Why did 43 of my skills never fire?

Answer

Hermes shows the model only the first 60 characters of a skill description. Forty-three of my 176 skills put their trigger words past that cut, so they competed for attention with a description the model never finished reading. Rewriting all 43 to fit the window — and re-parsing each with Hermes's own frontmatter parser — moved the library from 43 over-window descriptions to 0.

The window is SKILL_PROMPT_DESC_LIMIT = 60 in agent/skill_utils.py, not the 500 characters an imported skill asserted.
The worst case was impeccable at 895 characters; the model saw "Use when the user wants to design, redesign, shape, critique…" and nothing else.
A rewrite only counts if it is measured again: the audit tool now reports the over-window count directly.
12 FAILs remain, all content debt I deliberately did not touch in this pass.

Applies todeepseek seatM4176 catalog skillsexcludes .archive2026-11-07

How far past the cut did the worst ones run?

Skill	Description length	What the model actually saw
`impeccable`	15× the window895 chars	"Use when the user wants to design, redesign, shape, critique…"
`scroll-craft`	15×888 chars	"Build premium scroll-driven landing pages for service…"
`ai-influencer-lifecycle`	12×745 chars	"Create and manage a consistent AI influencer using…"
`archify`	12×740 chars	"Create validated architecture, workflow, sequence…"
`ssh-monitor-dashboard`	10×588 chars	"Build a monitoring dashboard that polls a remote machine…"
`done-before`	9×532 chars	"Query the one brain (Wiki_test vault) BEFORE starting…"

A description is a trigger, not a summary: its job is to win the moment of choice. Everything past character 60 is invisible at that moment, however well written it is.

What did the pass change?

Criterion	Before	After
Descriptions over the 60-char window	invisible43	all visible0
Longest description	895 chars	59 chars
Description collisions	competing2	1
WARN total	55	27
FAIL total	12	untouched axis12
Eval suite	—	passing13/13

Every original description is preserved verbatim under a ## Trigger surface heading inside the same file, so the wording was moved, not deleted.

What is the loop that produced it?

One axis per iteration

Diagram structure

Snapshot the library snap
Triage FAIL, then WARN triage
Re-run the analyzer again
Analyze every SKILL.md scan
Rewrite one axis fix
Record the delta log
again → log · moved?
fix → again
log → triage · next axis
scan → triage · exit 1 on FAIL
snap → scan
triage → fix · worst first

The loop is the whole method: snapshot, analyze, fix one axis, re-measure, record. Bundling a description rewrite with a body restructure destroys the only evidence you have about which change helped.

How do the issue counts compare?

Skill library issues before and after the pass

WARN fell by 28 because 43 rewritten descriptions removed 43 over-window warnings while the collision fix removed one more. FAIL stayed at 12: those are dead links and name mismatches, a different axis.

What is still broken?

Item	Count	What it needs
Dead links in skill bodies	factual15 across 10 skills	write the missing reference, or delete the pointer
`name` not matching directory	factual2 (`audiocraft`, `segment-anything`)	rename the directory, or align the field
References over 300 lines with no TOC	judgement51	add a contents block where the file is genuinely long
Bodies over 500 lines	judgement19	move depth into references behind a trigger
Duplicate skill pair	judgement1 (`cape-town` vs `civic-recourse`)	merge, or give each an exclusion clause

Fix the 15 dead links — each is either a missing file or a stale pointer.
Resolve the 2 name/directory mismatches before any hermes skills update.
Decide whether the duplicate recourse pair merges or splits.

What does the fix look like in practice?

+1 Added · −3 Removed

Old	New	Change
		`--- appledesign/SKILL.md`
		`+++ appledesign/SKILL.md`
		`-description: "Apple's approach to interface design and fluid, physical motion,`
		`- translated for the web. Use when building or reviewing a marketing site,`
		`- a product surface, or any interaction where the motion should feel native."`
		`+description: "Design interfaces in Apple's restrained, fluid idiom."`

The rewrite is hand-written, not truncated. Mechanical capping produced fragments like "…polls a remote machine." — grammatical debris that signals carelessness to a human reviewer and drops the verb that made the trigger work.