Orchestrate GitHub Spec Kit's core commands as bundled, human-gated phases tailored to Claude/Cowork. Use when the user wants semi-automated spec-driven development: requirements preparation, implementation preparation, an explicit approval gate before any source-code change, scoped implementation, then verification and closeout. This skill embeds and invokes the installed core Spec Kit skills (constitution, specify, clarify, checklist, plan, tasks, analyze, implement) rather than replacing them. It uses Claude-native affordances: AskUserQuestion for gates, the task list for progress, the MATLAB MCP server for verification, and subagents for independent review.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Orchestrate GitHub Spec Kit's core commands as bundled, human-gated phases tailored to Claude/Cowork. Use when the user wants semi-automated spec-driven development: requirements preparation, implementation preparation, an explicit approval gate before any source-code change, scoped implementation, then verification and closeout. This skill embeds and invokes the installed core Spec Kit skills (constitution, specify, clarify, checklist, plan, tasks, analyze, implement) rather than replacing them. It uses Claude-native affordances: AskUserQuestion for gates, the task list for progress, the MATLAB MCP server for verification, and subagents for independent review.
Requires a Spec Kit project initialized with .specify/ and the core speckit skills installed under .claude/skills/. Designed for Claude Code / Cowork. Honors the project constitution at .specify/memory/constitution.md, including its implementation gate and implementation-receipt ledger.
You MUST consider the user input before proceeding (if not empty).
Spec Kit Human Loop (Claude-tailored)
Run the ordinary GitHub Spec Kit workflow as a small number of human-gated
bundles, so the user does not chain eight commands by hand. This skill is an
orchestrator: it must not replace, summarize away, or silently skip the core
Spec Kit commands. At each phase it invokes the installed core skill and follows
its instructions.
This version is adapted for Claude/Cowork and for projects whose constitution
already enforces a strict implementation gate and an implementation-receipt
ledger. Where this skill and the project constitution overlap, the
constitution wins and this skill defers to it (see "Constitution reconciliation").
Non-negotiable rules
No source edits before approval. Do not modify source code, tests, build
files, or other implementation artifacts before the human explicitly approves
implementation at Gate 2 and the approval is expressed as a real
implementation invocation (see Constitution reconciliation). Spec Kit
artifacts under specs/<feature>/ and the orchestration state file are not
"source code" and may be written during preparation.
Embed, never emulate, core commands. Each bundle invokes the corresponding
installed core skill via the Skill tool (e.g. speckit-specify). Do not
hand-write spec.md, plan.md, or tasks.md unless the core skill is
unavailable and the user explicitly asks for a temporary fallback.
Sparse gates only. Stop for a human decision only at the three gates, or on
a real blocker (ambiguity, failed quality gate, excessive scope, an attempted
source edit). Do not ask between routine phases.
Record every human decision in the orchestration state file.
spec.md is the authority for requirements. When it changes, re-derive the
downstream artifacts (plan.md, tasks.md, analysis) rather than editing them
in isolation.
Stop and report if the repository is not a Spec Kit project, a required core
command is missing, the active feature cannot be determined, or the approved
scope cannot be enforced. Do not invent replacements for missing core commands.
Constitution reconciliation (read first)
Before Bundle 0, read .specify/memory/constitution.md if present and obey it.
This skill is explicitly subordinate to it. In particular:
Implementation gate. If the constitution restricts implementation to an
explicit invocation, then a Gate 2 menu choice is not sufficient by itself.
As of constitution v1.4.0 two implementation invocations are sanctioned and
equivalent: the core implementer (/speckit-implement, $speckit-implement) and
the agent-assign pipeline (/speckit-agent-assign-assign ->
/speckit-agent-assign-validate -> /speckit-agent-assign-execute, run in series);
a stated exact override phrase is the only other authorization. Gate 2 records the
human's intent and approved scope; the skill then proceeds to implementation only
via one of the constitution's required invocations. If the user picked "approve"
in the gate but the constitution needs an explicit implement invocation, ask the
user which path to use and to confirm with that exact invocation before any edit.
Run record / receipt. If the constitution mandates an implementation-receipt
ledger at a specific path (for example
<FEATURE_DIR>/agent-runs/<UTC-timestamp>-<short-name>/implementation-receipt.md),
that receipt is the canonical record of an implementation run. Do not invent
a competing record. This skill's human-loop.md is a lightweight orchestration
index only, and must cross-reference (not duplicate) the constitutional receipt.
Follow the receipt's required sections and concision rules exactly.
Language/coding standards. If the constitution sets coding standards (for
example, MATLAB rules: no evalc that suppresses warnings or shadows built-ins,
warnings must stay visible, propagate error stacks, no nargin), surface those
as checks in the implementation and verification bundles.
Run-directory naming. If the mandated receipt directory name does not match the
active agent, do not silently rename it โ the path is fixed by the
constitution. Follow it; renaming is a speckit-constitution change, not an
ad-hoc edit. (As of constitution v1.5.0 this directory is the agent-neutral
agent-runs/.)
Claude-native affordances (how to run each step)
Invoke core skills with the Skill tool, not shell aliases. Use
speckit-constitution, speckit-specify, speckit-clarify, speckit-checklist,
speckit-plan, speckit-tasks, speckit-analyze, speckit-implement.
Track bundles and tasks with the task list. At the start of a run, create a
task per bundle with TaskCreate; mark each in_progress/completed with
TaskUpdate. During implementation, mirror tasks.md entries as tasks so the
user sees live progress.
Run human gates with AskUserQuestion, not a plain-text lettered menu.
Present the gate's options as structured choices (the tool always adds an
"Other" escape). Record the selected option verbatim in human-loop.md.
Verify with the MATLAB MCP server when MATLAB work is involved: prefer
mcp__matlab__check_matlab_code for static checks, mcp__matlab__run_matlab_test_file
for the narrowest relevant test, and mcp__matlab__run_matlab_file /
mcp__matlab__evaluate_matlab_code for reproducibility checks. Follow the
project's MATLAB workflow doc for diaries/logs. Capture solver status strings
and residuals faithfully.
Use a subagent for independent review (Gate 2 option "independent review")
via the Task tool, so the review does not share the implementing context.
Surface review artifacts to the user with present_files
(implementation-review.md, comparison reports, diffs).
Hook awareness
If .specify/extensions.yml exists with auto_execute_hooks: true, each embedded
core command may trigger before_*/after_* hooks (for example git auto-commit
and agent-context refresh). Because a bundle runs several phases in sequence, this
can produce several commit prompts. At the start of a bundle, tell the user that
per-phase hooks will fire, and offer to (a) let each hook prompt as configured, or
(b) defer commits and make one commit at the end of the bundle. Never disable
hooks silently.
Orchestration state file
Maintain specs/<feature>/human-loop.md once the feature directory exists (create
it immediately after speckit-specify creates the directory). Keep it lightweight;
it indexes the run and points at the constitutional receipt rather than copying it.
# Human Loop State## Current State- Status:
- Active feature directory:
- Last completed bundle:
- Source code modified by this workflow: yes/no
## Core Command Ledger- constitution: (checked | invoked | n/a)
- specify:
- clarify:
- checklist:
- plan:
- tasks:
- analyze:
- implement:
## Human Decisions
| Date (UTC) | Gate | Option chosen | Consequence |
|---|---|---|---|
## Approved Implementation Scope- Approved: no
- Scope: (all | slice:<phase/task ids>)
- Tasks approved:
- Tasks deferred:
- Files allowed:
- Files not allowed:
## Pointers- Implementation receipt(s): <path(s) under the constitution's receipt ledger>- Implementation review: specs/<feature>/implementation-review.md
## Open Risks and Ambiguities
Bundle 0: project and phase detection
Confirm the current directory is inside a Git repository and .specify/ exists.
If not, stop and ask the user to run specify init or give the correct root.
Read .specify/memory/constitution.md and apply "Constitution reconciliation".
Check git status; warn if uncommitted source changes already exist.
Determine the active feature from the current branch and specs/.
Confirm the core skills are installed under .claude/skills/speckit-* (fall
back to .agents/skills/ or .specify/templates/commands/ only if needed).
Decide the requested mode: start, prepare, implement, continue, or
closeout. If unclear, ask with AskUserQuestion:
Start a new human-gated workflow from this request.
Continue the active feature workflow.
Prepare the implementation review for the active feature.
Implement the approved task scope.
Verify and close out the active feature.
Bundle 1: requirements preparation
Embedded core skills, in order:
speckit-constitution โ check, do not regenerate. If the constitution is
missing, stop and ask for governing principles (or invoke it if the user
supplied enough detail). If present, read it and record constitution: checked.
Only invoke it to create/revise principles when the user explicitly asks.
speckit-specify with the feature request.
speckit-clarify. If it raises blocking questions, stop at Gate 1.
speckit-checklist. If failures are blocking, stop at Gate 1.
Ensure human-loop.md exists and records completed commands. Then Gate 1.
Approve only the first independently testable slice (name it, e.g. a single plan phase).
Revise the task list, then re-run analyze.
Revise the plan, then re-run tasks and analyze.
Run an independent review (subagent) before deciding.
Only "approve all" or "approve a slice" can authorize edits โ and only then via
one of the constitution's required implementation invocations (see reconciliation).
When approving, also capture which implementation path the human wants: the core
implementer (speckit-implement) or the agent-assign pipeline (assign -> validate ->
execute). Record the approved scope and the chosen path in both human-loop.md
and implementation-review.md.
Bundle 3: approved implementation
Preconditions: implementation-review.md exists; human-loop.md records explicit
approval; the approved scope is "all" or a named slice; and the user has supplied
the constitution's required implementation invocation if one is mandated.
Mirror the approved tasks.md entries into the task list.
Run the approved implementation path, constrained to the approved scope:
Core implementer: invoke speckit-implement.
Agent-assign pipeline: invoke speckit-agent-assign-assign, then
speckit-agent-assign-validate (it MUST pass โ no unassigned tasks or missing
agents), then speckit-agent-assign-execute. Every spawned subagent stays within
the approved scope and the constitution's read-only-path rules.
Do not edit source directly and do not bypass the chosen path.
Follow task order/dependencies from tasks.md. Stay inside the approved files;
if a change outside scope is genuinely required to keep tests meaningful, record
it as a deviation rather than silently expanding scope.
Apply constitution coding standards as you go (e.g. MATLAB evalc/warning rules).
Run the narrowest relevant test via the MATLAB MCP server (or the project's test
harness) after the change.
Write the constitutional implementation receipt at its mandated path with the
required sections (Prompt, Final response, Diff summary, Tests, Unresolved
issues). The Final response must be the actual final user-facing completion
text, not a paraphrase. If the agent-assign pipeline ran, a single receipt covers
the whole run and its Diff summary aggregates every file changed across all
spawned subagents. Point human-loop.md at the receipt.
Bundle 4: verification and closeout
Inspect git diff --stat and git diff.
Compare the work against tasks.md and implementation-review.md.
Run or report tests (narrowest relevant first; use the MATLAB MCP server when
applicable). Preserve exact solver status strings and residual scale labels.
Optionally re-run speckit-analyze post-implementation (does not replace the
pre-implementation analyze).
Update human-loop.md: diff summary, tests, unresolved issues, deviations,
next recommended choice. Confirm the constitutional receipt is complete.
Gate 3 with AskUserQuestion.
Gate 3 โ closeout decision
Options:
Accept and close the feature.
Continue with the next approved slice.
Fix failing tests.
Reconcile implementation drift by updating spec/plan/tasks.
Revert selected changes.
Open a follow-up specification.
Gate discipline
Use AskUserQuestion with at most four primary options (the tool adds "Other").
State the safe default when there is one.
Record the exact option chosen and its consequence in human-loop.md.
Continue automatically only after a choice that permits continuation.
Stop conditions
Stop rather than auto-continuing if: a core skill is missing; the active feature
is undeterminable; the constitution is missing and the user has not supplied
principles; clarify/checklist/analyze report blocking issues; source code would be
modified before Gate 2 approval (or before the constitution's required implement
invocation); the approved scope cannot be enforced; or tests fail when the mode
was not test repair.
