quest

Name: quest
Author: levi-qiao

28stars2forks

How to install quest?

levi-qiao/graphkit/quest

$npx -y skills add levi-qiao/graphkit --skill quest

Installs into the current project.

Summary

Compile a long-horizon coding task into ONE objective prompt for a host that drives a goal to done with its own verifier (grok /goal, a Codex task) — where the host's harness is the acceptance auditor, so no second supervisor loop is needed. Folds the octopus discipline (verifi

Files · 1

View on GitHub

SKILL.md

1	# quest — one objective, riding the host's own verifier
2
3	## What it does & why
4
5	Some hosts already drive the graph for you. grok's `/goal` plans acceptance
6	criteria, works across rounds, and only marks the goal complete after an
7	independent adversarial verifier reproduces the evidence (defaulting to
8	refuted if it can't). A Codex task self-drives an objective to done across rounds.
9	On such a host, generating a separate executor + supervisor + two loops is
10	redundant — you'd pay tokens to re-describe what the harness enforces for free.
11
12	The quest arm instead emits one objective prompt and lets the host drive it. Its
13	whole value is that the objective carries the octopus discipline the host's generic
14	harness doesn't know — verifiable acceptance criteria, no-test-theater, no
15	speculative building, forced convergence, owner red lines — so the host's verifier
16	has something real to check against.
17
18	## When to use quest vs loop-graph
19
20	Use the quest arm when all hold:
21	- one self-contained goal (no strict-ordered milestones with a non-skippable gate);
22	- the host has a goal command that self-drives to done (grok `/goal`, a **Codex
23	task**);
24	- executor and reviewer live in the same run (you're not splitting them across
25	hosts/models);
26	- no owner-only decision sits on the goal's critical path.
27
28	Use the loop-graph arm (`../loop-graph`) when any hold: multi-milestone phases,
29	executor and supervisor on different hosts/models/cadences, owner red-lines/gates
30	that must stop the run, cross-session durability, or a host that only loops and
31	has no goal command (Claude Code, Cursor, shell) — there loop-graph supplies
32	the verifier as the supervisor node. See `../../lib/host-dialects.md` for which host
33	has what.
34
35	## How to run it
36
37	Step 1 — Interview (fill the blanks, don't assume; surface tradeoffs, don't
38	silently pick). Ask in order — skip only if context already answers:
39
40	1. Host — where does the goal run? (grok `/goal`, a Codex task). This
41	picks the invocation dialect (`../../lib/host-dialects.md`). If the host has no
42	goal command — Claude Code, Cursor, shell — stop and switch to the loop-graph arm.
43	2. Objective — the goal in one sentence.
44	3. Acceptance criteria — the North Star, each as a check the host's verifier can
45	reproduce without the agent: a command that must pass, or an artifact at a
46	persistent path showing a bar met on a named eval set. Push until every criterion
47	is checkable ("`make test` green", "scorecard ≥ X on the frozen set"), not "make
48	it work". This is the single most important step — grok's verifier defaults to
49	refuted on anything it can't reproduce, and a Codex task has no separate refuter
50	at all, so the criteria are your only guardrail there.
51	4. Red lines — the non-negotiables that halt the work (push/commit auth,
52	destructive ops, secrets/real data, frozen contracts, metrics-only-go-up).
53	5. Expensive-op guard — is there a full-cohort eval / bulk sweep / migration? If
54	so, it pilots first.
55
56	Milestone check. If the answers reveal strict-ordered phases with a gate, or an
57	owner-only call on the critical path, say so and switch to the loop-graph arm — the
58	quest arm deliberately has no gate and no owner-stop loop.
59
60	Step 2 — Fill the template. Copy `templates/quest.md`, replace every
61	`{{PLACEHOLDER}}`, delete the guidance comments. Interview in the user's language and
62	mirror it in the prose; keep structural headings as in the template.
63
64	Step 3 — Deliver. Hand the filled objective to the host's goal command, and tell
65	the user how in the chat (print it, don't assume this session is the host):
66
67	- grok: `/goal <the filled objective>` — optionally `--budget <tokens>`. Manage
68	with `/goal status` · `pause` · `resume` · `clear`.
69	- Codex: delegate a task with the filled objective as its brief; it self-drives
70	across rounds. Since there's no separate refuter, the objective's acceptance
71	criteria carry the whole verification burden — make them reproducible.
72
73	Optional reliability upgrade on grok (from `../../lib/host-dialects.md`): a `Stop`
74	hook to hold the turn open until a gate passes; a `Notification` hook to ping the
75	owner when the goal parks on a real blocker.
76
77	Durability note. A goal harness's state is ephemeral (grok's scratch dir is
78	deleted when th

People who liked this also viewed

6 picks

Type

quest

28stars2forks

View on GitHub

How to install quest?

levi-qiao/graphkit/quest

$npx -y skills add levi-qiao/graphkit --skill quest

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# quest — one objective, riding the host's own verifier
2
3	## What it does & why
4
5	Some hosts already drive the graph for you. grok's `/goal` plans acceptance
6	criteria, works across rounds, and only marks the goal complete after an
7	independent adversarial verifier reproduces the evidence (defaulting to
8	refuted if it can't). A Codex task self-drives an objective to done across rounds.
9	On such a host, generating a separate executor + supervisor + two loops is
10	redundant — you'd pay tokens to re-describe what the harness enforces for free.
11
12	The quest arm instead emits one objective prompt and lets the host drive it. Its
13	whole value is that the objective carries the octopus discipline the host's generic
14	harness doesn't know — verifiable acceptance criteria, no-test-theater, no
15	speculative building, forced convergence, owner red lines — so the host's verifier
16	has something real to check against.
17
18	## When to use quest vs loop-graph
19
20	Use the quest arm when all hold:
21	- one self-contained goal (no strict-ordered milestones with a non-skippable gate);
22	- the host has a goal command that self-drives to done (grok `/goal`, a **Codex
23	task**);
24	- executor and reviewer live in the same run (you're not splitting them across
25	hosts/models);
26	- no owner-only decision sits on the goal's critical path.
27
28	Use the loop-graph arm (`../loop-graph`) when any hold: multi-milestone phases,
29	executor and supervisor on different hosts/models/cadences, owner red-lines/gates
30	that must stop the run, cross-session durability, or a host that only loops and
31	has no goal command (Claude Code, Cursor, shell) — there loop-graph supplies
32	the verifier as the supervisor node. See `../../lib/host-dialects.md` for which host
33	has what.
34
35	## How to run it
36
37	Step 1 — Interview (fill the blanks, don't assume; surface tradeoffs, don't
38	silently pick). Ask in order — skip only if context already answers:
39
40	1. Host — where does the goal run? (grok `/goal`, a Codex task). This
41	picks the invocation dialect (`../../lib/host-dialects.md`). If the host has no
42	goal command — Claude Code, Cursor, shell — stop and switch to the loop-graph arm.
43	2. Objective — the goal in one sentence.
44	3. Acceptance criteria — the North Star, each as a check the host's verifier can
45	reproduce without the agent: a command that must pass, or an artifact at a
46	persistent path showing a bar met on a named eval set. Push until every criterion
47	is checkable ("`make test` green", "scorecard ≥ X on the frozen set"), not "make
48	it work". This is the single most important step — grok's verifier defaults to
49	refuted on anything it can't reproduce, and a Codex task has no separate refuter
50	at all, so the criteria are your only guardrail there.
51	4. Red lines — the non-negotiables that halt the work (push/commit auth,
52	destructive ops, secrets/real data, frozen contracts, metrics-only-go-up).
53	5. Expensive-op guard — is there a full-cohort eval / bulk sweep / migration? If
54	so, it pilots first.
55
56	Milestone check. If the answers reveal strict-ordered phases with a gate, or an
57	owner-only call on the critical path, say so and switch to the loop-graph arm — the
58	quest arm deliberately has no gate and no owner-stop loop.
59
60	Step 2 — Fill the template. Copy `templates/quest.md`, replace every
61	`{{PLACEHOLDER}}`, delete the guidance comments. Interview in the user's language and
62	mirror it in the prose; keep structural headings as in the template.
63
64	Step 3 — Deliver. Hand the filled objective to the host's goal command, and tell
65	the user how in the chat (print it, don't assume this session is the host):
66
67	- grok: `/goal <the filled objective>` — optionally `--budget <tokens>`. Manage
68	with `/goal status` · `pause` · `resume` · `clear`.
69	- Codex: delegate a task with the filled objective as its brief; it self-drives
70	across rounds. Since there's no separate refuter, the objective's acceptance
71	criteria carry the whole verification burden — make them reproducible.
72
73	Optional reliability upgrade on grok (from `../../lib/host-dialects.md`): a `Stop`
74	hook to hold the turn open until a gate passes; a `Notification` hook to ping the
75	owner when the goal parks on a real blocker.
76
77	Durability note. A goal harness's state is ephemeral (grok's scratch dir is
78	deleted when th

People who liked this also viewed

6 picks

Type