no-slop

Files (3)

• SKILL.md
• examples.md · references/
• slop-taxonomy.md · references/

SKILL.md (first lines)

# no-slop — dense prose, nothing lost

AI prose fails in two directions at once: it is verbose (filler, restatement, decoration) and thin (claims without specifics). This skill fixes both. The output contract: **every sentence carries new information, and no information is lost.**

This is a reusable prose-transform workflow that composes with task workflows.
Once invoked, its rules apply to user-facing prose produced for the current task
— reports, summaries, PR bodies, README sections, briefings, and explanations.
It does not change what work you do; it changes how you write the result.

## Workflow

Activate the modifier for the invocation mode below, apply the three rules while
drafting, then run both self-verification passes before returning prose.

### Invocation modes

| How it was invoked | What to do |
|---|---|
| Alongside another skill or task ("write X, no-slop") | Do the task per its own skill/instructions; apply these rules to every prose deliverable it produces |
| Standalone with existing text ("deslop this") | Rewrite the given text under these rules; show the result (and, if asked, what was cut and why) |

Skill activation is scoped to the current task. For a durable style rule across
tasks, put equivalent guidance in the applicable `AGENTS.md`.

## The three rules

### 1. Delete slop (cut without loss)

Cut anything that a reader could delete without losing information. The recognition patterns live in [`references/slop-taxonomy.md`](references/slop-taxonomy.md) — the canonical, growable list. The headline categories:

- Ritual openings and closings ("In this report, I will…", "I hope this helps")
- Restating the question or the task before answering it
- Contentless emphasis ("this is very important" with no *why*)
- Stacked hedges ("it could potentially be possible that")
- Saying the same thing twice in different words; conclusions that re-summarize the body
- Structure that outweighs substance: bullet sprawl, heading-per-sentence, bold/emoji decoration
- Explaining the obvious to the audience at hand

### 2. Densify (thin content is also slop)

Deleting filler is not enough — the remaining sentences must carry weight:

- Every claim gets its ground: a fact, number, path, example, or reason. A claim you cannot ground gets cut or marked as an open question — not padded.
- Replace vague quantities with the actual items ("several issues" → name the three).
- Replace abstract verbs with the concrete action ("handle it properly" → what happens, when).
- Prefer one precise sentence over three approximate ones.

### 3. Preserve (compression is not omission)

**Compression removes redundancy. Omission removes information. Only compression is allowed.** Never cut:

- Facts, numbers, file paths, names
- Caveats, limitations, failure conditions
- Reasons behind judgments and decisions
- Steps the reader must perform

When unsure whether something is redundancy or information, keep it. A dense text that silently dropped a caveat is worse than a verbose one that kept it — the reader can skim verbosity but cannot recover a missing warning.

## Self-verification (must-fire, two passes)

After drafting any deliverable prose, before presenting it:

1. **Density pass** — read each sentence and ask "what new information does this carry?" Cut or merge any sentence with no answer.
2. **Preservation pass** — list the facts, numbers, caveats, and reasons in your pre-cut draft (or source text, in rewrite mode); confirm each one survives in the final text. Anything missing goes back in.

Do not skip pass 2 to save time — over-cutting is this skill's primary failure mode, and it is invisible without the check.

## Calibration

- Dense ≠ terse. Write complete sentences; drop fragments and arrow chains ("A → B → fails") that force the reader to reconstruct your meaning. Readability outranks brevity.
- Match the reader: an expert needs less scaffolding, a newcomer needs more explanation — explanation for a reader who needs it is information, not slop.
- Structure (headings, tables, bullets) is allowed when it carries organization the prose can't; it is slop when it decorates content that one paragraph would state better.
- Worked before/after examples: [`references/examples.md`](references/examples.md). Read them once per session when this skill activates — they calibrate the cut line better than the rules alone.

## Final Report

Return the requested deliverable after the density and preservation passes. Do
not add a separate summary of cuts unless the user asks for one. When they do,
name what was removed or merged and why, without restating the rewritten text.

From s-hiraoku/codex-harnesses, MIT.

More from s-hiraoku/codex-harnesses