five-whys

Name: five-whys
Author: tkellogg

83stars19forks

How to install five-whys?

tkellogg/open-strix/five-whys

$npx -y skills add tkellogg/open-strix --skill five-whys

Installs into the current project.

Summary

Structured root cause analysis for arriving at a concrete action. Use when something went wrong, a pattern keeps recurring, behavior has drifted, or you catch yourself resolving to "do better" / "remember to X" without a concrete artifact. Five-whys forces behavioral resolutions

Files · 1

View on GitHub

SKILL.md

1	# Five Whys
2
3	Five Whys is a method for finding root causes. You ask "why did this happen?" and
4	answer it honestly. Then you ask "why?" about that answer. You keep going until you
5	hit bedrock — the structural cause that, if changed, would prevent the problem from
6	recurring.
7
8	The goal is understanding, not blame. You're investigating a system (which
9	includes yourself). Systems have structural properties that produce outcomes. Find
10	the structural property, change it, and the outcome changes.
11
12	## When to Use This
13
14	- Something failed or went wrong and the surface-level explanation feels incomplete
15	- A pattern keeps recurring despite previous fixes
16	- You're about to make a change and want to make sure you're fixing the right thing
17	- Behavior (yours or a system's) has drifted from expectations
18	- Someone asks "why does this keep happening?"
19
20	## When NOT to Use This
21
22	- The cause is obvious and singular (typo, wrong config value, missing file)
23	- You're troubleshooting a known bug with a known fix
24	- The problem is "how do I do X?" not "why did X go wrong?"
25
26	## The Process
27
28	### Step 1: State the Problem
29
30	Write down what happened. Be specific. Not "it broke" but "the classifier produced
31	11 consecutive errors between iterations 83-93, spending 57 minutes with zero
32	successful runs."
33
34	The problem statement is the root of your tree.
35
36	### Step 2: Ask Why — And Answer It
37
38	For the problem (or any node in the tree), ask: "Why did this happen?"
39
40	Then answer it. Both the question and the answer matter equally. A question without
41	an answer is incomplete. An answer without a clear question is unanchored.
42
43	```
44	PROBLEM: 11 consecutive classifier errors, 57 min wasted
45	WHY: The proposer suggested CatBoost, which isn't in the allowed algorithm list
46	ANSWER: The proposer has no documentation of which algorithms are available
47	```
48
49	Rules for good answers:
50	- An answer must be factual and verifiable. Check logs, read code, look at data.
51	Don't speculate — investigate.
52	- If you can't verify an answer, say so. "I believe X but haven't confirmed" is
53	honest. "X" stated as fact when you haven't checked is not.
54	- An answer should be a mechanism, not a redescription. "It failed because it
55	was broken" is circular, not explanatory.
56
57	### Step 3: Go Deeper — Recursively
58
59	Take each answer and ask "why?" again. Keep going until you reach one of:
60
61	- A structural property of the system that can be changed (a missing
62	constraint, a bad default, an absent check)
63	- An external boundary you don't control (a vendor API, a hardware limit, a
64	policy)
65	- A known and accepted tradeoff that was made intentionally
66
67	That's bedrock. Stop there.
68
69	How to know you've hit bedrock: If you ask "why?" one more time and the answer
70	is either "that's how the universe works" or "because we chose to" — you're there.
71
72	How to know you haven't: If you can still point to a structural property that
73	someone could change — keep going.
74
75	### Step 4: The Tree Is Fractal
76
77	This is critical. At any single level, the answer to "why?" often has **two or
78	three independent causes*. That's not a problem — it's better*. Branch the tree.
79
80	```
81	PROBLEM: Agent was unresponsive for 62 minutes
82	│
83	├─ WHY-1: The LLM API call hung and never returned
84	│ └─ WHY: No turn-level timeout exists in the harness
85	│ └─ WHY: The harness assumes LLM calls always complete
86	│ → BEDROCK: Missing timeout (structural — fixable)
87	│
88	├─ WHY-2: Agent was already degraded from sync sleep-polling
89	│ └─ WHY: Agent doesn't trust the async callback mechanism
90	│ └─ WHY: Past flaky experiences created a hedging pattern
91	│ └─ WHY: No positive evidence of reliability in memory
92	│ → BEDROCK: Missing trust documentation (structural — fixable)
93	│
94	└─ WHY-3: The LLM endpoint was unstable all day
95	└─ WHY: Another team member deleted the deployment
96	→ BEDROCK: Shared resource, no access control (external boundary)
97	```
98
99	Multiple root causes are the norm, not the exception. Most problems are
100	overdetermined — several things conspired. Finding all of them matters because
101	fixing only one may not prevent recurrence.
102
103	Don't force your tree into a single chain. If you find yourself writing "and also"
104	— that's a second branch, not a conjunction.
105
106	### Step 5: Action Items
107
108	Every bedrock node should produce an action item. Here's what separates useful
109	actions from theater:
110
111	An action item must produce a concrete artifact. It must result in a file edit,
112	a config change

People who liked this also viewed

6 picks

Type

five-whys

83stars19forks

View on GitHub

How to install five-whys?

tkellogg/open-strix/five-whys

$npx -y skills add tkellogg/open-strix --skill five-whys

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Five Whys
2
3	Five Whys is a method for finding root causes. You ask "why did this happen?" and
4	answer it honestly. Then you ask "why?" about that answer. You keep going until you
5	hit bedrock — the structural cause that, if changed, would prevent the problem from
6	recurring.
7
8	The goal is understanding, not blame. You're investigating a system (which
9	includes yourself). Systems have structural properties that produce outcomes. Find
10	the structural property, change it, and the outcome changes.
11
12	## When to Use This
13
14	- Something failed or went wrong and the surface-level explanation feels incomplete
15	- A pattern keeps recurring despite previous fixes
16	- You're about to make a change and want to make sure you're fixing the right thing
17	- Behavior (yours or a system's) has drifted from expectations
18	- Someone asks "why does this keep happening?"
19
20	## When NOT to Use This
21
22	- The cause is obvious and singular (typo, wrong config value, missing file)
23	- You're troubleshooting a known bug with a known fix
24	- The problem is "how do I do X?" not "why did X go wrong?"
25
26	## The Process
27
28	### Step 1: State the Problem
29
30	Write down what happened. Be specific. Not "it broke" but "the classifier produced
31	11 consecutive errors between iterations 83-93, spending 57 minutes with zero
32	successful runs."
33
34	The problem statement is the root of your tree.
35
36	### Step 2: Ask Why — And Answer It
37
38	For the problem (or any node in the tree), ask: "Why did this happen?"
39
40	Then answer it. Both the question and the answer matter equally. A question without
41	an answer is incomplete. An answer without a clear question is unanchored.
42
43	```
44	PROBLEM: 11 consecutive classifier errors, 57 min wasted
45	WHY: The proposer suggested CatBoost, which isn't in the allowed algorithm list
46	ANSWER: The proposer has no documentation of which algorithms are available
47	```
48
49	Rules for good answers:
50	- An answer must be factual and verifiable. Check logs, read code, look at data.
51	Don't speculate — investigate.
52	- If you can't verify an answer, say so. "I believe X but haven't confirmed" is
53	honest. "X" stated as fact when you haven't checked is not.
54	- An answer should be a mechanism, not a redescription. "It failed because it
55	was broken" is circular, not explanatory.
56
57	### Step 3: Go Deeper — Recursively
58
59	Take each answer and ask "why?" again. Keep going until you reach one of:
60
61	- A structural property of the system that can be changed (a missing
62	constraint, a bad default, an absent check)
63	- An external boundary you don't control (a vendor API, a hardware limit, a
64	policy)
65	- A known and accepted tradeoff that was made intentionally
66
67	That's bedrock. Stop there.
68
69	How to know you've hit bedrock: If you ask "why?" one more time and the answer
70	is either "that's how the universe works" or "because we chose to" — you're there.
71
72	How to know you haven't: If you can still point to a structural property that
73	someone could change — keep going.
74
75	### Step 4: The Tree Is Fractal
76
77	This is critical. At any single level, the answer to "why?" often has **two or
78	three independent causes*. That's not a problem — it's better*. Branch the tree.
79
80	```
81	PROBLEM: Agent was unresponsive for 62 minutes
82	│
83	├─ WHY-1: The LLM API call hung and never returned
84	│ └─ WHY: No turn-level timeout exists in the harness
85	│ └─ WHY: The harness assumes LLM calls always complete
86	│ → BEDROCK: Missing timeout (structural — fixable)
87	│
88	├─ WHY-2: Agent was already degraded from sync sleep-polling
89	│ └─ WHY: Agent doesn't trust the async callback mechanism
90	│ └─ WHY: Past flaky experiences created a hedging pattern
91	│ └─ WHY: No positive evidence of reliability in memory
92	│ → BEDROCK: Missing trust documentation (structural — fixable)
93	│
94	└─ WHY-3: The LLM endpoint was unstable all day
95	└─ WHY: Another team member deleted the deployment
96	→ BEDROCK: Shared resource, no access control (external boundary)
97	```
98
99	Multiple root causes are the norm, not the exception. Most problems are
100	overdetermined — several things conspired. Finding all of them matters because
101	fixing only one may not prevent recurrence.
102
103	Don't force your tree into a single chain. If you find yourself writing "and also"
104	— that's a second branch, not a conjunction.
105
106	### Step 5: Action Items
107
108	Every bedrock node should produce an action item. Here's what separates useful
109	actions from theater:
110
111	An action item must produce a concrete artifact. It must result in a file edit,
112	a config change

People who liked this also viewed

6 picks

Type