self-consistency-check

Name: self-consistency-check
Author: UnpaidAttention

84stars18forks

How to install self-consistency-check?

unpaidattention/fable5-methodology/self-consistency-check

$npx -y skills add UnpaidAttention/fable5-methodology --skill self-consistency-check

Installs into the current project.

Summary

Force the combination step that a weaker model can't do in one pass — surface conflicts between constraints, and interactions between requirements and edge cases, using a pairwise sweep, a fresh-context cold read by a subagent that never saw your reasoning, and N-version divergen

Files · 1

View on GitHub

SKILL.md

1	# Self-Consistency Check
2
3	Writing every constraint and edge case down makes a cross-term conflict visible; it does not
4	make it noticed. A weaker model can read a complete list and still miss that requirement 3
5	breaks edge case 7 — because noticing the interaction requires holding both in mind at once,
6	which is exactly the single-pass depth that doesn't transfer. This skill replaces that depth
7	with procedure: enumeration (which a weaker model does reliably) forced into combination
8	(which it doesn't do spontaneously). Run it on a plan/design/spec before you build.
9
10	## Technique 1: Pairwise constraint sweep (always, for >3 constraints)
11
12	1. List every constraint, requirement, and known edge case and number them C1..Cn. Pull them
13	from the spec, the plan, and your assumptions log — don't rely on memory.
14	2. Build the grid. Down the side: each constraint. Across the top: each other constraint it
15	could plausibly interact with, plus each edge case. For every plausible pair (Ci, Cj), write
16	one of: compatible (one clause why) or CONFLICT (name it).
17	3. Rules that make it real, not ritual:
18	- Do not skip a pair because it "feels fine" — that feeling is the miss you're hunting.
19	- Bound it to plausible interactions, not a blind N² (that's ceremony); but when unsure
20	whether a pair interacts, include it — the cost of a filled cell is one line.
21	- Every CONFLICT is either resolved in the design now, or logged as an explicit open
22	decision. A found-but-unresolved conflict is still a win — it's a bug caught pre-code.
23
24	## Technique 2: Fresh-context cold read (for interacting/high-stakes designs)
25
26	You miss the conflict because you're anchored on the path that produced the design. Someone who
27	never walked that path is structurally better positioned to see it.
28
29	1. Delegate to a subagent. Give it ONLY the artifacts — the spec and the plan/design.
30	2. Withhold your reasoning trace. Do not tell it what you were thinking or why; that's the
31	anchor you're trying to escape. Hand it the reasoning and it inherits your blind spot.
32	3. Ask exactly: "Which two statements here cannot both be true? Which requirement has no
33	handling for which edge case? What does this design assume that it never states?"
34	4. Treat what it finds as findings to resolve, not opinions to weigh.
35
36	## Technique 3: N-version divergence (for the hard kernel)
37
38	Parallelism substitutes for depth on the one part that matters most.
39
40	1. Identify the hard kernel (problem-framing / §15.1) — the ~20% that decides the outcome.
41	2. Spawn 2–3 subagents to solve just the kernel independently — they must not see each
42	other's attempts.
43	3. Diff the results. Where they agree, you're probably safe. Where they diverge is the
44	load-bearing, uncertain decision — the exact spot a single pass would have committed to
45	blindly. Investigate every divergence; often one branch honored a constraint the others
46	dropped, which tells you both the right answer and which constraint was slippery.
47
48	## Worked example
49
50	Design: a "share link" feature. Constraints gathered: C1 links expire after 7 days; C2 links are
51	revocable instantly; C3 access is logged for audit; C4 anonymous (unauthenticated) users may
52	open a link; C5 GDPR — a user can delete their account and all associated PII.
53
54	- Sweep finds what the list alone hid: (C3 × C5) — audit logs record who accessed a link,
55	but account deletion (C5) must purge PII, and access logs of the deleted user's links are
56	PII → CONFLICT: the audit-retention policy and the deletion guarantee are unspecified together.
57	(C4 × C3) — anonymous access logged how? no user id exists → what does "who accessed" even
58	mean → underspecified. Neither conflict is visible staring at C1–C5 as a list; both surface
59	the instant you fill the (i,j) cells.
60	- Cold read subagent, given only the spec, asks: "C2 says instantly revocable, C1 says
61	7-day expiry — is a revoked link's record kept until day 7 or purged immediately? The spec
62	never says." A third gap, caught by the un-anchored reader.
63	- Outcome: three interaction gaps found and resolved (or logged

People who liked this also viewed

6 picks

Type

self-consistency-check

84stars18forks

View on GitHub

How to install self-consistency-check?

unpaidattention/fable5-methodology/self-consistency-check

$npx -y skills add UnpaidAttention/fable5-methodology --skill self-consistency-check

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Self-Consistency Check
2
3	Writing every constraint and edge case down makes a cross-term conflict visible; it does not
4	make it noticed. A weaker model can read a complete list and still miss that requirement 3
5	breaks edge case 7 — because noticing the interaction requires holding both in mind at once,
6	which is exactly the single-pass depth that doesn't transfer. This skill replaces that depth
7	with procedure: enumeration (which a weaker model does reliably) forced into combination
8	(which it doesn't do spontaneously). Run it on a plan/design/spec before you build.
9
10	## Technique 1: Pairwise constraint sweep (always, for >3 constraints)
11
12	1. List every constraint, requirement, and known edge case and number them C1..Cn. Pull them
13	from the spec, the plan, and your assumptions log — don't rely on memory.
14	2. Build the grid. Down the side: each constraint. Across the top: each other constraint it
15	could plausibly interact with, plus each edge case. For every plausible pair (Ci, Cj), write
16	one of: compatible (one clause why) or CONFLICT (name it).
17	3. Rules that make it real, not ritual:
18	- Do not skip a pair because it "feels fine" — that feeling is the miss you're hunting.
19	- Bound it to plausible interactions, not a blind N² (that's ceremony); but when unsure
20	whether a pair interacts, include it — the cost of a filled cell is one line.
21	- Every CONFLICT is either resolved in the design now, or logged as an explicit open
22	decision. A found-but-unresolved conflict is still a win — it's a bug caught pre-code.
23
24	## Technique 2: Fresh-context cold read (for interacting/high-stakes designs)
25
26	You miss the conflict because you're anchored on the path that produced the design. Someone who
27	never walked that path is structurally better positioned to see it.
28
29	1. Delegate to a subagent. Give it ONLY the artifacts — the spec and the plan/design.
30	2. Withhold your reasoning trace. Do not tell it what you were thinking or why; that's the
31	anchor you're trying to escape. Hand it the reasoning and it inherits your blind spot.
32	3. Ask exactly: "Which two statements here cannot both be true? Which requirement has no
33	handling for which edge case? What does this design assume that it never states?"
34	4. Treat what it finds as findings to resolve, not opinions to weigh.
35
36	## Technique 3: N-version divergence (for the hard kernel)
37
38	Parallelism substitutes for depth on the one part that matters most.
39
40	1. Identify the hard kernel (problem-framing / §15.1) — the ~20% that decides the outcome.
41	2. Spawn 2–3 subagents to solve just the kernel independently — they must not see each
42	other's attempts.
43	3. Diff the results. Where they agree, you're probably safe. Where they diverge is the
44	load-bearing, uncertain decision — the exact spot a single pass would have committed to
45	blindly. Investigate every divergence; often one branch honored a constraint the others
46	dropped, which tells you both the right answer and which constraint was slippery.
47
48	## Worked example
49
50	Design: a "share link" feature. Constraints gathered: C1 links expire after 7 days; C2 links are
51	revocable instantly; C3 access is logged for audit; C4 anonymous (unauthenticated) users may
52	open a link; C5 GDPR — a user can delete their account and all associated PII.
53
54	- Sweep finds what the list alone hid: (C3 × C5) — audit logs record who accessed a link,
55	but account deletion (C5) must purge PII, and access logs of the deleted user's links are
56	PII → CONFLICT: the audit-retention policy and the deletion guarantee are unspecified together.
57	(C4 × C3) — anonymous access logged how? no user id exists → what does "who accessed" even
58	mean → underspecified. Neither conflict is visible staring at C1–C5 as a list; both surface
59	the instant you fill the (i,j) cells.
60	- Cold read subagent, given only the spec, asks: "C2 says instantly revocable, C1 says
61	7-day expiry — is a revoked link's record kept until day 7 or purged immediately? The spec
62	never says." A third gap, caught by the un-anchored reader.
63	- Outcome: three interaction gaps found and resolved (or logged

People who liked this also viewed

6 picks

Type