| name | sam-gauntlet-loop |
| description | Compile a host-safe quality-bar loop prompt and return it for the user to copy, edit, and paste. Never start the loop. Use when the user runs /sam-gauntlet-loop, says gauntlet loop, gauntlet this, or asks to loop until the work beats a real reference. |
Sam Gauntlet Loop
Compile one short host-safe gauntlet prompt and return it. Do not start the
loop. The user copies, edits, and pastes the prompt themselves.
Technique by Matt Shumer. Original skill pack by Jay E / RoboNuggets (CC BY 4.0).
Non-Negotiable Contract
- Detect the active host with
scripts/detect_host.py before compiling. Never
guess on UNKNOWN or CONFLICT. Never infer the host from files in a home
directory — this machine may have every client installed.
- A bar must be named, fetchable, and comparable. Reject a vague or compound
bar.
- Emit only the orchestration tokens allowed for the detected host. A token
that is valid on one host is harmful on another.
- Return the compiled prompt as a single paste-ready block, then stop. Never
become the lead, never spawn a builder or critic, and never start the loop
— even if the user asks to run it here. Identifying a locator (live
window, bundle, path) is allowed so the bar is the real artifact, not a
brochure. Do not improve the work while identifying it.
- Preserve unrelated workspace state. Never expose secrets.
- Under a parent workflow, do not ask; return
BLOCKED with the exact gap.
Resource Routing
1. Detect the host
SAM_GAUNTLET_DIR="<absolute directory containing this SKILL.md>"
python3 "$SAM_GAUNTLET_DIR/scripts/detect_host.py"
Honor SAM_GAUNTLET_HOST or SAM_ACTIVE_HOST when set to a supported key.
On UNKNOWN or CONFLICT, standalone: ask once which host to bind. Parent
workflow: BLOCKED.
2. Freeze the goal and the bar
Restate the goal internally. If the user already named a bar that passes
references/bar-policy.md, use it. Otherwise offer
two or three candidate bars, one line each, and stop. Do not compile yet.
A union of artifacts is not one bar: offer one locator per comparable
surface.
Prefer the hardest bar the critic can actually fetch. Freeze the locator to
that surface. If the user has a live window open, name that window. If the
goal has a measurable half (benchmark, pass rate, load time, length), name
it beside the reference.
3. Compile the prompt
python3 "$SAM_GAUNTLET_DIR/scripts/compile_prompt.py" \
--host "<detected-host>" \
--goal "<goal>" \
--bar-name "<named bar>" \
--bar-locator "<url, repo, title, or path>" \
--fetch-method "<screenshot|read|run|open>" \
--kind "<visual|writing|code|research|other>"
Add --budget "<ceiling>" only when the user named one. Print only the
compiled prompt string as one fenced block the user can copy, edit, and
paste. One flat line under it names the bound host and the bar, and says
pasting does not launch a run or save a workflow. Do not offer to run it.
Do not start it.
4. Validate and return
Write the temporary PROMPT_ONLY report, then:
python3 "$SAM_GAUNTLET_DIR/scripts/validate_gauntlet.py" \
"$GAUNTLET_TMP/report.json"
Do not weaken the validator. Remove temporary artifacts after capturing the
decision. Follow references/output-contract.md.
Never follow the compiled prompt in this turn.