prompt-terser

What this skill does

# Prompt Terser

**Terser wording improves adherence.** For the same thought, fewer or clearer words usually produces more reliable model
behavior — less attention competition, smaller constraint surface, less noise in the instruction layer. The goal is not
to save tokens; the goal is to make the prompt work better. Token reduction is a side effect, not the target.

Prompts drift toward verbose wording across edit cycles. "Please make sure to" creeps in, rationale paragraphs stack,
hedges accumulate. This skill audits an existing prompt and proposes terser wording for the same semantic content —
every meaning preserved, every restatement compressed. Output is a diff-proposal with per-cut justification.

## When to invoke

- A skill, system prompt, output style, or agent prompt that has been edited many times
- Before shipping a skill update, as an adherence quality gate
- When a prompt feels verbose or repetitive on read
- When adherence is unreliable and verbose wording may be competing for attention

**Do not invoke for:**

- Newly authored prompts (use `prompt-engineering` — this skill is for drift, not authoring)
- One-shot user prompts (these don't drift)
- Prompts you intend to redesign — this skill preserves meaning, it does not change it
- Pure style or visual consistency refactors — terser is wording reduction for the same content, not format unification
- Token-budget squeezing that allows meaning loss — that's compression, a different operation

## Workflow

Three phases, ordered cheap-to-expensive:

1. **Wording pass** — mechanical substitution, near-zero risk
2. **Format pass** — mechanical whitespace and structure cleanup
3. **Structural pass** — drift-pattern detection with falsification gate

Each phase produces diff-proposal entries. Do not auto-apply structural cuts.

## Phase 1 — Wording pass (mechanical)

Apply these substitutions to the entire prompt body. Rewrite rules, not judgment calls.

- "in order to" / "so as to" → "to"
- "utilize" / "make use of" → "use"
- "implement a solution for" → "fix" / "solve"
- "It is worth noting that" / "It should be noted" → drop entirely
- "Please make sure to" / "Be sure to" / "Ensure to" → imperative verb
- "I want you to" / "You should" / "You need to" → imperative verb
- "might potentially" / "may possibly" → direct claim
- "very important" / "extremely critical" → "critical" or drop
- "the reason is that" / "due to the fact that" → "because"
- "extensive" / "comprehensive" / "thorough" → "full" or drop
- "however" / "furthermore" / "additionally" → drop or em-dash
- "in conclusion" / "to summarize" → drop
- Hedging adjectives ("appropriate", "reasonable") → concrete spec or drop
- Passive voice where active is clearer → active
- Articles where clarity survives → drop
- "A lot of" / "a number of" → "many" or specific count
- "the majority of" → "most"
- "be able to" → drop (where grammar holds)
- "absolutely" + modal verb ("absolutely must", "absolutely cannot") → drop "absolutely"
- Redundant "that" clauses → drop "that" where grammar holds
- "Remember that X" / "Note that X" / "Be aware that X" → state X directly (self-referential framing shifts attention to
  the reminder rather than the content)

### Emphasis exception

- **Keep emphasis** in critical-rule blocks at the top 20% or bottom 20% of the prompt (U-curve recency/primacy zones)
  and in safety guardrails — e.g., **"You must NEVER X"** rather than **"Never X"**
- **Keep emphasis** on rule-prefix bolds in bullet lists (`- **Lead with function.** ...`) and KV-list labels — these
  function as keys, not prose emphasis, regardless of zone
- **Strip emphasis** from descriptive prose, rationale paragraphs, and middle-zone content

## Phase 2 — Format pass (mechanical)

- Strip decorative blank lines between adjacent bullets of the same list
- Remove decorative separators (`---` used as section padding rather than as a real boundary)
- Collapse multi-line bullets that fit on one line without losing meaning
- Remove `**bold**` emphasis on terms that already appear as headings
- Convert tables to KV lists when entries are independent — apply the table test: if removing a column would lose
  comparative meaning, keep as table; otherwise convert
- Strip nested sub-bullets used only for visual breathing room (not for hierarchy)

## Phase 3 — Structural pass (falsification-gated)

Scan the prompt for the drift patterns below. For each candidate cut, apply the falsification gate before recommending.

### Two principles before scanning

**The U-curve.** Constraint compliance follows a U-shape against length: peaks at extreme compression (≤10 words), dips
in the ~20–40 word "ambiguity zone," then rises again for structured rubrics (≥150 words). The middle is worst —
medium-length narrative paragraphs degrade adherence even relative to either extreme. Every cut should push content out
of the ambiguity zone, not into it. Compress to <10-word imperative OR commit to a structured rubric; do not stop in the
middle.

**Narrative vs structural verbosity.** Not all verbosity is drift. Distinguish two types:

- **Narrative verbosity** — rationale paragraphs, background descriptions, "we believe / because past experience
  shows...". Low instructional density. Often ignored; sometimes triggers hallucinations. **Bloat — candidate for cut.**
- **Structural verbosity** — rubrics, checklists, decision tables, multi-step procedures with explicit constraints. High
  instructional density. Externalized memory the model can reference during generation. **Load-bearing — keep even when
  verbose.**

The test: does each token act as an active constraint, or as descriptive background? Active → structural. Descriptive →
narrative.

### Drift patterns to detect

- **Layered additions** — rules tacked on with "Also," / "And remember," / "One more thing" without merging into
  existing rule blocks. Sign: a rule block has rules that don't share a parent topic.
- **Rationale stacking** — inline "why" paragraphs added for each rule. If the rule is self-evident or its rationale
  lives in a reference, the inline rationale is bloat.
- **Style inconsistency that enables terser unification** — same content stated multiple times at different times in
  different formats. Flag ONLY when unification eliminates restated content (3 statements → 1 statement). Pure visual
  reformatting without word reduction is a style refactor, not a terser concern.
- **Duplicated constraints** — same rule restated in different sections. Sign: searching for the rule's key noun finds
  multiple imperatives saying the same thing.
- **Vestigial scaffolding** — examples or anti-patterns that no longer match the rules they were written for. Sign: an
  example references a term, behavior, or rule that no longer exists.
- **Persona / philosophy creep** — bookend "we believe" / "we value" / "the goal is" paragraphs grown over iterations.
  Sign: philosophy is >5% of total prompt length and contains 3+ such sentences.
- **Calibration prose** — paragraphs that explain to a human author why the rule exists, addressed to the reader rather
  than the model. Sign: the prose uses second-person addressing the human ("you might wonder why...").
- **Defensive hedging** — "in most cases" / "generally" / "as a rule of thumb" softening rules that should be hard.
  Sign: the surrounding context shows the rule is enforced strictly.
- **Ambiguity-zone paragraphs** — narrative paragraphs of ~20–40 words explaining a rule. Worst-case length for
  adherence. Either compress to a <10-word imperative or expand into a structured rubric/checklist. Sign: rule has
  one-paragraph rationale that isn't a checklist and doesn't fit on a single line.

### Falsification gate

Every structural cut must pass three checks, in order. State all three in the diff-proposal.

**1. Verbosity type:** is the content **structural** or **narrative**?

- **Structural** (rubric, checklist, decision table, multi-step procedure with explicit constra