Grill walk
Walk a live artifact or code-grounded substitute, ratify concrete product findings one at a time, and hand a structured evidence record to dispatch-compiler for proof-gated issue filing and closeout. Use when the user says "grill walk", "design walk", "walk the app/site", "punch list session", "grill #N", or wants agent-ready work from product taste and runtime evidence.From its SKILL.md
npx -y skills add lemun/dispatch-compiler --skill grill-walkAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
5.5 KB, ~1.1k tokens by cl100k_base, as published. Nobody here has run it
Grill Walk
Turn product taste into ratified evidence, then delegate packet generation to
the canonical compiler. This skill owns the walk and the owner conversation.
dispatch-compiler owns lane carving, issue schema, model routing, tracker
mutation, and source closeout.
Non-Negotiables
- Do not fix during the walk. Observe, map, ratify, and record.
- Ratify one consequential product decision at a time.
- Map every finding to concrete territory before it can enter the handoff.
- Preserve vetoes and deliberate non-items so later sessions do not relitigate them.
- Follow the target repository's
AGENTS.mdand selected dispatch profile for evidence location, tracker labels, runtime ownership, and closeout policy. - Do not reproduce issue templates, model names, effort rules, GitHub commands,
or closeout mechanics here. Those belong to
dispatch-compilerand the project profile.
Phase 0 - Contract
If invoked with an issue number, read the current issue first. Treat its body as the charter, but re-verify its scope, blockers, prior ratification, and evidence against current repository state.
Before walking:
- Read the repository contract and select the applicable dispatch profile.
- State the session mode:
owner-ratified,autonomous,code-grounded, orlive-walked. Combine origins when more than one is true. - State the output contract: this phase produces a structured compilation handoff; the compiler decides safe lane shape and tracker routing afterward.
- Choose a durable evidence location allowed by the repository contract.
- If the owner compares the artifact with an elevated vision, apply the recall test: only details they can name are admissible, with at most one or two elevation cherry-picks per session.
- If recommendations depend on research, gather or read attributable research before proposing decisions.
Phase 1 - Walk
- Drive the real artifact when practical: browser, simulator, device, docs, email, store listing, or generated preview. Use realistic locale, theme, authentication state, platform, viewport, and data.
- Walk daily-use surfaces before first-run flows that disturb state. Keep each screen brief: early reactions reveal taste; prolonged staring invents noise.
- Prime each surface with two to four concrete observations to confirm or dismiss, then ask what bothers the owner.
- For every candidate, immediately map files, routes, components, strings, tokens, tests, screenshots, recordings, logs, or artifact paths.
- Record the observed behavior, proposed outcome, buy or veto, evidence origin, confidence, and any deliberate non-item.
- If an entire surface needs judgment, capture obvious mechanical findings and record a separate unresolved owner call instead of forcing a full redesign.
- If live driving stalls, pivot to current code and artifacts. Record the
blocker and mark the origin
code-groundedorartifact-groundedrather than pretending it was live-walked.
Phase 2 - Rank
Rank findings for dispatch usefulness:
- Broken behavior a new user will encounter.
- Systemic problems and daily-surface fixes.
- High-risk correctness involving auth, data, native behavior, payments, or release/runtime compatibility.
- Visual and copy polish, riders, and verify-only checks.
- At most one elevation cherry-pick per session.
Ranking is evidence for the compiler, not a pre-carved lane plan. Record likely sequencing or territory collisions, but let the compiler produce the final file-disjoint issue set. Verify-only observations normally become proof gates inside related work rather than standalone implementation issues.
Phase 3 - Compilation Handoff
Produce this record without omitting empty or blocked sections:
## Compilation handoff
- Source repository and source issue:
- Session mode and evidence origin:
- Overall confidence:
- Project profile:
### Ratified findings
- Finding, expected outcome, confidence, and owner decision:
### Vetoes and deliberate non-items
- Decision and reason:
### Evidence and artifact paths
- Path or URL and what it proves:
### Territory anchors
- Finding -> files, routes, components, strings, tests, or artifacts:
### Ranking and sequencing observations
- Rank, likely collision, or dependency:
### Unresolved owner calls
- Decision still required and concrete options already considered:
### Runtime proof status and blockers
- What was driven, what was not, blocker, and acceptable substitute evidence:
REQUIRED SUB-SKILL: invoke dispatch-compiler in the same task after the
handoff is complete. Pass the handoff intact and name the selected project
profile. The compiler must re-verify current anchors, carve merge-safe lanes,
generate packet-specific Codex and Claude routing, file or update tracker work
as authorized by the repository policy, and close out the source charter when
its close conditions are satisfied.
Do not re-ratify bought decisions during compilation. If current evidence
contradicts a ratified finding, preserve the decision record, stop that finding
from becoming ready, and report the contradiction through the compiler's
normal status and closeout rules.
What ships with it: 1 file
273 B alongside SKILL.md
agents/
- openai.yaml273 B