Skip to content
Back to skills

Speckit Coach

ASecurity

Coach developers through Spec-Driven Development and SpecKit Pro. Use for SDD methodology, SpecKit command and gate guidance, technical-roadmap and workflow design, roadmap-MOC guidance, checklist selection, SpecKit project repair, selective formal-methods guidance and model/checker selection, SpecKit consensus protocol and voting rules, or SpecKit preset and extension discovery and configuration. Not for running autopilot, conducting grill-me, or unrelated coding (including MCP tool implemen...

  • 5 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added September 27, 2026
ai-agentsgogit

Works with

  • cli
  • mcp

Security analysis

A100/100

Pro scans all 20 files and shows the line behind each finding

Scanned October 6, 2026

npx -y skills add racecraft-lab/racecraft-plugins-public --skill speckit-coach --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Speckit Coach?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Speckit Coach
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/racecraft-lab-speckit-coach-racecraft-plugins-public/badge)](https://www.skillsdirectory.com/skills/racecraft-lab-speckit-coach-racecraft-plugins-public)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: speckit-coach
description: Coach developers through Spec-Driven Development and SpecKit Pro. Use for SDD methodology, SpecKit command and gate guidance, technical-roadmap and workflow design, roadmap-MOC guidance, checklist selection, SpecKit project repair, selective formal-methods guidance and model/checker selection, SpecKit consensus protocol and voting rules, or SpecKit preset and extension discovery and configuration. Not for running autopilot, conducting grill-me, or unrelated coding (including MCP tool implementation). Formal-check execution and autopilot resume belong to speckit-autopilot.
---

# SpecKit Coach

Coach the user through the official SpecKit CLI and complementary SpecKit Pro workflows. This skill provides guidance and bounded project repair; it does not replace SpecKit or run an autonomous workflow itself.
In Codex, use `$skill-name` syntax for plugin skills rather than legacy custom slash-command syntax.

## Start with the project that exists

- Use `../speckit-autopilot/references/capability-discovery.md` for capability selection and `../speckit-autopilot/references/grounding.md` for external-fact grounding.
- Discover the session's available tools and skills before recommending an external action. Ground external claims in an observed result; say when the available evidence is insufficient.
- Inspect `.specify/presets/*/preset.yml` and `.specify/extensions/*/extension.yml` when present. Read the installed configuration before advising, so the advice reflects the project's actual constraints.
- Treat mutable actions as user-controlled. Explain the intended change and request confirmation before installing, removing, enabling, disabling, or configuring extensions.

## Route the request

Use the smallest relevant reference; do not load the whole library.

| User goal | Read or do |
|---|---|
| Learn SDD, start a project, recover from a failed phase, or assess artifacts | [Getting started](./references/getting-started.md) and [SDD methodology](./references/sdd-methodology.md) as needed. |
| Use an installed official Spec Kit command | [Command guide](./references/command-guide.md) for inputs, artifacts, mistakes, and recovery. Inspect the active installed command definition for version-specific mechanics. For gate questions, read the [live autopilot contract](../speckit-autopilot/SKILL.md) as a reference only; do not execute or invoke it. |
| Choose or author checklists | [Checklist domains](./references/checklist-domains-guide.md). Read the feature's `spec.md` and `plan.md`, rank the relevant risks, and offer enriched prompts for two to four domains. |
| Decide whether a behavior needs formal verification, choose a checker, or learn a first model | [Selective formal methods](./references/formal-methods-guide.md). Assess risk and simpler checks first; record explicit selection and adapt coaching to experience. |
| Model a difficult rule in Quint, or look up Quint language, modeling, or witness guidance | [Optional Quint modeling](./references/quint-guide.md). Read only the sections the model needs; SpecKit still owns selection, bounds, and gate decisions. |
| Design a constitution | [Constitution guide](./references/constitution-guide.md). |
| Create or repair `.specify/quality-gates.json`, pick gate thresholds, or record a permanent gate skip | [Quality gates guide](./references/quality-gates-guide.md). Measure the existing code before proposing a ceiling; write the file only after the operator confirms. |
| Design a multi-spec roadmap or split a feature | Read the complete [technical-roadmap template](./templates/technical-roadmap-template.md) and [slicing heuristics](./references/slicing-heuristics.md), then instantiate the template without dropping required sections. `## Progress Tracking` is mandatory: scaffold and status use it as the shared lifecycle record. Derive independently executable vertical slices with observable outcomes and explicit real dependencies; review the graph before writing. |
| Create a PRD and roadmap from an unformed idea | Route to `$speckit-pro:speckit-prd`; use its PRD and roadmap templates. |
| Track one spec through phases or understand scaffold-created workflow files | [Workflow template](./templates/workflow-template.md). For scaffold creation, population, inputs, or output locations, also read the [live scaffold skill](../speckit-scaffold-spec/SKILL.md) as a reference only; do not execute or invoke it. Keep artifacts, gates, and decisions current. |
| Explain the roadmap home note | [Roadmap-MOC guide](./references/roadmap-moc-guide.md). The curated zone is author-owned; the generated index is regenerated, not hand-edited. |
| Explain or configure presets, extensions, hooks, catalogs, or upgrade recovery | [Presets and extensions guide](./references/presets-extensions-guide.md). Discover and inspect first; confirm before any mutation. |
| Explain autonomous execution, consensus, remediation, review loops, or configuration | Read [the live autopilot skill](../speckit-autopilot/SKILL.md) as a reference only; explain its current contract without executing or invoking it. When the user asks to run an existing workflow, use the bounded autopilot hand-off below. |
| Scope a raw brief interactively | Route to `$speckit-pro:grill-me`; do not conduct its interview here. |
| Scaffold a roadmap item, inspect status, resolve a review, or repair Codex agents | Route respectively to `$speckit-pro:speckit-scaffold-spec`, `$speckit-pro:speckit-status`, `$speckit-pro:speckit-resolve-pr`, or the SpecKit Pro `install` skill. |
| Archive merged-spec records or clean up archived specs | Route to `$speckit-pro:speckit-archive-cleanup`; do not copy or execute its cleanup workflow here. |

## Bounded autopilot hand-off

Coach never executes or invokes autopilot. When an execution request belongs to
autopilot, the redirect is a bounded operator handoff, not merely a command
name. Briefly explain that autopilot owns the durable phase task list, native
subagent lifecycle, gate checks, workflow and state updates, and
operator-visible progress. When the user supplies the workflow path, return one
explicit command bounded to the requested stage:

```text
$speckit-pro:speckit-autopilot <workflow-file> --stage plan|implement|full
```

Use `--stage full` only when the user asks for the full workflow; use
`--stage plan` or `--stage implement` only for that requested boundary. Preserve
the supplied workflow path exactly, and do not invent a path or broaden the
stage. Set the expectation that autopilot continues through the selected stage
until completion or a real gate/capability stop, owning its own preflight,
phase gates, durable state, and fail-closed stops; at a stop it preserves
evidence and reports the exact action required to resume. Do not imply that
Coach will run, monitor, or resume that lifecycle. If no workflow path exists,
route to
`$speckit-pro:speckit-scaffold-spec`;
if the path or requested stage is ambiguous, ask for that missing input instead
of executing anything.

## Keep adjacent contracts distinct

- Scaffold owns roadmap lookup, branch/worktree placement, the canonical
  Design Concept and workflow artifacts, and the explicit hand-off to
  autopilot. Grill Me owns the interactive interview that supplies the Design
  Concept. Grill Me never creates a branch, worktree, workflow, or roadmap
  status update, and scaffold must not synthesize interview answers when a
  human interaction surface is unavailable.
- Canonical scaffold artifacts are
  `docs/ai/specs/.process/SPEC-<ID>-design-concept.md`,
  `docs/ai/specs/.process/SPEC-<ID>-workflow.md`, and
  `specs/<branch-name>/SPEC-MOC.md`. The workflow's `Branch` field records the
  actual dedicated spec branch returned and verified by scaffold; it is never
  `main` and is never guessed from the spec title alone.
- A declared extension command is not evidence that a hook runs
  automatically. Inspect installed presence, host command registration, and
  enabled hook wiring separately. Report an empty declaration as empty and an
  unavailable registration or wiring check as unverified; never describe a
  command declaration as a hook.

## Essential SDD guidance

The normal flow is `constitution → specify → clarify (as needed) → plan → checklist (as needed) → tasks → analyze (as needed) → implement`. Preserve the traceability markers in the artifacts: user stories, functional requirements, clarifications, parallel-safe tasks, and checklist gaps. Do not advance past a failed gate; explain the artifact or evidence that must change first.

For roadmap decomposition, prefer delivery slices over technical layers when a slice can be independently verified. Keep cross-spec dependencies explicit, minimize them, and use mocks only when the handoff contract is defined.

## Project fixup

For a request to repair an existing SpecKit Pro project:

1. Confirm the target root and inspect `git status`; preserve unrelated changes.
2. Inspect presets and resolve each affected core template with `spec_kit.cli_argv + ["preset", "resolve", "<template>"]`, taking `cli_argv` from the `check-prerequisites` helper and never launching bare `specify`. Move durable local customizations into a project preset rather than modifying core templates.
3. Restore a core template only from a reviewed source (version control, verified backup, or official template); never reconstruct it by guesswork.
4. Preserve any host PR template and run the relevant project checks before reporting the resolved template paths, restored files, and remaining manual follow-up.

Files in this skill

  • SKILL.md8.7 KB
  • agents/openai.yaml295 B
  • examples/formal/counter-quint/Counter.cfg38 B
  • examples/formal/counter-quint/Counter.qnt308 B
  • examples/formal/counter-tlc/CounterTLC.cfg74 B
  • examples/formal/counter-tlc/CounterTLC.tla409 B
  • examples/formal/counter-tlc/catalog.json1016 B
  • examples/formal/counter-traces/Counter.cfg78 B
  • examples/formal/counter-traces/Counter.tla330 B
  • examples/formal/counter-traces/CounterTrace.swift2 KB
  • examples/formal/counter-traces/counter.mts1.3 KB
  • examples/formal/counter-traces/counter.py1.1 KB
  • examples/formal/counter/Counter.cfg57 B
  • examples/formal/counter/Counter.tla406 B
  • examples/formal/counter/catalog.json845 B
  • examples/formal/tooling/quint/package-lock.json30.2 KB
  • examples/formal/tooling/quint/package.json112 B
  • references/checklist-domains-guide.md7.4 KB
  • references/command-guide.md4 KB
  • references/constitution-guide.md2 KB

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.