extended-problem-solving

Name: extended-problem-solving
Author: UnpaidAttention

84stars18forks

How to install extended-problem-solving?

unpaidattention/fable5-methodology/extended-problem-solving

$npx -y skills add UnpaidAttention/fable5-methodology --skill extended-problem-solving

Installs into the current project.

Summary

Externalize reasoning to a scratchpad file for problems too large to hold in working memory — a persistent record of candidate approaches, eliminated branches and WHY they were eliminated, open questions, and current best hypothesis, updated as understanding evolves. Trigger this

Files · 1

View on GitHub

SKILL.md

1	# Extended Problem Solving
2
3	Some problems have more branches than working memory has slots. Held in-context, they cause the
4	signature failures: re-exploring a branch you already killed, half-remembering why you rejected
5	something, and losing the thread across a context break. The fix is a scratchpad on disk that
6	IS your working memory — you offload to it and re-derive from it rather than from recollection.
7
8	## Step 1: Open the scratchpad immediately
9
10	Create `REASONING.md` (scratchpad dir or repo root; not committed unless asked). Structure —
11	all sections present from the start:
12
13	```markdown
14	# REASONING — <problem>
15
16	## Problem statement
17	<the actual question, precisely; refine it as it sharpens>
18
19	## Current best hypothesis
20	<the leading answer/approach right now — one paragraph, updated as it changes>
21
22	## Candidate approaches / branches
23	- [ACTIVE] B1: <approach> — status, what's promising
24	- [KILLED] B2: <approach> — ELIMINATED because <specific reason/evidence>
25	- [OPEN] B3: <approach> — not yet explored
26
27	## Confirmed facts
28	<things established with evidence — each with how it was confirmed>
29
30	## Open questions
31	<what's unresolved, ranked by how much it would unblock>
32
33	## Decisions
34	<choices made and why>
35	```
36
37	## Step 2: Record eliminations with their reason — always
38
39	The highest-value entries are the KILLED branches. When you rule something out, write WHY with
40	the specific evidence: "B2 (in-memory cache) killed — 4 nodes, state must be shared; confirmed
41	per-node state allows 4× the limit." Without the reason on disk, you WILL wander back into a
42	dead branch an hour later and re-walk it. The reason is what makes the elimination permanent.
43
44	## Step 3: Update as understanding evolves
45
46	- New fact confirmed → Confirmed facts, with how.
47	- Branch explored → flip its status, record the outcome.
48	- Best hypothesis changes → rewrite that section (don't just append; the current best must be
49	unambiguous at a glance).
50	- Update at every meaningful move, not in one batch at the end — the scratchpad is only a
51	memory if it's current.
52
53	## Step 4: Re-derive, don't half-remember, on return
54
55	When you come back to a branch (after a context break, a compaction, or switching away and
56	back): RE-READ its scratchpad entry and re-derive from the confirmed facts. Do not proceed from
57	a fuzzy recollection of "where I was" — that recollection is exactly what's unreliable at this
58	scale, and trusting it is how contradictions creep in. The file is authoritative over memory
59	wherever they disagree.
60
61	## Step 5: Stop conditions — recognize grind and switch
62
63	Watch for these and change strategy rather than grinding:
64	- Circular reasoning: you're revisiting a branch whose KILLED entry you wrote yourself.
65	Stop — it's already answered; read the reason.
66	- Diminishing returns: the last three updates added detail but didn't move the best
67	hypothesis or close an open question. The current strategy is exhausted — switch technique
68	(backward reasoning, minimal case, or get more evidence) or escalate to the user.
69	- All branches killed: the problem as framed has no solution → the Problem statement or an
70	assumption is wrong. Re-examine the framing, not the branches.
71	- Thrashing: best hypothesis flipping back and forth between two options → you're missing a
72	distinguishing fact. Name it in Open questions and go get THAT, instead of re-arguing.
73
74	## Worked example
75
76	Problem: intermittent data corruption in a pipeline, cause unknown, spans ingestion →
77	transform → store across three services.
78
79	- Scratchpad opened. Branches: B1 ingestion race, B2 transform off-by-one, B3 store
80	serialization, B4 clock/ordering.
81	- B2 explored → KILLED: "transform is pure, property-tested over 10k inputs, no corruption
82	reproduced — eliminated with evidence." (Recorded so it's never re-opened.)
83	- Confirmed fact: "corruption only on records processed within 100ms of each other — confirmed
84	from log correlation." This kills B2/B3 further and promotes B1.
85	- Best hypothesis rewritten to B1 (concurrent ingestion). Open question: "does the ingestion
86	queue guarantee per-key ordering?" — ranked top, most unblocking.
87	- Context compaction happens. On return: re-read

People who liked this also viewed

6 picks

Type

extended-problem-solving

84stars18forks

View on GitHub

How to install extended-problem-solving?

unpaidattention/fable5-methodology/extended-problem-solving

$npx -y skills add UnpaidAttention/fable5-methodology --skill extended-problem-solving

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Extended Problem Solving
2
3	Some problems have more branches than working memory has slots. Held in-context, they cause the
4	signature failures: re-exploring a branch you already killed, half-remembering why you rejected
5	something, and losing the thread across a context break. The fix is a scratchpad on disk that
6	IS your working memory — you offload to it and re-derive from it rather than from recollection.
7
8	## Step 1: Open the scratchpad immediately
9
10	Create `REASONING.md` (scratchpad dir or repo root; not committed unless asked). Structure —
11	all sections present from the start:
12
13	```markdown
14	# REASONING — <problem>
15
16	## Problem statement
17	<the actual question, precisely; refine it as it sharpens>
18
19	## Current best hypothesis
20	<the leading answer/approach right now — one paragraph, updated as it changes>
21
22	## Candidate approaches / branches
23	- [ACTIVE] B1: <approach> — status, what's promising
24	- [KILLED] B2: <approach> — ELIMINATED because <specific reason/evidence>
25	- [OPEN] B3: <approach> — not yet explored
26
27	## Confirmed facts
28	<things established with evidence — each with how it was confirmed>
29
30	## Open questions
31	<what's unresolved, ranked by how much it would unblock>
32
33	## Decisions
34	<choices made and why>
35	```
36
37	## Step 2: Record eliminations with their reason — always
38
39	The highest-value entries are the KILLED branches. When you rule something out, write WHY with
40	the specific evidence: "B2 (in-memory cache) killed — 4 nodes, state must be shared; confirmed
41	per-node state allows 4× the limit." Without the reason on disk, you WILL wander back into a
42	dead branch an hour later and re-walk it. The reason is what makes the elimination permanent.
43
44	## Step 3: Update as understanding evolves
45
46	- New fact confirmed → Confirmed facts, with how.
47	- Branch explored → flip its status, record the outcome.
48	- Best hypothesis changes → rewrite that section (don't just append; the current best must be
49	unambiguous at a glance).
50	- Update at every meaningful move, not in one batch at the end — the scratchpad is only a
51	memory if it's current.
52
53	## Step 4: Re-derive, don't half-remember, on return
54
55	When you come back to a branch (after a context break, a compaction, or switching away and
56	back): RE-READ its scratchpad entry and re-derive from the confirmed facts. Do not proceed from
57	a fuzzy recollection of "where I was" — that recollection is exactly what's unreliable at this
58	scale, and trusting it is how contradictions creep in. The file is authoritative over memory
59	wherever they disagree.
60
61	## Step 5: Stop conditions — recognize grind and switch
62
63	Watch for these and change strategy rather than grinding:
64	- Circular reasoning: you're revisiting a branch whose KILLED entry you wrote yourself.
65	Stop — it's already answered; read the reason.
66	- Diminishing returns: the last three updates added detail but didn't move the best
67	hypothesis or close an open question. The current strategy is exhausted — switch technique
68	(backward reasoning, minimal case, or get more evidence) or escalate to the user.
69	- All branches killed: the problem as framed has no solution → the Problem statement or an
70	assumption is wrong. Re-examine the framing, not the branches.
71	- Thrashing: best hypothesis flipping back and forth between two options → you're missing a
72	distinguishing fact. Name it in Open questions and go get THAT, instead of re-arguing.
73
74	## Worked example
75
76	Problem: intermittent data corruption in a pipeline, cause unknown, spans ingestion →
77	transform → store across three services.
78
79	- Scratchpad opened. Branches: B1 ingestion race, B2 transform off-by-one, B3 store
80	serialization, B4 clock/ordering.
81	- B2 explored → KILLED: "transform is pure, property-tested over 10k inputs, no corruption
82	reproduced — eliminated with evidence." (Recorded so it's never re-opened.)
83	- Confirmed fact: "corruption only on records processed within 100ms of each other — confirmed
84	from log correlation." This kills B2/B3 further and promotes B1.
85	- Best hypothesis rewritten to B1 (concurrent ingestion). Open question: "does the ingestion
86	queue guarantee per-key ordering?" — ranked top, most unblocking.
87	- Context compaction happens. On return: re-read

People who liked this also viewed

6 picks

Type