thought-layer-speedrun — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited thought-layer-speedrun (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
The fast path to a build-ready spec. You still reach the same SHAPE of output the full framework produces - a PRD hardened by the grill - but the validation in front of it is compressed to five fast, unranked stages, and there is far less conviction behind it. No panel, no confidence score, no 0.85 gate: one honest read per stage, and you move on whenever you want.
For the impatient, and a good way to see the whole arc - validation to spec - in one sitting (a demo). It trades the panel's defensible conviction for speed: the spec is real and buildable, but the validation behind it is a gut-check, not an adversarial panel. For anything you will actually build, run the full thought-layer-framework (/tl) - the rigor is the part that was ever scarce. Speedrun is the taste and the on-ramp.
For each of the five stages, give ONE honest read - not three personas, no tl_score, no grade, no gate:
Stay honest and stay fast: a few sentences, not an audit. Fast does not mean kind - a fatal flaw gets named plainly even if everything else gets one sentence. Park concerns that belong to the design phase in a line ("we'll catch that in the grill") rather than raising them now.
This no-panel, no-score, no-gate, user-driven rule holds for the ENTIRE speedrun - the brand pass, the PRD, and the grill included. Every sub-skill you invoke below runs in this unranked, user-driven mode: ignore its panel, tl_score, 0.85-gate, disqualifier-loop, and "categories covered" stop machinery, and let the user's "move on" end each step. The one exception is the grill's metric-honesty rule, kept below - it is cheap and load-bearing.
One stage per turn, in order, waiting for the user between stages. Record each answer to the shared state file as you go (op answer; see "Persisting"). The question ids match the web app exactly, so a speedrun loads straight into it. The spine deliberately skips the framework's domain-knowledge, time, costs, scale, acquisition, relationships, and support stages; those stay blank by design, so a loaded file that reads as partially complete is expected, not an error. If a skipped stage holds something that genuinely blocks the spec, raise it in one line instead of re-opening the stage.
what-statement) - "One sentence: a thing that does what, for whom." Clear and specific, not the pitch.paid-today, evidence) - "Has anyone paid you to solve this, or what is the strongest signal they will?" Liking is not buying.target-market, incumbent-gap) - "Who specifically is this for, and why won't an incumbent just copy it?" Start with the smallest market you can dominate.pricing-model) - "What is the number, and can you defend it in one sentence without apologizing?"bm-who-buys, bm-who-supplies) - "Name the price, a plausible customer count, and the main monthly costs as real numbers: does revenue clear cost, and roughly when?" The honest read is whether those are real estimates or hopeful blanks. Use the tl_project tool for a quick projection and store it as the bizModel artifact if useful. (bm-parties is intentionally left for the full mode, so the /tl upgrade stays lossless.)After stage 5, before the design phase, branch on brand.
Ask the user once: "Want to name it and sketch a quick brand before we spec it?"
tl_domains tool) - for its stages and questions only: suppress its panel, tl_score, 0.85 gate, and disqualifier loop; one fast unranked read per brand step, and the user moves on when they like a direction. A speedrun brand is a directional sketch off an unvalidated idea - re-test it under /tl before you print it on anything. Its output feeds the PRD's identity and UX notes and persists as the naming and brand artifacts.The speedrun gets you all the way to a real spec, fast. The two design steps are not optional.
prd artifact.prd artifact when done.What you get is a fast first draft built on a gut-check, not a validated spec. Re-running the grill under /tl, with the panels behind it, is what turns it into something to bet on.
With the PRD hardened you have a build-ready spec - run the thought-layer-build skill (/tl-build) to build it, or the tl_scaffold tool for an instant deployable landing page. (A speedrun spec is a gut-check; re-grill under /tl before you bet on it.)
Write to the state file as you go, like the full framework (see the thought-layer-framework skill's "Saving and resuming"): answers via op answer, the bizModel / naming / brand / prd / grill via op artifact, and a cursor via op cursor after each stage so the file resumes and upgrades cleanly (use the backbone stage numbers - the spine maps to stages 1, 3, 4, 9, 10, then the PRD is 14 and the grill 15; set phase to speedrun). The default file is .thought-layer/state.json; to keep several ideas apart, pass --path .thought-layer/<name>.json (or the tool's path) on every op, or set THOUGHT_LAYER_STATE, and use list to see what is already there. Do not write graded feedback - the speedrun does not rank, so there are no panel verdicts to store and the answers persist ungraded.
This keeps a speedrun loadable in the web app and upgradable. If you resume a file that already has panel feedback or a /tl cursor, do not discard it: tell the user which stages are already graded and only speedrun the still-open ones. If the user stops mid-spine, the answers so far are already saved - tell them the file location and that they can resume here or upgrade to /tl anytime.
If the user just wants to see how this works, offer to run the whole thing on a sample idea (suggest one, for example a scheduling tool for mobile dog groomers) so they can watch the arc - five fast reads, an optional brand, a PRD, a quick grill - in a couple of minutes. It is still one stage per turn, pausing for the user: demo mode is fast, not unattended. The sample writes a real state.json, so clear it (or run it in a throwaway directory) before the user starts on their own idea.
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.