# prd-taskmaster grades the PRD before it lets a task ship > prd-taskmaster is a Python engine for Claude Code that turns a one-line goal into a graded PRD, a dependency-ordered tasks.json graph and evidence-gated execution ending in a SHIP_CHECK_OK token, with the author labelling the whole project pre-alpha and its cost savings as estimates. **anombyte93/prd-taskmaster** — Zero-config goal-to-tasks engine for Claude Code (the Atlas engine). Graded PRD validation, dependency-ordered task graph, evidence-gated execution. - Repository: https://github.com/anombyte93/prd-taskmaster - Stars: 604 · Forks: 59 - Language: Python - License: MIT - Published: 2026-09-10 · Updated: 2026-09-10 · Language: en - Canonical page: https://hysenlabs.com/projects/anombyte93-prd-taskmaster ## From goal to SHIP_CHECK_OK in five phases The pipeline is written as one line: goal, then a discovery interview, then a graded PRD, then a dependency-ordered task graph, then verified execution. Each step is a phase the engine actually runs, and they are named Preflight, Discovery, Generate, Handoff and Execute. Preflight detects the environment, meaning which backend is native, whether the optional TaskMaster backend is present, which model CLIs exist and whether research tooling is available, and configures it with zero setup questions. Discovery is an adaptive interview that asks one question at a time to capture real constraints rather than your first guess. Generate writes the PRD, scores it against deterministic quality checks for a letter grade, then parses it into a task graph with complexity scores and full subtask coverage. Handoff detects what is installed and recommends one execution mode. Execute is the loop that implements each task and proves it with evidence, ending in the deterministic SHIP_CHECK_OK token. The split matters because it puts the two expensive steps, the interview and the execution loop, on either side of an artifact you can read and reject before any code is written. ## Two install paths, and the free engine needs no API key Path 1 is the one-liner the README recommends, and it installs the skill plus the prd_taskmaster package: ```bash curl -fsSL https://atlas-ai.au/install | bash ``` Path 2 is the Claude Code plugin, which adds a marketplace and installs from it, with a separate optional npm step to unlock the TaskMaster backend: ```bash # add the marketplace, then install the plugin /plugin marketplace add anombyte93/prd-taskmaster /plugin install prd ``` Either way, the first run happens inside Claude Code with the slash command /prd:atlas, or /prd:go, or by saying plainly that you want to build something. Requirements are Python 3.11 or newer on Linux, macOS or WSL. The free engine needs no paid API key because it drives the model CLIs you already have, and the optional local research proxy is bring your own rather than bundled. One install wrinkle is documented rather than hidden: npm runs a postinstall step that pip-installs the MCP server's Python dependencies. ## Placeholders floor the grade to NEEDS WORK Every generated spec is scored against deterministic checks and lands on one of four grades: EXCELLENT, GOOD, ACCEPTABLE or NEEDS WORK. The scoring is not advisory, because one class of content overrides it. Placeholders, meaning TBD, {{...}} or TODO in bare or bracketed form, are a hard fail. The grade floors to NEEDS WORK and validate-prd exits non-zero, so a spec with holes in it cannot be graded as acceptable and handed to the execution loop. The sample output the README prints shows what a pass looks like: grade GOOD, 49 of 57 checks at 86 percent, 11 checks passed across structure, testability and metrics, 2 warnings that are quoted and located rather than merely counted, a clean placeholder scan, and 14 tasks parsed into 52 subtasks with dependencies mapped. The task graph itself is described as backend-neutral tasks.json entries carrying dependencies, complexity scores and full subtask coverage, and the README is explicit that this is not a flat checklist. What that sample output does not show is which model produced it, or what it cost. Those numbers come later, if at all. ## One non-zero exit status blocks the ship token Evidence is what turns done from a claim into a state. Each task is implemented and has to produce execution evidence before it counts, and the completion token is emitted only when every gate passes. The blocking rule is narrow and checkable: a single non-zero Exit status in any evidence file blocks SHIP_CHECK_OK. There is exactly one way around it, and the README describes that way in enough detail to audit. It is an explicit admin override flag. It is written to the audit log. And it marks the token as [OVERRIDE] on stdout, so an override announces itself in the one place a downstream script is already reading. That last detail is what keeps the gate from decaying into a formality. A silent override would let the token mean whatever the operator wanted it to mean, while an override that changes the token text is something a pipeline can grep for, and something a reviewer can find in the log later without asking anyone. ## tasks.json outlives whichever vendor you route through The persistent state is deliberately vendor neutral. The PRD, the task graph and the execution state stay as plain files in your repository, which is what lets them survive a vendor swap. Above them sits a cross-vendor fleet in which Claude, Codex and Gemini run as separate quota pools rather than one brittle model lane, so a rate limit on one of them does not stop the graph. TaskMaster is an option rather than a requirement, despite what the name suggests. Native mode works with no TaskMaster install at all, and installing task-master-ai 0.43.0 or newer unlocks TaskMaster's model-agnostic AI, meaning any API you configure such as Anthropic, OpenAI, Perplexity, Gemini or an OpenAI-compatible endpoint, plus isolated workdir expansion when that backend is the one selected. In package.json this is a peer dependency on task-master-ai at 0.43.0 or newer, marked optional, so an npm install does not fail when it is missing. The top-level tree also shows how little lives inside the Python package. Alongside prd_taskmaster/ and script.py there are agents/, hooks/, skills/, phases/, templates/, reference/, skel/, mcp-server/, docs/ and tests/. ## token_economy decides how fast the engine escalates Model routing is one setting in .atlas-ai/fleet.json: ```json // .atlas-ai/fleet.json { "token_economy": "conservative" } // or "balanced" (default) / "performance" ``` The default is balanced, and the value controls how aggressive escalation is. The rule behind it is that every job runs on the cheapest model that can do the job, and escalation happens only when a validator says the attempt failed. Task decomposition and research run through whichever backend is selected, so the routing sits above the vendors rather than inside any one of them. The examples given are concrete. Complexity 2 scaffolding gets a haiku-class model, the hardest long-running work gets the frontier model, and nothing defaults to expensive. A local report called economy-report shows your real success rate and latency per model, with the stated purpose of making the routing better on your workload rather than on the author's. The caveat sits in the project's own status section, and it matters more than the feature list: the cost numbers are verified-rate estimates, not measured guarantees. ## The postinstall warning names its own failure The npm postinstall step pip-installs the MCP server's Python dependencies, and the script is written to survive a machine without pip. It ends by echoing the consequence and the fix: the MCP tools will not start, and the command to repair it is pip install -r node_modules/prd-taskmaster/mcp-server/requirements.txt. The README calls that warning non-fatal, which is accurate in the narrow sense that the install continues. The rest of the tooling follows the same pattern of naming its own failure mode. Publishing is guarded by prepublishOnly, which runs an npm auth check and then a version sync check, and release:preflight runs the auth check on its own. Tests are pytest, with test:fast skipping anything marked integration and test:integration pointed at tests/mcp/test_integration.py. Two dates are worth keeping in mind before you plan around this. The repository was last pushed on August 14, 2026, and the most recent GitHub release is v5.2.0 from June 13, 2026, while package.json already declares 5.3.0. Main is ahead of the tagged releases. ## Pre-alpha, and the savings are the author's own estimates The status banner is unambiguous: pre-alpha, under active development, with Atlas recently consolidated into this engine and the newer systems not fully tested in the wild. It asks you to expect rough edges and breaking changes between releases, tells you to pin a version if you need stability, and states there is no warranty beyond the MIT license. Atlas Pro is a private pilot and is not generally available. The status section then draws a line inside the project. The deterministic core, meaning graded PRD validation, the task graph, the ship-check gate and the CLI, is covered by roughly 300 tests and is called the most stable surface. Around it, the cross-vendor fleet, the backend abstraction, the token-economy ledger and the bundled Pro MCPs are described as recently built and not yet battle-tested. Their numbers, cost savings in particular, are labelled verified-rate estimates rather than measured guarantees, with docs/product/MODEL-ECONOMY.md holding the reasoning. That split is also the fairest way to read the feature list. The gates are the part with tests behind them. The routing economy is the part whose numbers the author himself marks as unverified. ## Conclusion prd-taskmaster suits a team that wants the spec and the task graph as reviewable files and would rather have a gate that says no than a summary that says done. It does not suit anyone who needs a stable release cadence today, because the project calls itself pre-alpha and asks you to pin a version. Before you commit, read the status section for what is covered by tests and what is not, check that Python 3.11 or newer is available on Linux, macOS or WSL, and treat every cost figure in the model economy document as an estimate rather than a saving. ## FAQ ### What does prd-taskmaster do with a one-line goal? It interviews you one question at a time, writes a PRD, scores that PRD against deterministic checks for a letter grade, then parses it into a dependency-ordered tasks.json graph with complexity scores and full subtask coverage. ### Does prd-taskmaster need a paid API key to run? No. The free engine uses the model CLIs you already have, and requires Python 3.11 or newer on Linux, macOS or WSL. An optional local research proxy can be plugged in, but it is not bundled. ### What happens if a prd-taskmaster PRD still contains TBD or TODO? Placeholders are a hard fail. The grade floors to NEEDS WORK and validate-prd exits non-zero, so a spec containing TBD, TODO or {{...}} cannot pass the gate. ### Can prd-taskmaster bypass its SHIP_CHECK_OK gate? There is one documented way, and it leaves a trace: an explicit admin override flag, which is audit-logged and marks the token as [OVERRIDE] on stdout. Otherwise a single non-zero Exit status in any evidence file blocks the token. ### Does prd-taskmaster require task-master-ai to be installed? No. It is an optional peer dependency pinned at 0.43.0 or newer. Installing it unlocks TaskMaster's model-agnostic AI and isolated workdir expansion, while native mode works without it and keeps the same validated task graph. ### How much does prd-taskmaster cost to run? The engine needs no paid API key, and routing prefers the cheapest model that can do a job, escalating only when a validator says it failed. The cost savings the project quotes are verified-rate estimates rather than measured guarantees. ## Sources - [anombyte93/prd-taskmaster on GitHub](https://github.com/anombyte93/prd-taskmaster) - [Issues](https://github.com/anombyte93/prd-taskmaster/issues) - [License: MIT](https://github.com/anombyte93/prd-taskmaster/blob/main/LICENSE) - [README](https://github.com/anombyte93/prd-taskmaster/blob/main/README.md) - [Releases](https://github.com/anombyte93/prd-taskmaster/releases) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/anombyte93-prd-taskmaster