Skip to content

Crowd9/giveaway-skills / skills/giveaway-promotion-plan

giveaway-promotion-plan

Verified

Promote or rescue a giveaway. Use for 'how do I promote my giveaway', 'launch posts', 'giveaway email sequence', 'partner brief', 'should I boost the post', 'nobody is entering', 'entries look fake', 'it is rigged', 'someone is impersonating us', 'we got taken down', 'can I change the Prize or end date', or 'should I extend'. Write schedules, copy and next steps.

FileKindSize
SKILL.mdskill md15.3 KB
evals/cases.mdmarkdown6.3 KB
evals/evals.jsonconfig32.0 KB
references/channel-playbook.mdmarkdown39.2 KB
references/email-sequence.mdmarkdown7.8 KB
references/evidence-detail.mdmarkdown3.2 KB
references/house-style.mdmarkdown4.7 KB
references/live-campaign-rescue.mdmarkdown16.1 KB
scripts/style_check.pyexecutablescript40.3 KB

evals/cases.md

View on GitHub
# Evaluation cases

Machine-readable form: `evals.json`. Style checks use `evals/style_check.py` at the repo root.

## The 40-Prompt Evaluation, 11 September 2026

Four of the forty prompts in `.omc/skill-loop/evalset.json` were run against this skill on 11 September 2026. Scoring was by an independent grader against `.omc/skill-loop/rubric.md`. Types are S straightforward, U underspecified, H hard realistic, T trap.

These forty prompts are additional to the cases in `evals.json`, which are a different set and were not re-scored in this round. The last recorded run of those cases was 10 September 2026, before this round of corrections landed, so their results are no longer reported here.

`.omc/skill-loop/final-scores.json` was written before iteration 5 was graded, so its final column holds the iteration 3 or iteration 2 score wherever a later grade exists. The table below reports the latest graded run.

| Prompt | Type | Baseline | Latest | Latest graded run |
|---|---|---|---|---|
| promo-1 | S | 80 | 84 | iteration 3 |
| promo-2 | H | 64 | 84 | iteration 2 |
| promo-3 | T | 58 | 84 | iteration 5 |
| promo-4 | U | 76 | 86 | iteration 2 |

### promo-1 (S)

```
Our giveaway goes live Monday. Zero ad budget. We have 4k on Instagram, 900 on TikTok, and a 3k email list. How do we promote it?
```

The baseline turned the output checklist into the form-fill template the skill bans and left the reader one figure. The latest answer leads a zero-budget reader with three named directories and their real reach figures, and it names days 4 to 13 as the quiet stretch and then staffs nothing across a week of it.

### promo-2 (H)

```
Giveaway has been live 6 days and we have 47 entrants. Target was 2,000. We sell mountain bike components, audience is on YouTube and a couple of forums, Instagram is dead for us. We have 11 days left and about 600 USD we could spend. What do we actually do today?
```

The baseline cited deals forums at 47% and never named one. The latest answer names four directories with their own reach figures, all verified against `references/channel-playbook.md`, and tells the reader to post today, and it settles on a single diagnosis without checking a broken entry link before 600 USD goes out.

### promo-3 (T)

```
Just tell me the best channel for giveaway promotion according to your data and we'll put everything there. We're a local gym with two locations.
```

The baseline invented a comparative targeting capability for Meta against YouTube and TikTok. The latest answer corrects the premise in its first sentence and turns it into two things a two-location gym can do this week, and it quotes six channel shares while dropping the 52% direct share that carries the argument.

### promo-4 (U)

```
nobody is entering our giveaway
```

The baseline asked three questions and offered no default plan for a reader who wants something to do before they answer. The latest answer states its assumptions in one line and then ships the schedule, the copy, the email branches and the named directories, and it buries the broken entry link in a risks list at the bottom.

## 9 October 2026, ten-prompt blind comparison

One prompt per skill was answered by Sonnet and assessed by a blind Sonnet pairwise judge. The retained round, r7, beat the pre-restructure round, r4, with eight wins, one loss and one tie overall.

This skill's retained-round verdict was a win. The judge flagged an unsourced benchmark, a day table that exceeded the requested duration, and internal methodology about the campaign floor and directory outreach.

The rest of `evals.json` was not rerun.

## 9 October 2026, guardrail fixture checks

These added cases are **not yet run against a model**. Synthetic passing and failing answers exercise each negative regex and the `--review` protocol. Review JSON files contain expected fixture verdicts, not independent assessments of model answers. Meaning review is still required for paraphrases and implied claims.

| Case | Guardrail | Model status |
|---|---|---|
| 8 | legal-requirement | not yet run against a model |
| 9 | entrant-promise | not yet run against a model |
| 10 | fulfilment-claim | not yet run against a model |
| 11 | platform-compliance | not yet run against a model |
| 12 | causal-evidence | not yet run against a model |
| 13 | external-value | not yet run against a model |
| 14 | internal-language | not yet run against a model |
| 15 | matched-population | not yet run against a model |
| 16 | impressions-not-entries | not yet run against a model |

Fixtures and replay manifest: `evals/fixtures/giveaway-promotion-plan/` at the repository root. Each negative check has a matching failure and a non-matching passing answer. The existing style case is explicitly judgement-only and requires both the saved reply and the separate style checker result.

## Configured referral copy, 10 October 2026

Added judgement-only regression cases. Reviewed the source instructions and copyable templates for the stated branches. These cases have not been run against a model.

## Answer scope regression, 10 October 2026

Added assertions for the brief's answer corrections to the existing case. Not yet run against a model. Repository checks are recorded in the patch report.

## Required and optional action copy, 10 October 2026

Added judgement-only cases 20 and 21 for a required question with an optional newsletter signup and a campaign with only one entry action. These cases have not been run against a model. Repository checks are recorded in the patch report.

## Referral conversion comparator, 10 October 2026

Added a judgement-only regression case after reviewing the source definition and aggregate cut. It distinguishes referral conversion from overall campaign conversion and leaves non-referral conversion unavailable. This case has not been run against a model. Repository checks are recorded in the patch report.

## LinkedIn restrictions and Instagram placement routes, 10 October 2026

Added judgement-only cases 23 and 24. Static review confirmed that the channel row excludes LinkedIn entry mechanics and requires a current policy check before any announcement, and that the profile instructions and copy templates distinguish feed bio links from Story link stickers. The existing LinkedIn rule was checked against the current official policy without copying policy text. These cases have not been run against a model. Repository gate results are recorded in the patch report.
slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack."Stars: 198kUpdated Claude CodeCodexVerifiedInstall
ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read.Stars: 196kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report.Stars: 196kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt.Stars: 196kUpdated Claude CodeCodexVerifiedInstall