| name | prompt-writing-discipline |
| description | Author discipline for LLM-consumed prose: accurate, precise, low noise, consistent, coherent; prefer clear English names over opaque codes/abbreviations (expand first use if a house label must appear). Use when writing or editing skills, rules, prompts, agent hubs, design notes, or other text aimed at models -- not product wire formats. Triggers: prompt writing, writing discipline, low noise prose, precise wording, consistent terminology, coherent docs, expand abbreviations, clear English names, opaque tier codes, LLM-facing text, skill prose, agent hub copy. |
| metadata | {"pattern":"pipeline","domain":"doc","version":"1.4"} |
Prompt writing discipline
Author discipline for LLM-consumed text (skills, rules, prompts, agent hubs, design notes). Not a product or wire-format skill.
When to apply
Writing or editing prose that a model will read. Skip for human-only marketing fluff or proprietary serialisation dialects unless the user asks.
Seven disciplines (MUST)
| Discipline | MUST |
|---|
| Accurate | State only what is true and checkable; no invented facts or soft hedges that hide uncertainty. |
| Precise | Prefer the exact term, bound, or step; avoid vague fillers ("somehow", "various", "appropriately"). |
| Low noise | Cut throat-clearing, repetition, and ornamental asides; every sentence earns its place. |
| Consistent | One term per concept; same voice, tense, and naming across the piece. |
| Coherent | Order so later lines follow from earlier ones; no orphan claims or topic jumps. |
| Abbreviations | Prefer clear English names over opaque codes; drop needless abbreviations; expand house labels on first use (see below). |
| Positive | Name the wire structure in positive terms, then the per-item gate. |
Abbreviation rule
- Default: write the expanded, plain-English name -- not an internal code or cryptic tier.
- Keep an abbreviation only if it is standard in the domain and clearer than the expansion after first use.
- If a house label must appear (design-doc shorthand, legacy pin): expand on first use with the meaning, then prefer the plain name thereafter.
- Do not invent cryptic tiers/codes (
T1, LAW-PIPE20, unglossed house labels) as the primary agent-facing term.
- Optional example (MemNet): if a design doc says "Tier A" or "Write = display", gloss once as obsolete house labels, then prefer GQL wire (shaped pin_map + openCypher-shaped mutate).
- Prefer ASCII in LLM-consumed skill/rule/hub text (
-> not arrows; -- not em dashes; no smart quotes). See pack rule R16 in LLM.md.
- When spawning Task/subagents: set
model per sub-agent-policy / R17 (thinking/unclear -> cursor-grok-4.5-low; visual items review -> kimi-k3-max; MemNet snapshot -> gpt-5.6-luna-medium). MUSTNOT use any *-fast / FAST slug.
Good vs bad (one-liners)
| Bad | Good |
|---|
| "Leverage the util to somehow sync stuff ASAP." | "Run scripts/sync.py to copy config into the target folder." |
| "Use CFG / ENV / RT interchangeably." | "Use config for files; environment variable for process env." |
| "Follow Tier A / T1 / LAW-PIPE20." | "Follow the GQL wire (shaped pin_map + openCypher mutate); expand any house label on first use." |
| "The NFT of the SKL is TBD w.r.t. QoS." | "Skill metadata version is unset; quality targets are not defined yet." |
| Three paragraphs restating the same MUST. | One MUST line; no restatement. |
Checklist before emit