Tessl • The management layer for AI agent skills

An exemplarily dense, expert-facing skill body: exact formulas, numeric invariants, a file map, and an explicit validation checklist with zero padding and clearly signaled cross-skill references. The only soft spot is that the change→validate workflow is implied by section order rather than given as an ordered sequence with a re-validate loop.

Dimension	Reasoning	Score
Conciseness	The 39-line body is lean and assumes Claude's competence: it never explains kernels, softplus, or triangular solves, and every sentence carries non-derivable project specifics such as "inspect the affected entry point's signature before changing its implementation" and "The safe intra path uses 16-token diagonal blocks and midpoint offsets before exponentiation". No padding to trim.	5 / 5
Actionability	Guidance is fully concrete: exact activation formulas ("-exp(A_log) * softplus(g + dt_bias)", "lower_bound * sigmoid(exp(A_log) * (g + dt_bias))"), exact bounds ("-5 <= lower_bound < 0", "tests cover [-5, 0]"), named files, and precise directives ("Preserve these offsets rather than exponentiating the full cumulative gate", "Test through the public entry point") covering both gate-mode cases and safe/non-safe paths. Per the rubric's code-vs-instruction note, the absence of code is not penalized for this instruction-only skill.	5 / 5
Workflow Clarity	A coherent change-then-validate flow is present with an explicit case-selection checklist ("Select the KDA cases affected by the change:" plus six categories) and edge-case guidance ("include the lower-bound limits, saturated gate inputs, extreme gate scales..."), but the sequence is distributed across topical sections rather than explicitly ordered, and there is no test-run command or fix-and-revalidate loop. Not 5: checkpoints, while detailed, are not sequenced with feedback loops; not 3: validation is explicit and thorough, not implicit.	4 / 5
Progressive Disclosure	No bundle files exist (references/, scripts/, assets/ are absent), and the under-50-line body is organized into clear sections (Code map, Gate contracts, Numerical invariants, Validation) with well-signaled one-level-deep cross-links ("[backend dispatch](../fla-dispatch-backends/SKILL.md)", "[correctness coverage](../fla-correctness-coverage/SKILL.md)", AGENTS.md). Per the rubric guideline, a skill under 50 lines with well-organized sections scores 5.	5 / 5
	Total	19 / 20 Passed

Description

66%Weight 40%Scale 1-5

Based on the skill's description, can an agent find and select it at the right time? Clear, specific descriptions lead to better discovery.

A tightly scoped, domain-precise description that names four concrete component areas with minimal fluff. Its main weakness is the missing explicit "when to use" trigger clause, which caps completeness, and the absence of entry-point names that users would most naturally mention.

Suggestions

Add an explicit trigger clause, e.g. "Use when changing files under fla/ops/kda/, updating KDA gate or decay math, or touching chunk/recurrent kernel backends."

Include the public entry point names (chunk_kda, fused_recurrent_kda) and the test path (tests/ops/test_kda.py) so users mentioning those functions naturally match this skill.

Scope "backend routes" to KDA (e.g. "KDA backend routes") to reduce overlap with the sibling fla-dispatch-backends skill.

Dimension	Reasoning	Score
Specificity	"Modify or review" applied to four enumerated areas — "KDA gates", "chunk and recurrent kernels", "backend routes", and "their correctness coverage" — gives several specific targets with near-complete scope coverage. It falls short of 5 because the verbs themselves are generic (two actions) rather than the fully enumerated concrete action lists of the top anchor.	4 / 5
Completeness	The "what" is clear (modify or review KDA gates, kernels, backend routes, coverage), but there is no "Use when..." clause or equivalent explicit trigger guidance, which caps completeness at 3 per the judging guidelines. It is not 4 because the "when" is entirely absent rather than merely imprecise.	3 / 5
Trigger Term Quality	"KDA", "gates", "chunk and recurrent kernels", and "backend routes" are the natural phrases a developer in this codebase would say. A few natural terms are missing: the entry-point names (chunk_kda, fused_recurrent_kda), the fla/ops/kda path, and synonyms such as the gated-delta expansion.	4 / 5
Distinctiveness Conflict Risk	"KDA gates" pins a clear niche, so confusion with unrelated skills is unlikely. Minor overlap risk remains with closely related fla skills: "backend routes" and "correctness coverage" are shared vocabulary with the sibling fla-dispatch-backends and correctness-coverage skills referenced in the body.	4 / 5
	Total	15 / 20 Passed

Validation

93%

Warnings & errors only

Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.

Validation — 15 / 16 Passed

Validation for skill structure

Criteria	Description	Result
relative_links	Relative link issues: 3 suspicious	Warning

	Total	15 / 16 Passed

Repository: fla-org/flash-linear-attention
Path: .agents/skills/fla-kda/SKILL.md
Revision: 9819311

Reviewed: 1 month ago

Discovery Implementation Validation

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.

fla-kda

Table of Contents

Is this your skill?