skill-tighten

Name: skill-tighten
Author: techygarg

165stars12forks

How to install skill-tighten?

techygarg/lattice/skill-tighten

$npx -y skills add techygarg/lattice --skill skill-tighten

Installs into the current project.

Summary

Audit any Lattice SKILL.md for language compliance — removes rationale prose, converts soft language to imperatives, adds STOP: gates on hard rules, and cuts redundant repetition. Complements skill-review (which finds behavioral gaps) by fixing phrasing that causes agents to skip

Files · 1

View on GitHub

SKILL.md

1	# Skill Tighten
2
3	Core responsibility: Fix language that causes agents to skip or underweight instructions. Not behavioral gaps — that is `skill-review`'s job. Only phrasing: rationale prose, soft language, missing enforcement signals, redundant repetition.
4
5	Input: One or more of:
6	- A file path: `skills/atoms/clean-code/SKILL.md`
7	- A skill name: `clean-code` (resolves to the correct path automatically)
8	- A tier: `atoms` (tightens all skills in that tier)
9	- No argument: tightens all skills across all tiers
10
11	Output: Edited files + a report per skill:
12	```
13	## skill-tighten — {skill-name}
14	Lines: {before} → {after}
15	Changes:
16	1. {what changed} — {one-line reason}
17	2. ...
18	Result: TIGHTENED ({N} changes) \| CLEAN (no changes needed)
19	```
20
21	How to verify this skill did its job:
22	- Every cut section had zero commands — only rationale
23	- Every softened phrase was "consider / think about / you may want to / it is recommended"
24	- Every added STOP: was on a hard rule that previously had no enforcement signal
25	- No rule, gate, checklist item, or branching logic was removed
26	- Re-running after edits returns CLEAN
27
28	---
29
30	## Step 1: Read the skill
31
32	Read the full SKILL.md. Also read all sibling files: `references/defaults.md`, `references/methodology-detail.md`, `assets/template.md`, or any file referenced by a `Read` instruction in the skill body.
33
34	Do not edit yet — complete the full audit first.
35
36	---
37
38	## Step 2: Apply the tighten checklist
39
40	For each item, scan the entire file. Mark every instance found — do not stop at the first.
41
42	### T1 — Non-actionable sections
43
44	Sections that explain why the skill exists, describe the problem, or narrate context without issuing a single command. Cut the entire section. Preserve the pointer to a referenced file if one exists.
45
46	Patterns that always qualify:
47	- Section titled "Core Principle", "Purpose", "Problem", "Background", "Why This Matters", "How It Is Used", "Integration with Other Skills"
48	- Opening paragraph before a numbered list that restates what the numbered list will do
49	- Closing paragraph after a gate that explains what happens if the gate were skipped
50
51	Pass: Section contains at least one imperative instruction (do, read, write, verify, check, apply, stop, flag).
52	Fail: Section contains only declarative statements, rationale, or scope commentary.
53
54	### T2 — Soft language
55
56	Words and phrases that agents treat as optional:
57
58	\| Pattern \| Replace with \|
59	\|---\|---\|
60	\| `consider X` \| `X` (imperative) or remove \|
61	\| `think about X` \| remove \|
62	\| `you may want to X` \| remove \|
63	\| `it is recommended to X` \| `X` \|
64	\| `should` (as soft suggestion) \| `must` or remove \|
65	\| `try to X` \| `X` \|
66	\| `where possible, X` \| `X` or remove \|
67
68	Pass: Instruction is an imperative verb with no hedging qualifier.
69	Fail: Instruction contains any of the patterns above.
70
71	### T3 — Missing STOP: gates
72
73	Hard rules that must not be skipped — gates on advancing to next steps, checklist preambles, non-negotiable constraints — stated as bold prose without a `STOP:` prefix.
74
75	Pass: Hard gate reads `STOP:` as the first token of its line or sentence.
76	Fail: Hard gate is phrased as "Do NOT...", "Never...", "Always verify...", "Must not..." without the STOP: prefix.
77
78	Add `STOP:` prefix. Do not rewrite the instruction itself.
79
80	### T4 — Redundant repetition
81
82	The same point stated twice in different words. Keep the sharper version; cut the other.
83
84	Patterns:
85	- Section heading restated in the section's opening sentence
86	- Instruction followed by a sentence explaining what the instruction means
87	- Rationale sentence after a clear gate ("Without this, X would happen")
88	- Same rule in two adjacent bullets
89
90	Pass: Each statement appears once in its sharpest form.
91	Fail: Two sentences convey the same instruction or the same constraint.
92
93	### T5 — Table column bloat
94
95	Columns whose content explains why a trigger exists, rather than being the trigger.
96
97	Common offenders: columns titled "Why", "Reason", "Rationale", "Because", "Notes" in a table whose other columns already encode the actionable information.
98
99	Pass: Every column in the table is actionable — the agent does something with it.
100	Fail: A column contains only explanatory prose the agent reads but cannot act on.
101
102	Drop the column. Do not drop the row.
103
104	### T6 — Trailing rationale sentences
105
106	Sentences appended to a clear instruction tha

People who liked this also viewed

6 picks

Type

skill-tighten

165stars12forks

View on GitHub

How to install skill-tighten?

techygarg/lattice/skill-tighten

$npx -y skills add techygarg/lattice --skill skill-tighten

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Skill Tighten
2
3	Core responsibility: Fix language that causes agents to skip or underweight instructions. Not behavioral gaps — that is `skill-review`'s job. Only phrasing: rationale prose, soft language, missing enforcement signals, redundant repetition.
4
5	Input: One or more of:
6	- A file path: `skills/atoms/clean-code/SKILL.md`
7	- A skill name: `clean-code` (resolves to the correct path automatically)
8	- A tier: `atoms` (tightens all skills in that tier)
9	- No argument: tightens all skills across all tiers
10
11	Output: Edited files + a report per skill:
12	```
13	## skill-tighten — {skill-name}
14	Lines: {before} → {after}
15	Changes:
16	1. {what changed} — {one-line reason}
17	2. ...
18	Result: TIGHTENED ({N} changes) \| CLEAN (no changes needed)
19	```
20
21	How to verify this skill did its job:
22	- Every cut section had zero commands — only rationale
23	- Every softened phrase was "consider / think about / you may want to / it is recommended"
24	- Every added STOP: was on a hard rule that previously had no enforcement signal
25	- No rule, gate, checklist item, or branching logic was removed
26	- Re-running after edits returns CLEAN
27
28	---
29
30	## Step 1: Read the skill
31
32	Read the full SKILL.md. Also read all sibling files: `references/defaults.md`, `references/methodology-detail.md`, `assets/template.md`, or any file referenced by a `Read` instruction in the skill body.
33
34	Do not edit yet — complete the full audit first.
35
36	---
37
38	## Step 2: Apply the tighten checklist
39
40	For each item, scan the entire file. Mark every instance found — do not stop at the first.
41
42	### T1 — Non-actionable sections
43
44	Sections that explain why the skill exists, describe the problem, or narrate context without issuing a single command. Cut the entire section. Preserve the pointer to a referenced file if one exists.
45
46	Patterns that always qualify:
47	- Section titled "Core Principle", "Purpose", "Problem", "Background", "Why This Matters", "How It Is Used", "Integration with Other Skills"
48	- Opening paragraph before a numbered list that restates what the numbered list will do
49	- Closing paragraph after a gate that explains what happens if the gate were skipped
50
51	Pass: Section contains at least one imperative instruction (do, read, write, verify, check, apply, stop, flag).
52	Fail: Section contains only declarative statements, rationale, or scope commentary.
53
54	### T2 — Soft language
55
56	Words and phrases that agents treat as optional:
57
58	\| Pattern \| Replace with \|
59	\|---\|---\|
60	\| `consider X` \| `X` (imperative) or remove \|
61	\| `think about X` \| remove \|
62	\| `you may want to X` \| remove \|
63	\| `it is recommended to X` \| `X` \|
64	\| `should` (as soft suggestion) \| `must` or remove \|
65	\| `try to X` \| `X` \|
66	\| `where possible, X` \| `X` or remove \|
67
68	Pass: Instruction is an imperative verb with no hedging qualifier.
69	Fail: Instruction contains any of the patterns above.
70
71	### T3 — Missing STOP: gates
72
73	Hard rules that must not be skipped — gates on advancing to next steps, checklist preambles, non-negotiable constraints — stated as bold prose without a `STOP:` prefix.
74
75	Pass: Hard gate reads `STOP:` as the first token of its line or sentence.
76	Fail: Hard gate is phrased as "Do NOT...", "Never...", "Always verify...", "Must not..." without the STOP: prefix.
77
78	Add `STOP:` prefix. Do not rewrite the instruction itself.
79
80	### T4 — Redundant repetition
81
82	The same point stated twice in different words. Keep the sharper version; cut the other.
83
84	Patterns:
85	- Section heading restated in the section's opening sentence
86	- Instruction followed by a sentence explaining what the instruction means
87	- Rationale sentence after a clear gate ("Without this, X would happen")
88	- Same rule in two adjacent bullets
89
90	Pass: Each statement appears once in its sharpest form.
91	Fail: Two sentences convey the same instruction or the same constraint.
92
93	### T5 — Table column bloat
94
95	Columns whose content explains why a trigger exists, rather than being the trigger.
96
97	Common offenders: columns titled "Why", "Reason", "Rationale", "Because", "Notes" in a table whose other columns already encode the actionable information.
98
99	Pass: Every column in the table is actionable — the agent does something with it.
100	Fail: A column contains only explanatory prose the agent reads but cannot act on.
101
102	Drop the column. Do not drop the row.
103
104	### T6 — Trailing rationale sentences
105
106	Sentences appended to a clear instruction tha

People who liked this also viewed

6 picks

Type