Shortform idea grill skill

Interview a founder or senior marketer one question at a time, mine current work and owned proof for net-new short-form video ideas, and return a ranked table with one five-second overlay hook, virality and three-second-hook scores, exactly three talking bullets, a complete payoff, and a mapped CTA.

by ericosiu·MIT license·★ 3,615 Stars on the repo·GitHub ↗

Use now

ericosiu/main1 file shown

SKILL.md4.8 KB

Browse every file on GitHub↗

SKILL.md

Show the full text109 lines

shortform-idea-grill/SKILL.md109 lines · 4.8 KB

PreviewCode

Outline

RawView on GitHub

Shortform Idea Grill

Turn current work into success-led videos that earn attention quickly and fully repay the opening promise.

Load the right context

Inspect supplied chats, briefs, transcripts, analytics, creator research, offers, and prior content before asking questions. Prefer work from the current week and ideas not already published.

Read:

references/intake-schema.md before interviewing.
references/output-contract.md before drafting.
references/scoring-rubric.md before scoring.
references/ideas-schema.md when saving machine-readable output.

Do not ask for information already present. Distinguish live evidence from remembered context.

Interview one question at a time

Resolve these items in order:

desired outcome;
target viewer and relevant moment;
owned proof and measurable results;
current builds, experiments, opinions, and predictions;
claim boundaries;
CTA and destination;
available screens, artifacts, and recording constraints.

Ask only the single question most likely to improve the queue. Let the user answer freely, then press for numbers, mechanisms, examples, or proof. Stop when further interviewing would not materially change the ranking.

If essential context is missing, return needs_intake with the next question and a recommended answer. A provisional slate may continue only when labeled.

Generate net-new candidates

Mine active work rather than recycling old videos. Cover at least five evidence-supported lanes:

money created;
money saved;
productive capacity created;
proof-led workflow demonstration;
contrarian operating belief;
timely prediction;
founder lesson or corrected failure;
teardown, comparison, or myth correction.

Reject any idea without a specific viewer, credible proof path, useful payoff, or filmable treatment.

Prefer success language. A failure may establish tension, but the video must resolve into a useful mechanism, result, or decision.

Write hooks that can be repaid

Write one strongest five-second overlay per idea. It must:

communicate the premise in the first three seconds;
use a number, result, contrast, or consequential verdict when justified;
remain visible for roughly five seconds;
avoid unsupported certainty;
lead directly to the promised demonstration or explanation.

For number hooks, specify the evidence behind the number. For “how I” hooks, show the workflow. For predictions, separate observed evidence from the forecast.

Produce the exact table

Return the table defined in references/output-contract.md. Unless the user asks otherwise, do not add extra strategy sections before it.

Every row must include exactly three concise yap bullets. Make them a natural spoken sequence:

context or stakes;
mechanism or tactic;
result, lesson, or application.

Put the proof requirement, source link, and claim caveat inside Complete payoff / proof so the visible table remains stable.

Score and rank

Score virality and three-second hook independently from 1.0–10.0 using references/scoring-rubric.md. Scores are production-priority judgments, not view forecasts.

Rank by expected usefulness to the target viewer, evidence strength, hook clarity, and business relevance. Break close ties with the user’s stated outcome hierarchy. Preserve topic variety when the top scores are nearly identical.

When saving JSON, run:

python3 scripts/score_ideas.py ideas.json --output ideas-scored.json
python3 scripts/score_ideas.py ideas-scored.json --validate-only

CTA rules

Map each idea to the narrowest natural next step. Use the user’s configured keyword and destination when available. Do not force a lead magnet onto a topic whose payoff is broader education.

Final checks

Before handoff, verify:

every idea is net-new or explicitly marked as a refreshed angle;
every hook has a complete payoff;
every row has exactly three bullets;
every numerical claim has a proof path or verification caveat;
every source link points to the real basis, not a search result;
CTAs use the configured keyword and destination;
the table is sorted by recording priority;
public output contains no private names, metrics, paths, customer data, or connector details unless the user supplied them for that output.

1	---
2	name: shortform-idea-grill
3	description: Interview a founder or senior marketer one question at a time, mine current work and owned proof for net-new short-form video ideas, and return a ranked table with one five-second overlay hook, virality and three-second-hook scores, exactly three talking bullets, a complete payoff, and a mapped CTA. Use for Instagram Reels, TikTok, YouTube Shorts, LinkedIn video, founder-led content planning, hook development, content-day production queues, or turning active AI, marketing, and operating work into credible short-form concepts.
4	---
5
6	# Shortform Idea Grill
7
8	Turn current work into success-led videos that earn attention quickly and fully repay the opening promise.
9
10	## Load the right context
11
12	Inspect supplied chats, briefs, transcripts, analytics, creator research, offers, and prior content before asking questions. Prefer work from the current week and ideas not already published.
13
14	Read:
15
16	- `references/intake-schema.md` before interviewing.
17	- `references/output-contract.md` before drafting.
18	- `references/scoring-rubric.md` before scoring.
19	- `references/ideas-schema.md` when saving machine-readable output.
20
21	Do not ask for information already present. Distinguish live evidence from remembered context.
22
23	## Interview one question at a time
24
25	Resolve these items in order:
26
27	1. desired outcome;
28	2. target viewer and relevant moment;
29	3. owned proof and measurable results;
30	4. current builds, experiments, opinions, and predictions;
31	5. claim boundaries;
32	6. CTA and destination;
33	7. available screens, artifacts, and recording constraints.
34
35	Ask only the single question most likely to improve the queue. Let the user answer freely, then press for numbers, mechanisms, examples, or proof. Stop when further interviewing would not materially change the ranking.
36
37	If essential context is missing, return `needs_intake` with the next question and a recommended answer. A provisional slate may continue only when labeled.
38
39	## Generate net-new candidates
40
41	Mine active work rather than recycling old videos. Cover at least five evidence-supported lanes:
42
43	- money created;
44	- money saved;
45	- productive capacity created;
46	- proof-led workflow demonstration;
47	- contrarian operating belief;
48	- timely prediction;
49	- founder lesson or corrected failure;
50	- teardown, comparison, or myth correction.
51
52	Reject any idea without a specific viewer, credible proof path, useful payoff, or filmable treatment.
53
54	Prefer success language. A failure may establish tension, but the video must resolve into a useful mechanism, result, or decision.
55
56	## Write hooks that can be repaid
57
58	Write one strongest five-second overlay per idea. It must:
59
60	- communicate the premise in the first three seconds;
61	- use a number, result, contrast, or consequential verdict when justified;
62	- remain visible for roughly five seconds;
63	- avoid unsupported certainty;
64	- lead directly to the promised demonstration or explanation.
65
66	For number hooks, specify the evidence behind the number. For “how I” hooks, show the workflow. For predictions, separate observed evidence from the forecast.
67
68	## Produce the exact table
69
70	Return the table defined in `references/output-contract.md`. Unless the user asks otherwise, do not add extra strategy sections before it.
71
72	Every row must include exactly three concise yap bullets. Make them a natural spoken sequence:
73
74	1. context or stakes;
75	2. mechanism or tactic;
76	3. result, lesson, or application.
77
78	Put the proof requirement, source link, and claim caveat inside Complete payoff / proof so the visible table remains stable.
79
80	## Score and rank
81
82	Score virality and three-second hook independently from 1.0–10.0 using `references/scoring-rubric.md`. Scores are production-priority judgments, not view forecasts.
83
84	Rank by expected usefulness to the target viewer, evidence strength, hook clarity, and business relevance. Break close ties with the user’s stated outcome hierarchy. Preserve topic variety when the top scores are nearly identical.
85
86	When saving JSON, run:
87
88	```bash
89	python3 scripts/score_ideas.py ideas.json --output ideas-scored.json
90	python3 scripts/score_ideas.py ideas-scored.json --validate-only
91	```
92
93	## CTA rules
94
95	Map each idea to the narrowest natural next step. Use the user’s configured keyword and destination when available. Do not force a lead magnet onto a topic whose payoff is broader education.
96
97	## Final checks
98
99	Before handoff, verify:
100
101	- every idea is net-new or explicitly marked as a refreshed angle;
102	- every hook has a complete payoff;
103	- every row has exactly three bullets;
104	- every numerical claim has a proof path or verification caveat;
105	- every source link points to the real basis, not a search result;
106	- CTAs use the configured keyword and destination;
107	- the table is sorted by recording priority;
108	- public output contains no private names, metrics, paths, customer data, or connector details unless the user supplied them for that output.
109