Ready to supercharge your AI agents?

Install any skill in seconds. Free to start, no credit card required.

Get Started Free →

Refine · sharpdeveye | DecimalAI

sharpdeveye/refine

SkillScore73Maintained

Use when the workflow works but needs polish, or as the final step in a diagnose → fix → refine cycle before shipping.

agent-toolingby sharpdeveye

Links this skill to your agents and keeps it current with the author’s updates.

Automatic

AI Engineer pack

1 · One-time — install the CLI (docs ↗)

pip install decimalai

2 · Get the skill — writes .claude/skills/sharpdeveye-refine/SKILL.md

decimalai skills pull sharpdeveye-refine --out .claude/skills/

No CLI? Get the same file — curl · raw file

curl -fsSL --create-dirs -o .claude/skills/sharpdeveye-refine/SKILL.md https://app.decimal.ai/s/sharpdeveye-refine/SKILL.md

Raw file: /s/sharpdeveye-refine/SKILL.md ↗ · machine-readable summary at /s/sharpdeveye-refine.json ↗

In productionGlobal

how this skill behaves on real traffic · last 30 days

No production data for this skill yet — usage numbers appear here once agents start running it.

Anonymized aggregate across every org using this skill — proof from real production use, not a demo.

Does this skill help? · Benchmark: with vs. withoutsign in to run this on your agents

22 test cases from the publisher’s eval.yaml · each run twice on gemini-3.6-flash — with the skill injected, then without · last run 32d ago

—+7% tokens0% turns73% of whole cases pass with skill (16/22)

Test case	Without → With	Effect	Δ tokens	Δ turns
case-06	✗→✓	▲ Improved	14%	0%
case-02	✗→✓	▲ Improved	-16%	0%
case-09	✗→✓	▲ Improved	-24%	0%
case-15	✗→✓	▲ Improved	20%	0%
case-19	✗→✓	▲ Improved	14%	0%

Showing 5 of 22 cases

SkillSafety

✓ 3 of 3 passed

Code scan15 of 15 checks clean

0 critical · 0 warning · 0 info — scanner v4 — 8/3/2026

live_secretremote_code_executionreverse_shelldata_exfiltrationssrf_metadataagent_snoopinganti_refusaldestructive_commandsobfuscated_unicodeunicode_homoglyphopaque_blobsuspicious_urltool_overreachtrigger_abuseprompt_injection_phrasing

IntentSafe

LLM judge · does the skill do what it claims

ContentClean

LLM judge · policy & harmful-content review

Source

Imported

sharpdeveye / maestro

source/skills/refine/SKILL.md

View on GitHub ↗

License MITImported 8/3/2026Version v1

Skill contents

SKILL.md

MANDATORY PREPARATION

Invoke /agent-workflow — it contains workflow principles, anti-patterns, and the Context Gathering Protocol. Follow the protocol before proceeding — if no workflow context exists yet, you MUST run /teach-maestro first.

This is the final quality pass. The workflow works — now make it excellent.

Refinement Checklist

Prompts

] Every prompt follows the 4-zone pattern (role, context, instructions, output)
] Output schemas are explicit and validated
] Negative instructions clarify what NOT to do
] No contradictory instructions
] Few-shot examples included for ambiguous tasks
] Chain-of-thought used for multi-step reasoning tasks

Tool Descriptions

] Every tool has a multi-line description: what, when to use, when NOT to use, returns
] Input parameters have descriptions and types
] Error responses are documented
] At least one example input/output in the description

Error Messages

] Error messages are specific (not "an error occurred")
] Error messages suggest corrective action
] Errors include context (what was being attempted)
] Errors are structured (code + message + details)

Logging

] Every model call is logged (input tokens, output tokens, latency, cost)
] Tool calls are logged with inputs and outputs
] Errors are logged with full context
] PII is redacted from logs
] Workflow ID traces through all log entries

Configuration

] All magic numbers are named constants
] Environment-specific values are in config, not code
] Defaults are sensible — config is for overrides
] Cost ceilings are set
] Timeout values are set for all external calls

Output

For each checklist item that fails, provide:

What's wrong (specific finding)
Where it is (file, line, or component)
How to fix it (concrete suggestion)
Priority (critical / important / nice-to-have)

Priority Matrix

| Priority | Criteria | Maestro Action | |----------|---------|----------------| | Critical | Affects correctness or safety | /fortify or /guard before shipping | | Important | Affects quality or maintainability | /calibrate in current cycle | | Nice-to-have | Cosmetic or minor inconsistency | Note for next /refine pass |

Recommended Next Step

After refinement is complete, run /evaluate to verify the polished workflow against realistic scenarios.

NEVER:

Skip the checklist — go through every item
Mark items as passing without checking
Suggest changes that alter behavior (this is polish, not redesign)
Refine before the workflow is functionally correct (fix first, refine last)

Illustrative Want to see what sharpdeveye/refine would put in front of your model? Preview the rendered body and check its trigger phrases — no LLM call, no signup.

Latest benchmark · publisher runVerified

8/21/2026, 04:08 UTC · 22 cases

Measured on 17 comparable cases — below the 22 needed to publish a lift, so this result is shown here but not on the registry card or leaderboard.

Lift

+14%

73% vs 59%

Agent turns

0%

1.0 vs 1.0

Tokens

+7%

2,071.06 vs 1,929

fail → pass · 7pass → pass · 9fail → fail · 2pass → fail · 4

Per-case results (22)

Case	Status	Duration (ms)			Turns			Tokens			Tool calls
Case	Status	Without	With	Δ	Without	With	Δ	Without	With	Δ	Without	With	Δ
case-22 We want to completely rewrite our agent's decision-making state machine in `agents/router.ts` to use a different business logic workflow. Is changing core functional behavior part of a refinement pass?	pass→pass	9,578	7,656	-20%	1	1	0%	1,458	1,840	+26%	0	0	—
case-06 In `services/api_client.py`, when a call fails we throw `Error('Failed')`. We want to polish error responses. What attributes should structured errors include?	fail→pass	12,723	10,970	-14%	1	1	0%	2,342	2,670	+14%	0	0	—
case-01 Our AI agent implementation is working functionally, but we need to do a thorough refinement check on the prompt structures, tool documentation, error messages, logging details, and configuration settings. Can you review our codebase, identify any quality gaps across these areas, specify where each problem is, give actionable recommendations to fix them, and categorize their priority?	fail→fail	22,445	4,463	-80%	1	1	0%	3,831	920	-76%	0	0	—
case-02 We just completed building our multi-step workflow pipeline and want a final polish review. Please audit our prompts, tool specs, error handling structures, logging coverage, and hardcoded config parameters. Highlight any missing best practices or quality defects, point out the exact component or file involved, explain how to correct each issue, and assign appropriate impact priorities to the audit findings.	fail→pass	33,187	25,968	-22%	1	1	0%	6,214	5,220	-16%	0	0	—
case-03 I need a final quality audit on this LLM workflow project before release. Please evaluate our prompt definitions, tool descriptions, structured error responses, observability/logging setups, and magic numbers/config values. For anything that doesn't meet production standards, list the specific defect, exact location, proposed remedy, and severity level for fixing.	fail→fail	22,766	5,285	-77%	1	1	0%	3,973	1,023	-74%	0	0	—
case-04 We have an LLM prompt template in `src/prompts/summarizer.ts` that defines system instructions and output rules. Perform a refinement pass on it. Should we add explicit role, context, instruction, and output formatting sections?	pass→fail	12,872	5,218	-59%	1	1	0%	2,193	888	-60%	0	0	—
case-05 Our custom tool `db_search` in `tools/search.py` has a brief single-sentence docstring: 'Searches database for queries.' How should we expand this docstring during our polish pass?	pass→pass	13,928	10,606	-24%	1	1	0%	2,210	2,539	+15%	0	0	—
case-07 We are adding structured logging to `llm_orchestrator.ts`. We currently log raw user input strings and completion tokens. What telemetry and security requirements should be verified during refinement?	pass→pass	16,638	11,655	-30%	1	1	0%	2,724	2,669	-2%	0	0	—
case-08 In `agent_runner.py`, we have inline numbers like `30` for timeouts, `0.02` for cost limits, and `3` for retries. What is the refinement standard for these values?	pass→pass	12,300	8,159	-34%	1	1	0%	1,980	1,825	-8%	0	0	—
case-09 During audit of `guardrails/checker.ts`, we found an unhandled prompt injection vulnerability that allows user input to bypass safety checks. What priority level should this receive and what action command should be taken?	fail→pass	10,375	11,210	+8%	1	1	0%	1,565	1,184	-24%	0	0	—
case-10 We noticed that `tools/weather.ts` lacks input parameter type descriptions, making model parameter extraction inconsistent. What priority level should this receive and what action command applies?	pass→pass	5,965	6,790	+14%	1	1	0%	922	1,624	+76%	0	0	—
case-11 We found minor inconsistent casing in log messages inside `utils/logger.py`. How should this finding be categorized and handled?	pass→pass	10,202	5,389	-47%	1	1	0%	1,523	1,404	-8%	0	0	—
case-12 Our agent prompt in `agents/writer.ts` describes desired formatting but users keep getting conversational filler like 'Sure, here is your article:'. How should negative constraints be refined?	pass→pass	13,188	25,768	+95%	1	1	0%	1,962	3,057	+56%	0	0	—
case-13 We have an ambiguous extraction prompt in `extractors/entity.py` that fails on edge cases. What should be added to improve extraction accuracy without changing logic?	pass→pass	10,957	11,180	+2%	1	1	0%	1,753	2,306	+32%	0	0	—
case-14 Our reasoning pipeline in `chains/reasoner.py` performs multi-step mathematical evaluation in a single output step, leading to errors. What refinement rule applies here?	pass→fail	8,736	2,898	-67%	1	1	0%	1,420	811	-43%	0	0	—
case-15 When auditing model invocation logs in `observability/tracer.py`, which specific token and performance metrics must be recorded for every model call?	fail→pass	8,296	5,799	-30%	1	1	0%	1,391	1,671	+20%	0	0	—
case-16 In `prompts/classifier.ts`, the model is asked to return JSON but no explicit JSON Schema is defined. What refinement requirement applies?	pass→fail	6,526	1,968	-70%	1	1	0%	952	902	-5%	0	0	—
case-17 Our new workflow pipeline in `workflows/pipeline.ts` currently throws runtime exceptions and fails fundamental unit tests. Should we run a full refinement audit now?	pass→fail	7,182	4,700	-35%	1	1	0%	1,063	1,133	+7%	0	0	—
case-18 While reviewing `agents/planner.ts` during refinement, should we redesign the routing state machine to use a different algorithm that changes workflow behavior?	pass→pass	9,296	4,449	-52%	1	1	0%	1,457	1,360	-7%	0	0	—
case-19 Can we skip checking error messages and configuration settings during this quality pass if we only care about prompt formatting?	fail→pass	10,926	10,127	-7%	1	1	0%	1,502	1,711	+14%	0	0	—
case-20 We are planning a new agent workflow project and have written high-level architectural notes in `docs/architecture.md`, but no executable code exists yet. Should we perform a refinement pass on these planning notes?	fail→pass	12,748	7,731	-39%	1	1	0%	1,892	1,827	-3%	0	0	—
case-21 Our agent script `services/runner.ts` is failing unit tests due to an unhandled exception on API timeout. Should we execute the workflow refinement quality pass to resolve this unit test failure?	fail→pass	5,649	4,884	-14%	1	1	0%	946	1,399	+48%	0	0	—

These cases were also screened by the skill’s intent / safety pass.See the Trust review ↗Reproduce this benchmark in Colab ↗

How this skill was measured

DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 17 counted toward the lift figure. The other 5 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +14 percentage points is the difference between those two pass rates over the 17 comparable cases. 4 cases got worse with the skill loaded, and they are included in that figure.

Model measured: gemini-3.6-flash
Graded by: gemini-3.6-flash
Grading method: judged
Verification: verified
Pass rate without the skill: 59.1%
Pass rate with the skill: 72.7%
Cases attempted: 22
Cases counted toward lift: 17
Cases that regressed: 4
Skill version measured: v1
Run completed: 8/21/2026, 04:08 UTC

Measured example — same prompt, with vs. without the skill · from the benchmark run · case case-06

Prompt

In `services/api_client.py`, when a call fails we throw `Error('Failed')`. We want to polish error responses. What attributes should structured errors include?

Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.

Show both answers

SkillScore blends 4 signals1 of 4 complete

BenchmarkPass rate WITH skill on a verified run73

Live evalNeeds ≥10 evaluated traces in 30 days

AI ratingNeeds ≥10 rated traces

AdoptionNeeds ≥10 serves across ≥2 orgs in 30 days

Efficiency · separate axis (at equal correctness)+0% turns+17% tokens

Shown for transparency; not credited unless a verified run shows no correctness regression.

Backed by 1 of 4 signals — each added signal makes the score more robust.

Verified on gemini-3.6-flash · re-tested 31 days ago

Other measured skills in the registry, with their headline benchmark lift.