CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

workflow-from-eval

Build or improve a Roboflow Workflow using eval-driven search. Prefer this for open-ended business problems where the approach is unclear, or when comparing alternatives. A business goal plus images or video is enough to start; no existing Workflow or Eval is required. Build an example suite and scorecard the user approves, compare candidate Workflows, then tune and validate the winner. Use direct tools for straightforward builds and edits, Workflow Evals tools for measuring one existing Workflow once, and training skills for training a model.

Quality

90%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Low

Low-risk findings worth noting

SKILL.md
Quality
Evals
Security

Quality

Content

85%Weight 40%Scale 1-5

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

The body is an unusually tight, highly actionable orchestration protocol: exact tool names, file layouts, budgets, retry and validation loops, and explicit user-facing checkpoints, with essentially no filler. Its one material defect is that the four guide files it delegates phases to (intake.md, ground-truth.md, search.md, gotchas.md) are absent from this bundle, so the progressive-disclosure structure is not actually navigable as shipped.

Suggestions

Ship the four referenced guides (intake.md, ground-truth.md, search.md, gotchas.md) in the bundle — Phases 1, 3, and 4 currently instruct "Follow <guide>.md" for content that is nowhere in the package.

Provide a fallback for the missing guides the way the draw-vision reference does ("If unavailable, use simple captioned media"), or inline the minimum needed from each guide into SKILL.md.

Trim the duplicated cost-announcement guidance — "say what it is and what it costs" appears in both the "Talk like this" section and "Budget and cost".

DimensionReasoningScore

Conciseness

The body is dense and operational throughout — named tools ("workflow_specs_validate, then workflow_specs_run on three varied examples"), exact naming patterns ("<slug>-spec, <slug>-suite, <slug>-eval"), and numeric budgets with almost no concept-explanation padding. It is not 5 because there is minor repetition and trimming opportunity (e.g., "say what it is and what it costs" appears in both "Talk like this" and "Budget and cost", and a few multi-clause passages could be tightened); it is clearly above 3 since nearly every line carries non-obvious operational content.

4 / 5

Actionability

Guidance is fully concrete for an instruction-only skill: exact tool calls and their caveats ("Delete by url slug (workflows_delete takes the slug)"), file structures ("created.json, answer-key-draft.json, cases/all.json, candidates/*.json and rounds/*/summary.json"), specific budgets ("20 to 40 candidates", "A Run accepts up to 100 Executions"), retry policies, and even user-facing phrasing examples ("round 2: 14 candidates × 52 examples"). This matches anchor 5 — specific, executable guidance covering the common cases.

5 / 5

Workflow Clarity

Five phases with four explicitly defined checkpoints ("A checkpoint is a message where you show something and wait. Never spend inference past a checkpoint before the user answers") and validation woven in: validate-then-run in Phase 2, smoke before save in Phase 4, noise measurement before ranking, a finalists run with "execution_reuse: never", and confirmation-gated soft deletion at Checkpoint 4. Feedback loops are explicit ("Retry failed Cases up to twice", "If a joint Run is rejected, split it in halves and retry once; then fall back to smaller Runs"), matching anchor 5.

5 / 5

Progressive Disclosure

The references themselves are well-signaled and one level deep ("Guides in this skill: intake.md… ground-truth.md… search.md… gotchas.md" with a one-line description each), but none of these files exist in the actual bundle — no references/ directory is present — and the protocol's phases depend on them ("Follow intake.md", "Follow ground-truth.md", "Follow search.md"). Scored against the real bundle structure, the split is not realized; this fits anchor 3 (structure present, but the organization does not actually work as written) rather than 4, where content would be appropriately placed with only minor gaps.

3 / 5

Total

17

/

20

Passed

Description

92%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 strong description: it states concrete capabilities across the full build/compare/tune/validate lifecycle, gives an explicit "Prefer this for…" trigger clause, and cleanly separates itself from adjacent direct-tool and training skills. The only weakness is modest synonym coverage for natural trigger phrases. It is written in third person and stays concise for its scope.

DimensionReasoningScore

Specificity

Multiple concrete actions are listed with comprehensive coverage: "Build an example suite and scorecard the user approves, compare candidate Workflows, then tune and validate the winner", plus build/improve and explicit routing ("Use direct tools for straightforward builds and edits, Workflow Evals tools for measuring one existing Workflow once, and training skills for training a model"). It exceeds anchor 4 because the actions span the full lifecycle rather than leaving minor coverage gaps.

5 / 5

Completeness

Both halves are explicit: what — "Build or improve a Roboflow Workflow using eval-driven search… Build an example suite and scorecard… compare candidate Workflows, then tune and validate the winner"; when — "Prefer this for open-ended business problems where the approach is unclear, or when comparing alternatives", reinforced by the contrast with direct tools and training skills. The when-clause is explicit with concrete trigger conditions, matching anchor 5 rather than anchor 4 where the when could be more specific.

5 / 5

Trigger Term Quality

Good natural-term coverage: "open-ended business problems where the approach is unclear", "when comparing alternatives", "images or video", "training a model" — phrases users would plausibly say. Not 5 because common variations users might use ("benchmark", "compare models", "which approach is best", "test my workflow") are absent; not 3 because several relevant natural phrases are present, not just one or two generic keywords.

4 / 5

Distinctiveness Conflict Risk

"Roboflow Workflow using eval-driven search" carves out a clear niche, and the description explicitly disambiguates from neighboring skills ("Use direct tools for straightforward builds and edits, Workflow Evals tools for measuring one existing Workflow once, and training skills for training a model"), minimizing wrong-skill triggering. Not 4: there is no meaningful overlap risk with closely related skills.

5 / 5

Total

19

/

20

Passed

Validation

93%

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

CriteriaDescriptionResult

relative_links

Relative link issues: 1 suspicious

Warning

Total

15

/

16

Passed

Repository

roboflow/computer-vision-skills

Path

skills/workflow-from-eval/SKILL.md

Reviewed

Table of Contents

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.