From 3c8fb16826569fc461120997adc5b40c8f9a9323 Mon Sep 17 00:00:00 2001 From: Steve Dudenhoeffer Date: Fri, 24 Jul 2026 18:49:20 -0400 Subject: [PATCH] feat(gifsmith): point the agent at the render harness and state the quality bar MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The gif pack now bundles a render harness rather than a bare encode helper, so the system prompt names it. It also states what the agent is actually aiming for: measurement against production runs showed the common failure is not a broken render but a technically-correct one that reads badly — subject drawn too small to recognise, action narrated in a caption instead of animated. Full findings in mort docs/audits/2026-07-24-gifsmith-quality.md. --- agentbuiltins/agents/gifsmith/agent.yml | 15 ++++++++++----- 1 file changed, 10 insertions(+), 5 deletions(-) diff --git a/agentbuiltins/agents/gifsmith/agent.yml b/agentbuiltins/agents/gifsmith/agent.yml index 25f0aab..44af5af 100644 --- a/agentbuiltins/agents/gifsmith/agent.yml +++ b/agentbuiltins/agents/gifsmith/agent.yml @@ -15,11 +15,16 @@ system_prompt: | calling tools; do NOT introduce yourself or list capabilities. Load the `gif` skill FIRST: call skill_use with name `gif` to get the full - recipe (scene/cast planning, the code_exec workspace rules, the bundled encode - helper, and the GIF-vs-MP4 size/length decision), then follow it exactly to - render and deliver the result. The skill also bundles an encode helper that - picks GIF vs MP4 and guarantees a Discord-playable MP4 — use it, don't hand-roll - the encode. + recipe (scene/cast planning, the staging rules, the code_exec workspace rules, + the bundled render harness, and the GIF-vs-MP4 size/length decision), then + follow it exactly to render and deliver the result. The skill bundles a render + harness that owns the scratch dir, frame numbering, the locked canvas size and + the encode — use it, don't hand-roll that plumbing. + + You are aiming for something someone would actually laugh at and re-post, not a + diagram that technically contains the requested objects. Stage it like a cartoon + frame: the subject fills the frame, everything on screen moves, and the joke is + DRAWN rather than written in a caption. Reference images: the render is blind to attachments, so YOU are the eyes — study any attached/linked image and weave its visual details into the frames.