AI Skills for Tenstorrent Products
Overview
Skills for agents working on tt-metal, TTNN, Metalium, models, and related projects. Tell your agent what you want to do; the finder helps it choose and install the right plugin. A plugin is a collection of related skills that you can enable or disable together. — use AutoDebug, AutoFix, and model bring-up. This is the normal way to use this repository. If the required plugin is missing, the finder recommends it and asks for your approval. Approve the installation and your agent follows the host's installation flow, then uses the skills to do the work. Restart the session if prompted. For this example, it installs tt-autodebug, uses AutoDebug to investigate, and AutoFix to repair. Or describe the job, such as “Review this TTNN pull request,” and let the finder recommend a suitable plugin. You choose which plugins to install. Once installed, their skills can be selected automatically for matching tasks; you can also ask your agent to make them explicit-only.
README
Tenstorrent skills
Skills for agents working on tt-metal, TTNN, Metalium, models, and related projects. Tell your agent what you want to do; the finder helps it choose and install the right plugin. A plugin is a collection of related skills that you can enable or disable together.
Getting started with Agentic Research skills — use AutoDebug, AutoFix, and model bring-up.
Get started
Ask your Codex or Claude Code agent:
Read https://github.com/tenstorrent/skills and set up the marketplace and
its tt-skills finder plugin for this agent.
Once the finder is active, just ask for the skill you want to use:
Use AutoDebug to fix .
This is the normal way to use this repository. If the required plugin is missing, the finder
recommends it and asks for your approval. Approve the installation and your agent follows the
host’s installation flow, then uses the skills to do the work. Restart the session if prompted.
For this example, it installs tt-autodebug, uses AutoDebug to investigate, and AutoFix to repair.
You can also ask:
Use tt-model-bringup on .
Or describe the job, such as “Review this TTNN pull request,” and let the finder recommend a suitable plugin. You choose which plugins to install. Once installed, their skills can be selected automatically for matching tasks; you can also ask your agent to make them explicit-only.
Plugin catalogue
| Plugin | Purpose |
|---|---|
tt-skills |
The finder: recommends relevant plugins and helps you install them with your approval |
tt-review-skills |
PR and diff review for TTNN, Metalium, LLK, models, serving, multi-chip, trace, precision, testing, and L1 changes |
tt-autodebug |
AutoDebug and AutoTriage investigate code issues and hangs; AutoFix tests hypotheses and repairs the cause |
tt-model-bringup |
Eleven stages from HF decoder through TTNN/vLLM benchmarking, with chunked-prefill guidance, prefill/TTFT optimization, trace warmup/reuse and allocation validation, targeted path/serving checks, and fixed standard accuracy subsets plus 4K-input vLLM benchmarks at concurrency 1 and 32. Includes a standalone TTI release skill. Requires tt-autodebug. |
tt-debug-tools |
Drive the Tenstorrent debug tools and read their output: tt-triage, dprint, watcher, asserts, etc. See plugins/tt-debug-tools/README.md |
tt-buddy |
A coding agent with Tenstorrent operating principles: takes notes all the time, learns when needed, keeps the diff minimal. Device runs need tt-device-mcp. See plugins/tt-buddy/README.md |
tt-serve-model |
Take a model from host discovery to a verified server on Tenstorrent hardware: discover the host (boards vs ASICs, --tt-device string, driver/firmware, hugepages, docker, device holders, reset verdict), retrieve the model (HF gating, tt-inference-server spec entry, image spec drift, GHCR tag/ancestry, tt-model-manager bundles, exact pull command), diagnose a launch (stage/hang/failure classification, mandatory docker flags, override safety, timing, displaced services) and verify the running server (health, correctness, mesh utilization, throughput/batching, reasoning-model checks). |
AutoDebug investigates in a fresh agent process and writes AUTODEBUG.md. AutoFix handles source
changes and validation. For examples and expected outputs, see the
Agentic Research guide.
AutoDebug and AutoTriage prompts and investigation runners are maintained here. Case curation, backtests and improvement flows live in skills-autoimprove, which runs a recorded skills revision through the shared runner interface.
If AutoDebug’s Codex child sandbox fails, the calling agent assesses the error and existing
authorization before explicitly retrying with AUTODEBUG_SKIP_CHILD_SANDBOX=1; see the
sandbox guidance.
Model bringup serving uses the standalone TT vLLM plugin and resolves its recommended upstream vLLM version for each new run. Before each local launch, the runner rejects known legacy vLLM and bundled-plugin source layouts. See serving stack setup.
Model bring-up
Ask your agent to use tt-model-bringup with the HF model ID in your target checkout. The finder
recommends both tt-model-bringup and its tt-autodebug dependency if needed, and asks for approval
to install them.
The stage skills support Codex and Claude Code; automated multi-goal execution uses Codex.
See the Agentic Research guide for how to
run a bring-up, and startup and stage orchestration
for detailed setup, dry-run, and resume instructions.
For new Codex multigoal runs, model-bringup automatically installs and uses the telemetry
plugin from tenstorrent/ar-dashboard when the execution host’s existing credentials can
read that repository. Users without access continue silently, without login prompts or
an extra dependency. Explicit opt-outs and disabled integrations are respected. The
setup instructions
connect the installed plugin through generic lifecycle hooks;
dashboard endpoints and upload code stay in the private plugin.
Alternative manual skill installation
If you prefer to install plugins yourself, use the commands below. GitHub authentication is required while the repository has restricted visibility.
Codex
In a terminal, register the marketplace and install the finder:
codex plugin marketplace add https://github.com/tenstorrent/skills.git
codex plugin add tt-skills@tenstorrent-skills
Choose any optional plugins you want:
codex plugin add tt-autodebug@tenstorrent-skills
codex plugin add tt-review-skills@tenstorrent-skills
codex plugin add tt-debug-tools@tenstorrent-skills
codex plugin add tt-buddy@tenstorrent-skills
# tt-buddy ships session hooks: trust them in Codex with /hooks, then t.
codex plugin add tt-serve-model@tenstorrent-skills
# Model bring-up also requires tt-autodebug:
codex plugin add tt-model-bringup@tenstorrent-skills
You can also select plugins under Tenstorrent Skills in the Plugins Directory.
Claude Code
Inside the session, register the marketplace and install the finder:
/plugin marketplace add https://github.com/tenstorrent/skills.git
/plugin install tt-skills@tenstorrent-skills
Choose any optional plugins you want:
/plugin install tt-autodebug@tenstorrent-skills
/plugin install tt-review-skills@tenstorrent-skills
/plugin install tt-debug-tools@tenstorrent-skills
/plugin install tt-buddy@tenstorrent-skills
/plugin install tt-serve-model@tenstorrent-skills
For model bring-up, install tt-autodebug above and then:
/plugin install tt-model-bringup@tenstorrent-skills
Adding the marketplace does not bulk-install the optional plugins. Codex marks the finder as installed by default; Claude requires the explicit finder installation shown above.
Direct gh-aw use
The review skills remain directly pin-able by name in a gh-aw workflow:
skills:
- tenstorrent/skills/tt-review-core@
- tenstorrent/skills/ttnn-op-kernel-review@
- tenstorrent/skills/tt-l1-memory-review@
Pins resolve by skill name, not path — the bucket a skill lives in is invisible to the resolver, so skills can move between buckets without breaking a pin. Always pin a 40-character SHA: a pin that fails to resolve is reported as a non-fatal warning, so a typo degrades the review silently rather than failing the run.
See .github/workflows/tt-pr-review.md for a complete worked workflow.
How tt-review-skills composes
Load tt-review-core first — it carries the severity vocabulary, the evidence rule, the scope
rules, and the do-not-flag guards that every other skill assumes and does not restate. Then load
at most two domain skills. A reviewer holding fourteen checklists applies all of them shallowly.
tt-review-router maps changed paths to the right subset.
Reference
common — cross-cutting review discipline
| Skill | Reviews |
|---|---|
tt-review-core |
The shared contract: severity, evidence, scope, output shape, false-positive guards |
tt-review-router |
Maps changed paths to the domain skills that apply |
tt-test-coverage-review |
PCC bars, tile-boundary cases, program-cache tests, regression tests on bug fixes |
tt-perf-claim-review |
Whether a stated performance number is supported by its measurement |
tt-comment-hygiene-review |
Iteration-journey comments, tribal knowledge, magic values, op docstrings |
tt-split-pr-by-codeowners |
Whether a PR should be split so each piece needs fewer approvals |
models — model bringup, TTNN consumers
| Skill | Reviews |
|---|---|
tt-model-bringup-review |
Residual contract, QKV topology, logical batch vs tile padding, hidden host fallbacks |
tt-multichip-ccl-review |
num_links vs topology, bias before all-reduce, distributed RMSNorm, gather axes |
tt-trace-review |
Capture safety, program-cache warmup signatures, device-owned autoregressive state |
tt-precision-review |
Per-tensor-group dtype policy, the prefill/decode cache asymmetry, PCC-collapse triage |
ttnn — TTNN op authors
| Skill | Reviews |
|---|---|
ttnn-op-kernel-review |
The eight structural categories: init, TRISC sync, tile_regs, CB UB, work split, semaphores, control flow, in-place |
metal — tt-metal host and kernel infrastructure
| Skill | Reviews |
|---|---|
tt-l1-memory-review |
Buffer inventory discipline, data-movement tiers, CB sizing, accumulator capacity |
llk — low-level kernels
| Skill | Reviews |
|---|---|
llk-race-audit-review |
Nine race hazard classes and the cross-class seams, under a monotonic join contract |
llk-perf-audit-review |
Static Tensix perf under a provenance lens and a semantic-equivalence gate |
inference — serving
| Skill | Reviews |
|---|---|
tt-vllm-serving-review |
Generator contracts, plugin registration, the tt_data_parallel ambiguity |
meta — catalogue maintenance
| Skill | Reviews |
|---|---|
tt-skills-upstream-audit |
Drift between vendored skills and their upstream sources (user-invoked) |
Attributions
**This repo is an aggregation. ** The skills here are vendored, reshaped and re-framed from work other people did — often work that took years of debugging to learn. The structure is borrowed too.
Everything below was consulted while building this, whether or not content was ultimately taken. Primary author is the top contributor to that path by commit count.
Content sources
| Source | Primary author | What came from it |
|---|---|---|
tenstorrent/tt-buddy — skills/ |
@ppetrovicTT | The reviewer contract: evidence rule, severity taxonomy, read-past-the-diff discipline, the reviewer cast |
tenstorrent/tt-buddy — knowledge/ |
@viktorpusTT | CCL and matmul knowledge, vLLM recipes |
tenstorrent/tt_ops_code_gen |
@mstaletovicTT | The eight-category structural kernel checklist, L1 footprint discipline, memory and precision references |
tt-metal — .agents |
@yieldthought | Optimization rules, multichip, tracing, datatype sweep, vLLM integration — and the Codex PR-review skill this catalogue’s output format came from |
tt-metal — tt-llk/.claude |
@ndivnicTT | The LLK audit suite as a whole |
⤷ race-audit-all |
@amahmudTT | Nine hazard classes, the monotonic JOIN contract, per-architecture divergence |
⤷ perf-optimization-audit |
@fvranicTT | The provenance lens, semantic-equivalence gate, SIMD false-positive guards |
tt-metal — .github/bug_checker |
@stevendae | Rules distilled from ~1,398 merged fix PRs: program-cache correctness, op validation, CCL ring buffers, stale LLK config. Strong evidence of which failures recur |
tt-metal — tech_reports/Handling_Special_Value |
@ttmtrajkovic | NaN/Inf/denormal semantics and the FPU/SFPU divergence |
tt-metal — docs/source/tt-metalium/tools, tt_metal/impl/debug, tt_metal/tools, tt_metal/hw/inc/api/debug |
see note below | The tt-debug-tools plugin: watcher, DPRINT, NoC debug dump, asserts and the LLK sanitizer, Inspector, triage, the device profiler, and TTNN’s host-side debug flags |
tenstorrent/tt-exalens |
see note below | The tt-exalens skill: the command set, the server/remote and GDB modes, JTAG, and NoC selection |
Structure and tooling
| Source | Primary author | What came from it |
|---|---|---|
mattpocock/skills |
@mattpocock | The shape of the review catalogue. Bucketed skills///, progressive disclosure, trigger-style descriptions, in-progress/ and deprecated/, invocation bifurcation, changesets, the install-block convention |
githubnext/gh-aw |
@dsyme, @pelikhan, @mnkiefer | The consumer. skills: frontmatter, safe-outputs, and the mattpocock-skills-reviewer triage pattern the reference workflow follows |
If your work is here and the attribution is wrong, thin, or you would rather it were not — open an issue and we will fix or remove it. Everything vendored is Apache-2.0, but licence compliance and proper credit are different things, and we care about the second one.
Validate changes
python3 scripts/sync_review_plugin.py --check
pytest # repo and every plugin
claude plugin validate . --strict
Agent-driven evals are separate — they cost money per run and never run in CI. Each plugin invokes
its own; see plugins/tt-debug-tools/README.md for that plugin’s.
Contributing
See CONTRIBUTING.md for the plugin layout, contribution workflow, versioning, validation, and attribution requirements. Maintainers are listed in CODEOWNERS.
Security
For reporting security vulnerabilities privately, see SECURITY.md.
Code of Conduct
This project follows the Contributor Covenant Code of Conduct.
License
Licensed under the Apache License, Version 2.0, with third-party notices and license details in NOTICE. See LICENSE_understanding.txt for the accompanying Tenstorrent rights clarification.
Recommended Tools
Try a different keyword or remove a filter.
Install
npx skillfish add tenstorrent/skills