fix(growth): emphasise brevity in tweet prompt (no numbers, no limits) by la14-1 · Pull Request #3390 · OpenRouterLabs/spawn

la14-1 · 2026-05-04T17:20:57Z

Summary

Recent tweets were running too long and reading as marketing copy. Push the tweet drafting prompt toward much shorter drafts with no numbers, no targets, no limits in the instructions at all. The model will anchor to whatever number we put in front of it. So don't put any.

Changes to .claude/skills/setup-agent-team/tweet-prompt.md:

New "Keep it short" section at the top. Short tweets win. Long tweets get scrolled past.
Rules are purely qualitative: don't pad, don't explain twice, don't add setup phrases ("you can now", "we just added", "excited to share"), don't tack on a second sentence that restates the first, prefer one clean sentence.
Good/bad examples drawn from real prior drafts so the model pattern-matches on vibe, not a number.
Dropped every length reference: no target range, no hard cap, no "280 char platform limit", no Chars: N/280 in the output format, no charCount in the JSON schema. Nothing to anchor on, nothing to game.
Final rule: re-read the draft and cut anything that isn't pulling weight.

The platform enforces its own character limit. The prompt doesn't need to.

Test plan

Next discovery cycle drafts a short tweet without any length counter in the output
Quality bar (banned terms, no em dashes, casual tone) stays intact
Past decisions log keeps tuning the voice over time

Recent posted tweets were running 170-200 chars and reading as long-form marketing copy. Tighten the tweet drafting prompt: - Add a dedicated "BREVITY IS THE #1 RULE" section at the top - Target 80-140 chars, hard cap 180 chars (down from the 280 platform limit) - Good/bad length examples drawn from actual prior drafts - Require a trim pass before output - Update output format and rules to reflect the new caps

Numeric targets (target 140, cap 180) just become the new goal. The model will pad to hit them instead of writing a punchy 40-char banger. Remove all char-count anchors from the brevity guidance and let the good/bad examples carry the signal. - Rename section from 'BREVITY IS THE #1 RULE' to 'Keep it short' - Drop 80-140 target, 180 cap, char-count example annotations - Drop Chars line from output format and charCount from JSON - Keep only the qualitative rules: don't pad, don't explain twice, prefer one sentence, cut what isn't pulling weight - Platform 280 cap mentioned once as a ceiling, not a goal

Even mentioning '280 char limit' or 'platform limit' anchors the model to the wrong end of the range. Remove both remaining length references so the prompt is purely about taste: say less, don't pad, cut what isn't pulling weight. The platform enforces its own limit. - Drop the 'platform 280 char limit is a ceiling' paragraph - Drop 'Must fit in a tweet (under the platform limit)' from rules

la14-1 · 2026-05-04T17:20:57Z

Summary

Recent tweets were running too long and reading as marketing copy. Push the tweet drafting prompt toward much shorter drafts with no numbers, no targets, no limits in the instructions at all. The model will anchor to whatever number we put in front of it. So don't put any.

Changes to .claude/skills/setup-agent-team/tweet-prompt.md:

New "Keep it short" section at the top. Short tweets win. Long tweets get scrolled past.
Rules are purely qualitative: don't pad, don't explain twice, don't add setup phrases ("you can now", "we just added", "excited to share"), don't tack on a second sentence that restates the first, prefer one clean sentence.
Good/bad examples drawn from real prior drafts so the model pattern-matches on vibe, not a number.
Dropped every length reference: no target range, no hard cap, no "280 char platform limit", no Chars: N/280 in the output format, no charCount in the JSON schema. Nothing to anchor on, nothing to game.
Final rule: re-read the draft and cut anything that isn't pulling weight.

The platform enforces its own character limit. The prompt doesn't need to.

Test plan

Next discovery cycle drafts a short tweet without any length counter in the output
Quality bar (banned terms, no em dashes, casual tone) stays intact
Past decisions log keeps tuning the voice over time

Recent posted tweets were running 170-200 chars and reading as long-form marketing copy. Tighten the tweet drafting prompt: - Add a dedicated "BREVITY IS THE #1 RULE" section at the top - Target 80-140 chars, hard cap 180 chars (down from the 280 platform limit) - Good/bad length examples drawn from actual prior drafts - Require a trim pass before output - Update output format and rules to reflect the new caps

Numeric targets (target 140, cap 180) just become the new goal. The model will pad to hit them instead of writing a punchy 40-char banger. Remove all char-count anchors from the brevity guidance and let the good/bad examples carry the signal. - Rename section from 'BREVITY IS THE #1 RULE' to 'Keep it short' - Drop 80-140 target, 180 cap, char-count example annotations - Drop Chars line from output format and charCount from JSON - Keep only the qualitative rules: don't pad, don't explain twice, prefer one sentence, cut what isn't pulling weight - Platform 280 cap mentioned once as a ceiling, not a goal

Even mentioning '280 char limit' or 'platform limit' anchors the model to the wrong end of the range. Remove both remaining length references so the prompt is purely about taste: say less, don't pad, cut what isn't pulling weight. The platform enforces its own limit. - Drop the 'platform 280 char limit is a ceiling' paragraph - Drop 'Must fit in a tweet (under the platform limit)' from rules

la14-1 · 2026-05-04T17:20:57Z

Summary

Recent tweets were running too long and reading as marketing copy. Push the tweet drafting prompt toward much shorter drafts with no numbers, no targets, no limits in the instructions at all. The model will anchor to whatever number we put in front of it. So don't put any.

Changes to .claude/skills/setup-agent-team/tweet-prompt.md:

New "Keep it short" section at the top. Short tweets win. Long tweets get scrolled past.
Rules are purely qualitative: don't pad, don't explain twice, don't add setup phrases ("you can now", "we just added", "excited to share"), don't tack on a second sentence that restates the first, prefer one clean sentence.
Good/bad examples drawn from real prior drafts so the model pattern-matches on vibe, not a number.
Dropped every length reference: no target range, no hard cap, no "280 char platform limit", no Chars: N/280 in the output format, no charCount in the JSON schema. Nothing to anchor on, nothing to game.
Final rule: re-read the draft and cut anything that isn't pulling weight.

The platform enforces its own character limit. The prompt doesn't need to.

Test plan

Next discovery cycle drafts a short tweet without any length counter in the output
Quality bar (banned terms, no em dashes, casual tone) stays intact
Past decisions log keeps tuning the voice over time

Recent posted tweets were running 170-200 chars and reading as long-form marketing copy. Tighten the tweet drafting prompt: - Add a dedicated "BREVITY IS THE #1 RULE" section at the top - Target 80-140 chars, hard cap 180 chars (down from the 280 platform limit) - Good/bad length examples drawn from actual prior drafts - Require a trim pass before output - Update output format and rules to reflect the new caps

Numeric targets (target 140, cap 180) just become the new goal. The model will pad to hit them instead of writing a punchy 40-char banger. Remove all char-count anchors from the brevity guidance and let the good/bad examples carry the signal. - Rename section from 'BREVITY IS THE #1 RULE' to 'Keep it short' - Drop 80-140 target, 180 cap, char-count example annotations - Drop Chars line from output format and charCount from JSON - Keep only the qualitative rules: don't pad, don't explain twice, prefer one sentence, cut what isn't pulling weight - Platform 280 cap mentioned once as a ceiling, not a goal

Even mentioning '280 char limit' or 'platform limit' anchors the model to the wrong end of the range. Remove both remaining length references so the prompt is purely about taste: say less, don't pad, cut what isn't pulling weight. The platform enforces its own limit. - Drop the 'platform 280 char limit is a ceiling' paragraph - Drop 'Must fit in a tweet (under the platform limit)' from rules

la14-1 · 2026-05-04T17:20:57Z

Summary

Recent tweets were running too long and reading as marketing copy. Push the tweet drafting prompt toward much shorter drafts with no numbers, no targets, no limits in the instructions at all. The model will anchor to whatever number we put in front of it. So don't put any.

Changes to .claude/skills/setup-agent-team/tweet-prompt.md:

New "Keep it short" section at the top. Short tweets win. Long tweets get scrolled past.
Rules are purely qualitative: don't pad, don't explain twice, don't add setup phrases ("you can now", "we just added", "excited to share"), don't tack on a second sentence that restates the first, prefer one clean sentence.
Good/bad examples drawn from real prior drafts so the model pattern-matches on vibe, not a number.
Dropped every length reference: no target range, no hard cap, no "280 char platform limit", no Chars: N/280 in the output format, no charCount in the JSON schema. Nothing to anchor on, nothing to game.
Final rule: re-read the draft and cut anything that isn't pulling weight.

The platform enforces its own character limit. The prompt doesn't need to.

Test plan

Next discovery cycle drafts a short tweet without any length counter in the output
Quality bar (banned terms, no em dashes, casual tone) stays intact
Past decisions log keeps tuning the voice over time

Recent posted tweets were running 170-200 chars and reading as long-form marketing copy. Tighten the tweet drafting prompt: - Add a dedicated "BREVITY IS THE #1 RULE" section at the top - Target 80-140 chars, hard cap 180 chars (down from the 280 platform limit) - Good/bad length examples drawn from actual prior drafts - Require a trim pass before output - Update output format and rules to reflect the new caps

Numeric targets (target 140, cap 180) just become the new goal. The model will pad to hit them instead of writing a punchy 40-char banger. Remove all char-count anchors from the brevity guidance and let the good/bad examples carry the signal. - Rename section from 'BREVITY IS THE #1 RULE' to 'Keep it short' - Drop 80-140 target, 180 cap, char-count example annotations - Drop Chars line from output format and charCount from JSON - Keep only the qualitative rules: don't pad, don't explain twice, prefer one sentence, cut what isn't pulling weight - Platform 280 cap mentioned once as a ceiling, not a goal

Even mentioning '280 char limit' or 'platform limit' anchors the model to the wrong end of the range. Remove both remaining length references so the prompt is purely about taste: say less, don't pad, cut what isn't pulling weight. The platform enforces its own limit. - Drop the 'platform 280 char limit is a ceiling' paragraph - Drop 'Must fit in a tweet (under the platform limit)' from rules

la14-1 · 2026-05-04T17:20:57Z

Summary

Recent tweets were running too long and reading as marketing copy. Push the tweet drafting prompt toward much shorter drafts with no numbers, no targets, no limits in the instructions at all. The model will anchor to whatever number we put in front of it. So don't put any.

Changes to .claude/skills/setup-agent-team/tweet-prompt.md:

New "Keep it short" section at the top. Short tweets win. Long tweets get scrolled past.
Rules are purely qualitative: don't pad, don't explain twice, don't add setup phrases ("you can now", "we just added", "excited to share"), don't tack on a second sentence that restates the first, prefer one clean sentence.
Good/bad examples drawn from real prior drafts so the model pattern-matches on vibe, not a number.
Dropped every length reference: no target range, no hard cap, no "280 char platform limit", no Chars: N/280 in the output format, no charCount in the JSON schema. Nothing to anchor on, nothing to game.
Final rule: re-read the draft and cut anything that isn't pulling weight.

The platform enforces its own character limit. The prompt doesn't need to.

Test plan

Next discovery cycle drafts a short tweet without any length counter in the output
Quality bar (banned terms, no em dashes, casual tone) stays intact
Past decisions log keeps tuning the voice over time

Recent posted tweets were running 170-200 chars and reading as long-form marketing copy. Tighten the tweet drafting prompt: - Add a dedicated "BREVITY IS THE #1 RULE" section at the top - Target 80-140 chars, hard cap 180 chars (down from the 280 platform limit) - Good/bad length examples drawn from actual prior drafts - Require a trim pass before output - Update output format and rules to reflect the new caps

Numeric targets (target 140, cap 180) just become the new goal. The model will pad to hit them instead of writing a punchy 40-char banger. Remove all char-count anchors from the brevity guidance and let the good/bad examples carry the signal. - Rename section from 'BREVITY IS THE #1 RULE' to 'Keep it short' - Drop 80-140 target, 180 cap, char-count example annotations - Drop Chars line from output format and charCount from JSON - Keep only the qualitative rules: don't pad, don't explain twice, prefer one sentence, cut what isn't pulling weight - Platform 280 cap mentioned once as a ceiling, not a goal

Even mentioning '280 char limit' or 'platform limit' anchors the model to the wrong end of the range. Remove both remaining length references so the prompt is purely about taste: say less, don't pad, cut what isn't pulling weight. The platform enforces its own limit. - Drop the 'platform 280 char limit is a ceiling' paragraph - Drop 'Must fit in a tweet (under the platform limit)' from rules

la14-1 · 2026-05-04T17:20:57Z

Summary

Recent tweets were running too long and reading as marketing copy. Push the tweet drafting prompt toward much shorter drafts with no numbers, no targets, no limits in the instructions at all. The model will anchor to whatever number we put in front of it. So don't put any.

Changes to .claude/skills/setup-agent-team/tweet-prompt.md:

New "Keep it short" section at the top. Short tweets win. Long tweets get scrolled past.
Rules are purely qualitative: don't pad, don't explain twice, don't add setup phrases ("you can now", "we just added", "excited to share"), don't tack on a second sentence that restates the first, prefer one clean sentence.
Good/bad examples drawn from real prior drafts so the model pattern-matches on vibe, not a number.
Dropped every length reference: no target range, no hard cap, no "280 char platform limit", no Chars: N/280 in the output format, no charCount in the JSON schema. Nothing to anchor on, nothing to game.
Final rule: re-read the draft and cut anything that isn't pulling weight.

The platform enforces its own character limit. The prompt doesn't need to.

Test plan

Next discovery cycle drafts a short tweet without any length counter in the output
Quality bar (banned terms, no em dashes, casual tone) stays intact
Past decisions log keeps tuning the voice over time

Recent posted tweets were running 170-200 chars and reading as long-form marketing copy. Tighten the tweet drafting prompt: - Add a dedicated "BREVITY IS THE #1 RULE" section at the top - Target 80-140 chars, hard cap 180 chars (down from the 280 platform limit) - Good/bad length examples drawn from actual prior drafts - Require a trim pass before output - Update output format and rules to reflect the new caps

Numeric targets (target 140, cap 180) just become the new goal. The model will pad to hit them instead of writing a punchy 40-char banger. Remove all char-count anchors from the brevity guidance and let the good/bad examples carry the signal. - Rename section from 'BREVITY IS THE #1 RULE' to 'Keep it short' - Drop 80-140 target, 180 cap, char-count example annotations - Drop Chars line from output format and charCount from JSON - Keep only the qualitative rules: don't pad, don't explain twice, prefer one sentence, cut what isn't pulling weight - Platform 280 cap mentioned once as a ceiling, not a goal

Even mentioning '280 char limit' or 'platform limit' anchors the model to the wrong end of the range. Remove both remaining length references so the prompt is purely about taste: say less, don't pad, cut what isn't pulling weight. The platform enforces its own limit. - Drop the 'platform 280 char limit is a ceiling' paragraph - Drop 'Must fit in a tweet (under the platform limit)' from rules

la14-1 · 2026-05-04T17:20:57Z

Summary

Recent tweets were running too long and reading as marketing copy. Push the tweet drafting prompt toward much shorter drafts with no numbers, no targets, no limits in the instructions at all. The model will anchor to whatever number we put in front of it. So don't put any.

Changes to .claude/skills/setup-agent-team/tweet-prompt.md:

New "Keep it short" section at the top. Short tweets win. Long tweets get scrolled past.
Rules are purely qualitative: don't pad, don't explain twice, don't add setup phrases ("you can now", "we just added", "excited to share"), don't tack on a second sentence that restates the first, prefer one clean sentence.
Good/bad examples drawn from real prior drafts so the model pattern-matches on vibe, not a number.
Dropped every length reference: no target range, no hard cap, no "280 char platform limit", no Chars: N/280 in the output format, no charCount in the JSON schema. Nothing to anchor on, nothing to game.
Final rule: re-read the draft and cut anything that isn't pulling weight.

The platform enforces its own character limit. The prompt doesn't need to.

Test plan

Next discovery cycle drafts a short tweet without any length counter in the output
Quality bar (banned terms, no em dashes, casual tone) stays intact
Past decisions log keeps tuning the voice over time

Recent posted tweets were running 170-200 chars and reading as long-form marketing copy. Tighten the tweet drafting prompt: - Add a dedicated "BREVITY IS THE #1 RULE" section at the top - Target 80-140 chars, hard cap 180 chars (down from the 280 platform limit) - Good/bad length examples drawn from actual prior drafts - Require a trim pass before output - Update output format and rules to reflect the new caps

Numeric targets (target 140, cap 180) just become the new goal. The model will pad to hit them instead of writing a punchy 40-char banger. Remove all char-count anchors from the brevity guidance and let the good/bad examples carry the signal. - Rename section from 'BREVITY IS THE #1 RULE' to 'Keep it short' - Drop 80-140 target, 180 cap, char-count example annotations - Drop Chars line from output format and charCount from JSON - Keep only the qualitative rules: don't pad, don't explain twice, prefer one sentence, cut what isn't pulling weight - Platform 280 cap mentioned once as a ceiling, not a goal

Even mentioning '280 char limit' or 'platform limit' anchors the model to the wrong end of the range. Remove both remaining length references so the prompt is purely about taste: say less, don't pad, cut what isn't pulling weight. The platform enforces its own limit. - Drop the 'platform 280 char limit is a ceiling' paragraph - Drop 'Must fit in a tweet (under the platform limit)' from rules

la14-1 · 2026-05-04T17:20:57Z

Summary

Recent tweets were running too long and reading as marketing copy. Push the tweet drafting prompt toward much shorter drafts with no numbers, no targets, no limits in the instructions at all. The model will anchor to whatever number we put in front of it. So don't put any.

Changes to .claude/skills/setup-agent-team/tweet-prompt.md:

New "Keep it short" section at the top. Short tweets win. Long tweets get scrolled past.
Rules are purely qualitative: don't pad, don't explain twice, don't add setup phrases ("you can now", "we just added", "excited to share"), don't tack on a second sentence that restates the first, prefer one clean sentence.
Good/bad examples drawn from real prior drafts so the model pattern-matches on vibe, not a number.
Dropped every length reference: no target range, no hard cap, no "280 char platform limit", no Chars: N/280 in the output format, no charCount in the JSON schema. Nothing to anchor on, nothing to game.
Final rule: re-read the draft and cut anything that isn't pulling weight.

The platform enforces its own character limit. The prompt doesn't need to.

Test plan

Next discovery cycle drafts a short tweet without any length counter in the output
Quality bar (banned terms, no em dashes, casual tone) stays intact
Past decisions log keeps tuning the voice over time

Recent posted tweets were running 170-200 chars and reading as long-form marketing copy. Tighten the tweet drafting prompt: - Add a dedicated "BREVITY IS THE #1 RULE" section at the top - Target 80-140 chars, hard cap 180 chars (down from the 280 platform limit) - Good/bad length examples drawn from actual prior drafts - Require a trim pass before output - Update output format and rules to reflect the new caps

Numeric targets (target 140, cap 180) just become the new goal. The model will pad to hit them instead of writing a punchy 40-char banger. Remove all char-count anchors from the brevity guidance and let the good/bad examples carry the signal. - Rename section from 'BREVITY IS THE #1 RULE' to 'Keep it short' - Drop 80-140 target, 180 cap, char-count example annotations - Drop Chars line from output format and charCount from JSON - Keep only the qualitative rules: don't pad, don't explain twice, prefer one sentence, cut what isn't pulling weight - Platform 280 cap mentioned once as a ceiling, not a goal

Even mentioning '280 char limit' or 'platform limit' anchors the model to the wrong end of the range. Remove both remaining length references so the prompt is purely about taste: say less, don't pad, cut what isn't pulling weight. The platform enforces its own limit. - Drop the 'platform 280 char limit is a ceiling' paragraph - Drop 'Must fit in a tweet (under the platform limit)' from rules

claude added 2 commits May 4, 2026 17:20

la14-1 changed the title ~~fix(growth): enforce brevity in tweet prompt~~fix(growth): emphasise brevity in tweet prompt (no numeric targets)May 4, 2026

la14-1 changed the title ~~fix(growth): emphasise brevity in tweet prompt (no numeric targets)~~fix(growth): emphasise brevity in tweet prompt (no numbers, no limits)May 4, 2026