# Coordinate Evidence-Driven Agent Workflows

Complex agent tasks often begin before scope, risk, and evidence requirements are clear. This skill adds gates, role discipline, verification, and explicit acceptance.

## Install

```bash
npx skillstore add jadeyingwah/gpt-series-reasoning-style
```

## Metadata

- Status: approved
- Slug: jadeyingwah-gpt-series-reasoning-style
- Version: 1.2.0
- Author version: 1.2.0
- Skillstore revision: r2
- Version status: valid
- Tree hash: 17bfc133d5acc2a15b45dd61d802356415fe6fc751b7ca9972bee164743123a3
- Author: jadeyingwah
- GitHub username: jadeyingwah
- License: MIT
- Repository: https://github.com/jadeyingwah/gpt-series-reasoning-style/tree/b6f054551a19b10b75cac58e51b295af2f7a6556/
- Ref: 5526951beaeca22519572a471618555839b53b05
- Supported tools: Claude, Codex, Claude Code
- Audit status: complete
- Agent install advisory: confirmation\_required
- Manual install advisory: allowed
- Artifact signature: available
- Audit attestation: unavailable
- Human verification: not\_verified
- Risk factors: network, external\_commands, filesystem
- Quality score: 76
- Public page: https://skillstore.pages.dev/skills/jadeyingwah-gpt-series-reasoning-style
- Manifest: https://skillstore.pages.dev/api/skills/jadeyingwah-gpt-series-reasoning-style/manifest

## Capabilities

- Classifies tasks by risk and defines a pre-implementation confirmation gate.
- Surveys installed skills, reusable templates, existing implementations, and reference sources.
- Supports single-agent, subagent, and commander coordination modes.
- Defines task packages, ownership, trust tiers, evidence records, and closure rules.
- Requires fresh verification and marks untested claims as UNVERIFIED.
- Provides optional scripts for self-checks, claim checks, probes, and mutation testing.

## Use Cases

- Structure a complex implementation: Define scope, risks, resources, verification, and acceptance before changing a codebase.
- Coordinate independent agents: Assign roles, package tasks, track evidence, and keep one accountable owner.
- Audit completion claims: Compare stated results with fresh tests, files, screenshots, and explicit coverage.

## Prompt Templates

### Plan a focused task

```
Use this skill to assess my task. State the risk tier, needed resources, plan, and verification before implementation.
```

### Prepare an evidence plan

```
Define acceptance criteria for this deliverable. List fresh evidence, tested input ranges, and anything that must remain UNVERIFIED.
```

### Run a rigorous implementation

```
Apply the full workflow to this high-impact change. Track scope, risks, decisions, review findings, tests, and final acceptance evidence.
```

### Coordinate multiple agents

```
Design a commander workflow for this task. Define roles, task packages, trust tiers, handoffs, evidence requirements, and final DRI closure.
```

## Limitations

- It improves execution discipline but does not increase model reasoning capability.
- Its primary rules and most supporting documents are Chinese-first.
- Full workflows add overhead to small or time-sensitive tasks.
- The claim checker executes reviewed commands and is not a security sandbox.

## Best Practices

- Use the light workflow only for specific, reversible, low-impact tasks.
- Attach fresh evidence to every completion claim and label untested areas UNVERIFIED.
- Keep one DRI accountable even when several agents contribute.

## Anti Patterns

- Do not treat a declared plan as permission to implement.
- Do not report test counts without naming the covered and uncovered input ranges.
- Do not accept another agent's completion statement without checking its evidence.

## Security Audit

- Audited at: 2026-09-12T11:20:04.186\+00:00
- Summary: Most static alerts are false positives caused by defensive examples, readable Chinese prose, Markdown syntax, SVG paths, and documented installation locations. One high-risk behavior is confirmed: claim-check executes claims-file commands through shell=True, so hostile or insufficiently reviewed input can run arbitrary code. Static review was capped at 400/554 representative findings; omitted static matches are unconfirmed, so automatic publishing stays disabled until manual review.

## Stats

- Views: 2
- Downloads: 1
- Favorites: 1
- Popularity score: 0