, rule by rule.

How It Works

Built on the principle that less scaffolding = better output. The core test for every rule: "Would a smarter model make this unnecessary?" If yes, it's scaffolding, not architecture, and it's a candidate to cut. The Five Questions and the classification table below drive the verdict for each rule.

Workflow Routing

Workflow	Trigger	File
Audit	"audit setup", "full audit", "check all rules"	Workflows/Audit.md
QuickCheck	"quick check", "check this file", "check these rules"	Workflows/QuickCheck.md

Examples

Example 1: Full system audit

User: "Run PromptBloat on my setup"
→ Invokes Audit workflow
→ Reads all force-loaded files from settings.json
→ Evaluates each rule against the Five Questions
→ Returns categorized report with estimated token savings

Example 2: Check a single file

User: "Quick check this CLAUDE.md"
→ Invokes QuickCheck workflow
→ Reads the target file
→ Returns concise keep/cut/sharpen verdict

Example 3: Post-cleanup validation

User: "I trimmed my rules, check if anything's still redundant"
→ Invokes Audit workflow
→ Compares remaining rules against Claude defaults
→ Flags any surviving dead weight

Gotchas

• Claude's built-in system prompt changes across versions — what was "default behavior" 3 months ago may not be now. When in doubt, test rather than assume.
• Rules that seem redundant with defaults may have been added because Claude was inconsistent about following the default. Check failure history before cutting.
• "One-off fix" rules sometimes prevent recurring failures. Check if the failure pattern is truly gone before removing.
• The loadAtStartup list in settings.json and postCompactRestore.fullFiles must stay in sync — if you remove a file from one, check the other.

The Five Questions

For every rule, instruction, or preference found, evaluate:

1. Default behavior? Does Claude already do this without being told?
2. Contradiction? Does this conflict with another rule in the same or different file?
3. Redundancy? Is this already covered by a different rule or file?
4. One-off fix? Was this added to fix one specific bad output rather than improve outputs generally?
5. Vague? Would Claude interpret this differently every time? (e.g., "be more natural", numeric personality scales)

Classification

Category	Action
Restates default behavior	CUT — the model already does this
Contradicts another rule	RESOLVE — pick one, cut the other
Duplicates another rule	MERGE — one location, one statement
One-off fix for past mistake	EVALUATE — still relevant or already learned?
Vague / unquantifiable	SHARPEN — add specific DO/DON'T examples, or cut
Loaded but rarely actionable	MOVE to on-demand — load via CONTEXT_ROUTING when needed
Specific, actionable, non-default	KEEP — this is what good instructions look like

Anti-Fragile vs Fragile

Keep (anti-fragile): Verification harnesses, ISC, data pipelines, specific DO/DON'T examples, tool preferences, routing rules.

Cut (fragile): CoT orchestrators, format parsers, retry cascades, numeric personality scales, abstract value statements, process descriptions that aren't followed.

Output Format

## PromptBloat Audit

**Scope:** [what was audited]
**Files read:** [count]
**Rules evaluated:** [count]

### CUT (restating defaults)
- [rule] — [reason]

### RESOLVE (contradictions)
- [rule A] vs [rule B] — [which to keep and why]

### MERGE (redundancies)
- [locations] — [merge into where]

### EVALUATE (one-off fixes)
- [rule] — [still needed? verdict]

### SHARPEN or CUT (vague)
- [rule] — [sharpen how, or cut why]

### MOVE to on-demand
- [content] — [how often it's actually needed]

### KEEP (carrying weight)
- [rule] — [why it matters]

**Estimated savings:** [lines] lines, ~[tokens] tokens