Ran the skill's own workflow against skills/deslopify/SKILL.md.
Scan findings in the skill's own prose (quoted catalog patterns excluded):
- 8 negative parallelisms, including two section headings and the
subtitle line
- em-dash density 23 in 1151 words against the skill's own ~1-per-150
budget, with 4 sentences carrying two dashes
- 4 mid-paragraph bolds
- one rule-of-three anaphora
- one throat-clearing opener ("not optional and not a formality")
Rewrote by meaning per Phase 2 triage, then re-scanned. Em dashes now 2
in 1120 words, both inside quoted example patterns; remaining regex hits
are the catalog quoting itself. Cadence spans 4-44 words with no runs of
3+ sentences within +/-4. Register held to the technical-docs profile in
references/voices.md, so the "Term: definition" bullets stay.
Content, phase structure, reference links, and trigger phrases in the
frontmatter description are unchanged; the description drops its false
range ("anything between") and em dash.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015nnqhunsBsEavSZbsXjjiq
The old name hurt discoverability and made the repo awkward to link
publicly. Renames the skill directory, frontmatter name, headings, and
all doc references, and drops the vulgar trigger phrases from the
description so the word no longer appears in indexed frontmatter.
Invocation is now /deslopify; "deslop", "de-slop this", "remove the AI
tells", and "humanize this" still trigger the skill.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YM6LL3db5t2gjNGGPsU4dt
New skill that counters the structural failure where agents decompose the
noun of a request (wiki -> ingestion, search, LLM) and the experience of
using the thing never appears in the decomposition, so it never becomes a
task. The skill decomposes the scene instead: a concrete magic moment,
walked step by step, with the residue specced at plumbing fidelity.
- skills/last-20-percent/SKILL.md: five-phase workflow (magic moment,
scene-vs-noun decomposition, plumbing-fidelity specs with golden
artifacts, front-loading, the walk) plus plan and audit modes
- references/last-20-catalog.md: eight recurring categories of skipped
work with probe questions and cross-domain examples
- README.md, CLAUDE.md: register the eighth skill
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019MYgZjVEB9jkRWL6eL1Uq5
Installs a per-turn first-line canary (user name + turn counter + honest
context self-check) so attention drift, compaction, and truncation become
visible the moment they happen, with a checkpoint/re-anchor/reset trip
protocol. References file documents the research grounding: Chroma's
context-rot study, Lost in the Middle (Liu et al.), instruction-drift
findings, compaction summary loss, and Breunig's context failure taxonomy.
Updates README, CLAUDE.md, and CONTEXT.md skill rosters to six skills.
https://claude.ai/code/session_01CWPytH8cfzYfCgU6eC9pX9
New skill built around the insight that the worst AI tells (negative
parallelism / 'not X but Y', rule-of-three, puffery, uniform cadence)
are emergent generative habits, not vocabulary mistakes — so a rewriting
model reintroduces them in paraphrase. The skill therefore runs a loop:
mechanical regex scan -> rewrite at the level of meaning -> re-scan the
output until clean, then a per-genre register check.
- skills/fuck-slop/SKILL.md: workflow (target, scan, meaning-first
rewrite triage, verify loop, register check, anti-overcorrection rules)
- references/tells.md: runnable grep -E catalog of tells, tested against
slop and clean samples
- references/voices.md: register guide for academic, tweet, reddit,
LinkedIn, email, blog, marketing, docs, and application prose
- README/CONTEXT/CLAUDE.md: list the sixth skill
https://claude.ai/code/session_014C4zq4r5oQv7427bdvzqd2
Treats agent-generated plans as junior work and constructs a senior
reviewer grounded in codebase research plus web research of current
best practices. Diagnoses altitude failures (fog: vague on hard parts;
tunnel: granular with no product vision), reviews adversarially with
evidence-backed findings, and rewrites the plan into a scoped,
state-of-the-art version with product decisions surfaced as open
questions.
Includes references for the review rubric (altitude tests, vague-word
blacklist, severity calibration) and the research playbook (code and
web research tracks, source quality ranking, stop conditions).
https://claude.ai/code/session_01EXN9BXWy6933G7RNQmmUU5
Rename the `ui-craft` skill to `interface-kit`: move/rename the skill directory and reference files (skills/ui-craft → skills/interface-kit), update SKILL.md (name and heading), and update repository docs (CLAUDE.md, CONTEXT.md, README.md, docs/adr/0001-repo-shape.md) to reflect the new name and wording. Keeps priorities and content intact while standardizing naming across the repo.
Introduce the new `loop-factory` skill and its documentation. Adds skills/loop-factory/SKILL.md plus reference docs (autonomous.md, commands.md, install.md, spec-authoring.md) describing the spec-driven inbox→active→archive loop, autonomous mode, CLI commands, install steps, and spec authoring guidance. Also updates README.md, CONTEXT.md, and CLAUDE.md to list and summarize the new skill.
Julius Brussee