Turn intent and repository evidence into native reviewed plans and atomic,
signed Task-Specs; run authorized recipes with TaskMesh, inspect canonical
acceptance, and retain release and maintenance evidence.
Installs into .claude/skills of the current project.
Are you the author of Claude Code?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/luanmorenommaciel-claude-code)
---
name: task-spec
description: |
Turn intent and repository evidence into native reviewed plans and atomic,
signed Task-Specs; run authorized recipes with TaskMesh, inspect canonical
acceptance, and retain release and maintenance evidence.
metadata:
version: "3.10.0"
---
# Task-Spec for Claude Code
Task-Spec is the atomic contract; Claude Code is a replaceable authoring or
execution harness. The canonical doctrine ships with the installed engine, and
the `taskspec` CLI owns deterministic validation and mutation.
Read current `taskspec agent-context` and task or initiative status before acting.
Use the same product skill and persisted state as other supported harnesses;
conversation history does not authorize work. Detailed workflows are available
through `taskspec guide intent`, `decomposition`, `recipes`, `execution`,
`acceptance`, `sdlc`, and `migration`.
## Prerequisites
The `taskspec` CLI must be on PATH and `taskspec doctor` must pass. Use the
repository installer with `--global` for user-level Claude, Codex/Kimi, and Grok
skills, or `--target <repo>` for project-local copies.
## What to do
1. Inspect repository instructions, the smallest relevant code slice, its tests,
schemas, and ownership boundaries.
2. Use optional research only when explicitly selected; normalize retained
results as credential-free `AuthoringEvidence/v1`.
3. Compose every declared unit in `TaskPlan/v1`, preview it, and generate only
the approved plan. Never invent omitted paths, evals, dependencies, or
children.
4. Drive the CLI; never hand-edit authorization or acceptance fields:
| Intent | Command |
|--------|---------|
| Prepare native decomposition | `taskspec decompose prepare <initiative> --recipe <file>` |
| Explicit topology review | `taskspec decompose review <initiative> --accept --reviewer <name> --reason <text>` |
| Compile reviewed plan | `taskspec decompose compile <initiative>` |
| Compose and preview | `taskspec plan --manifest <TaskPlan>` |
| Generate approved units | `taskspec batch --plan <TaskPlan>` |
| Direct scaffold | `taskspec new <slug> <effort> [agent]` |
| Prove authoring | `taskspec validate <spec>` then `taskspec dod <spec>` |
| PRE-gate authorization | `taskspec gate --stamp <spec>` |
| Portable handoff | `taskspec handoff <spec> --backend claude --out <file>` |
| POST-gate acceptance | `taskspec accept --handoff <file> --stamp <spec>` |
| Status and recovery | `taskspec status <id>` and `taskspec doctor --backlog` |
| Optional TaskMesh cockpit | `taskspec mesh frontier`, `run`, `watch`, `accept`, and `finish` |
| Managed recipe or native graph | `taskspec mesh run --initiative <id> --execute` |
| Opt-in evidence task | `taskspec new --format 4 <slug> <effort>` then `taskspec author-doctor <spec>` |
5. Respect the effort gate: XS/S/M/L are runnable leaves; L needs a backend
from `TASKSPEC_LONG_HORIZON_BACKENDS`. XL/XXL are non-runnable nodes with at
least two/three child Task-Specs.
## Hard rules
- `signed_off: true` comes ONLY from `taskspec gate --stamp`. Never hand-stamp
the envelope; HMAC v3 seals `TaskRevision/v1` and unknown authority fields.
- A task is DONE only when `taskspec accept` emits `ACCEPTED=1` — an agent's
own claim of GREEN is not evidence. Acceptance must bind `TaskHandoff/v3`, the
attempt, base commit, closure, repository scope, and required receipts.
- Format v4 acceptance must receive every policy-required external receipt;
never expose a private holdout bundle to the executor.
- Changed scope or proof policy requires explicit replanning and reauthorization;
never silently rewrite downstream dependencies.
- Resolve recipe guidance before sealing. Managed recipes use bounded sequential
rounds in the existing TaskMesh lifecycle, including one-task graphs. Workers
cannot sign or accept their own work. Reuse applicable recorded decisions and
ask only for missing decisions or authorization.
- An incident becomes candidate work. Pipeline success is reported evidence;
it does not establish target health or authorize deployment.
- Exit codes: `0` pass/accept · `1` failed/rejected · `2` usage error.
- When TaskMesh is installed, Claude may start or reconnect to the durable
cockpit, but the daemon owns the run and Task-Spec owns authority. Do not call
a supervised worktree a sandbox, self-accept an attempt, or merge the target
branch. Autonomous OMP requires externally verified isolation and expiring
credential evidence.