Go goal go
Help users write /goal objectives as concise natural-language prose in their own language, with a built-in Reviewer sub-agent check. Triggers on user requests in any language asking to write a goal.From its SKILL.md
npx -y skills add xiehuacheng/skills --skill go-goal-goAssembled 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
4.9 KB, ~1.1k tokens by cl100k_base, as published. Nobody here has run it
go-goal-go
A /goal plan is prose in the user's language, plus one <reviewer> block for independent verification. One standard shape, no tier picking, no dense parameter forms.
Rules
- Drafting is read-only until
CreateGoalis called. All shipping actions need explicit approval. - The Reviewer is a separate sub-agent (not the main loop). On
FAILroute back to the main loop iteration, decrement remaining budget; do not declare complete. On Reviewer dispatch error (timeout / OOM / refused / empty) surface to the user as a soft pause with the captured partial output, never as a code change. - Default Reviewer role:
verifier. Default budget: 50 iterations / 30 min wall-clock / $2 / 5-no-progress / 20% reserve for handoff. Budget exhausted terminates with a cost summary; never claimPASS. - Default boundaries: project-only paths, write allowed in-scope, deny-all network.
- A
<reviewer>block must be preceded by a one-line dispatch mandate naming the sub-agent and forbidding in-place self-judging. Seereferences/reviewer-template.mdfor canonical phrasings. - Verdict parsing:
^(PASS|FAIL:.{1,280})$on the last non-blank line. Anything else =REVIEW_INFRA_ERROR.
When to use / not
Use when the user wants to write a goal or describes a multi-step task that could run unattended. Skip for single-turn questions, exploration with no finish line, or any task where neither proof nor budget can be named.
Workflow
- Fit check. Queue-shaped + verifiable + bounded + repeatable + has-budget. At least 3 of 5, or push back.
- Confirm intent. For each of boundary / budget / Reviewer role: if the user has already named a value, echo back for confirmation; otherwise ask. Defaults above.
- Draft. Write the prose goal plan in the user's language using the standard shape below, including the dispatch mandate +
<reviewer>block. - Revise. Show, walk through, repeat until approved.
CreateGoalonly after explicit approval.
The standard shape
A goal plan is prose covering these sections, in the user's language, plus one <reviewer> block:
- Goal — what must become true.
- End state — one observable predicate the host can check.
- Proof — the command plus 1-2 supporting signals (e.g. "no new skipped tests", "no assertion weakening").
- Boundaries — paths the loop may touch; network/side effects in prose ("no network", "no installs", "no API calls").
- Budget — caps on iterations / wall-clock / cost / stall / reserve (last optional, default 20%).
- Stop conditions — hard stops (terminate) vs. soft stops (pause + ask).
- Reviewer —
[dispatch mandate]then<reviewer>block with 4 sections (Role, Inputs, Acceptance, Verdict).
Worked example (English)
Goal: stabilize 5 intermittently failing tests under test/auth so every rerun is solid.
End state: npm test --testPathPattern=test/auth --runInBand exits 0 three times in a row; all 5 tests pass; no new .skip / describe.skip / xit.
Proof: run that command; exit 0 + no new skipped tests + all 5 test IDs in passing list across the 3 runs.
Boundaries: only files under src/auth/ and test/auth/; no installs, no network, no test-runner config edits, no deleting or disabling tests.
Budget: 50 iterations / 20 min / $1 max; stop after 5 stalled iterations; 20% reserve.
Stop conditions:
- write outside in-scope paths: terminate and report.
- need to edit jest/vitest config, tsconfig, or CI workflow: stop and ask.
- user interrupt: graceful shutdown, persist checkpoint.
Reviewer: dispatch an independent sub-agent with the prompt below; the main loop must not self-judge this in place.
<reviewer> Role: criticInputs:
- final diff vs base
- the 3 proof runs' full output
- iterations log
Acceptance:
- all 5 tests pass across 3 runs; skipped = 0.
- no new
.skip/describe.skip/xit/xdescribeetc. - all changed files under
src/auth/ortest/auth/. - at least one iteration is a real root-cause change.
- no assertion weakening (
toEqual→toMatch,.resolvesfallback, nestedsetTimeoutretries, fake timers).
Verdict: last line of output must be exactly PASS or FAIL: <one-line reason>.
</reviewer>
References
references/reviewer-template.md— dispatch mandate canonical phrasings,<reviewer>block schema, host-side defaults.references/contract-glossary.md— prose checklist behind each paragraph (for tooling that auto-generates plans).references/example-zh.md— same example in Chinese.references/host-internals.md— index of files only host implementers need (dispatch-primitive.md,capability-boundary.md).
What ships with it: 6 files
22.7 KB alongside SKILL.md
references/
- capability-boundary.md5.1 KB
- contract-glossary.md4.6 KB
- dispatch-primitive.md6.2 KB
- example-zh.md1.8 KB
- host-internals.md941 B
- reviewer-template.md3.9 KB