MinimizePrompt

MinimizePrompt

Cut motivational, marketing, and emphasis-inflated prose from prompt-shaped documents. Preserve the directive content and the facts that make the artifact work.

Aligned with [MVPR-0001](docs/decisions/MVPR-0001 Minimum Viable Prompt.md) and [MVPR-0002](docs/decisions/MVPR-0002 Prompt Minimalization Metrics.md). Referenced by ForgeAdopt as the minimize transform.

What to remove

Pattern	Example
Motivational framing	"Don't hold back", "Claude is capable of extraordinary work"
Emphasis inflation	CAPS-LOCKED warnings, repeated **CRITICAL** or **IMPORTANT**
Marketing superlatives	"Production-ready", "Best-in-class", "Complete toolkit"
Filler transitions	"First of all", "It's worth noting", "As previously mentioned"
Tautology	"Use good practices" without defining good practices
Duplicate summaries	Opening paragraph repeats the description; closing repeats the opening
Decorative checkmarks	✅-bulleted feature lists that restate workflow content

What to preserve

• Workflow steps and decision points
• Anti-patterns with concrete examples ("avoid X because Y")
• Tables, code blocks, and command examples
• Constraints and hard rules
• Frontmatter fields that affect routing — description, allowed-tools, name, version

Red flags that a section should stay

A section earns its tokens if it does at least one of:

• Cites a specific source, rule, benchmark, or prior decision
• Names a concrete anti-pattern with a reason
• Is referenced by another section in the same document
• Carries a command, regex, schema, or structured data
• Encodes a decision that cannot be reconstructed from names alone

If none of the above applies, the section is a candidate for removal.

Procedure

1. Read the document end-to-end once before cutting.
2. Identify the load-bearing sections — workflow, constraints, examples, tables. These are immune.
3. Sweep the remaining prose for the patterns in "What to remove." Cut per-pattern, not per-paragraph.
4. After each sweep, re-read. If a sentence no longer makes sense without a neighbor, the cut was too aggressive.
5. Measure the delta. A healthy minimization on a community prompt cuts 20-40% of tokens. More than 50% suggests content loss, not just filler loss.

Constraints

• Never remove content that names a specific failure mode
• Never collapse a table or code block into prose
• Preserve anti-pattern lists even if worded emphatically — the emphasis is directive signal, not filler
• Keep one concrete example per abstract principle; cut additional restatements
• Do not rewrite in a different voice; remove, do not paraphrase