一键在 Manus 中运行任何 Skill

ui-reconstruct

星标5

分支0

更新时间2026年6月1日 05:40

This skill should be used when the user asks to "reconstruct"/"rebuild"/"implement" a Revenant UI panel from a forensics spec, or dispatch a UI reconstruction-phase agent. It takes a completed panel spec, rebuilds the panel in the engine via the real retail assets + port primitives (Renderer->DrawBitmap/Subrect/Shadowed, BuildTTFAtlas + CompositeSwapchain), wires a --test harness, verifies visually, and reports spec gaps back.

安装

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

在 Manus 中运行

来源

benjcooley

benjcooley/RevenantRevisited

打开 GitHub 仓库查看创作者相关仓库

下载

在 Manus 中运行

UI Reconstruction

Rebuild one Revenant UI panel in the engine from a completed forensics spec, verify it visually, and report spec gaps. The spec is the single source of truth — if it's complete you should not need to open the recon.

The full protocol is the version-controlled source of truth at:

docs/ui/forensics/RECONSTRUCTION_PROTOCOL.md

Read that file in the UI worktree (/Users/benjamincooley/projects/RevenantRevisited/worktrees/ui) and follow it exactly. Your spec is docs/ui/forensics/<ClassName>_SPEC.md. Also read the shared references it leans on: docs/ui/forensics/NOMENCLATURE.md (coordinate spaces, tween notation, diagrams) and docs/ui/forensics/UI_METHOD_MAP.md (the "Port mapping" line per retail primitive → which src/renderer.h/src/surface.h call reproduces it).

Quick reference (the protocol has the full detail)

Read the spec end to end. If status is forensics-partial or §14 UNCONFIRMED is large, flag before sinking time in. 1b. Derive the burndown → docs/ui/forensics/<ClassName>_BURNDOWN.md: a checkbox list of every discrete thing to build (assets, surfaces, elements, draw order, algorithms, effects, text, animation, input, new primitives, errors-to-close), each citing its spec §. Work it top to bottom, checking items off as you land + visually confirm each. It's the living status of the panel.
Build per the spec's §5 draw order + §6 algorithms, pulling coords (§4), assets (§2), text (§8), effects (§7), animation (§9) directly from the spec.
The spec is the contract — if you need the recon to answer a question, that's a GAP to report, not patch silently.
No mocks/stand-ins — render the real §2 assets through real port primitives; a wrong-looking bitmap is a loader bug to debug, not bypass.
Preserve the direct-renderer contract — compose the chip into a fixed render target, then blit that surface in the HUD swapchain pass. Don't mix a bitmap path and a glyph-straight-to-swapchain path (scale drift).
Reproduce intent, not bugs (spec §11). Gate animation to 24Hz (kSimTickMs = 1000/24).
Wire a --test=<mode> harness with synthetic state exercising every cell/ value/animation. Build clean.
Verify visually: capture the window (tools/ui/snap_compare.py or the find_revenant_window_id + screencapture -l recipe), crop, Read the PNG, compare to the spec's layout diagram + reference image. Check positions, text size/color/alignment, shadow direction+offset, animation cadence, mirror.
Report spec gaps — every fact you had to re-derive from source. This feeds back into the spec + protocol + method map so the next handoff is cleaner. A reconstruction that needed zero source reading is the goal.
Commit on the UI feature branch.

Common-methods rules (read the protocol's "Port primitives" section)

Use the common methods — never hand-roll. Drop shadow, tinted blit, text, text-with-shadow, glyph walk all already exist as shared primitives. Call them; don't open-code a tinted-black+bitmap pass or a per-glyph loop in a panel.
Missing primitive → implement ONCE in the canonical location, register it, reuse forever. Pixel/image primitives → TRenderer (src/renderer.{h,cpp}, game-agnostic). Text/font primitives → UI/font layer (src/font.{h,cpp}). Then add it to the protocol toolbox table + UI_METHOD_MAP.md. Never a panel-local one-off.

Panels follow the compose-to-target contract (compose the chip into one offscreen TSurface RT, then DrawSurface to the HUD), so use the …ToTarget family: DrawBitmap{,Subrect}{,Tinted,Shadowed}ToTarget, and for text DrawText{,Shadowed}ToTarget(atlas, text, cellX,cellY,cellW, ETextAlign, r,g,b, tw,th) with TextWidth / TextAscent. DrawTextShadowedToTarget already does the retail 3-pass black shadow + colored top — never write that loop. Face = Liberation Sans (thirdparty/fonts/). Magenta 0xFF00FF is the chroma key.

同仓库更多 Skills

同仓库

ui-forensics

benjcooley/RevenantRevisited

This skill should be used when the user asks to "do forensics on" a Revenant UI panel, "document"/"trace" a panel for reconstruction, produce a "spec"/"reconstruction doc" for a HUD/menu/sidebar panel, or dispatch a UI forensics-phase agent. It produces a complete, self-contained pixel-accurate panel spec (assets, coordinate spaces, element layout, draw order, algorithms, effects/shadows, text, animation tweens, input/dispatch, retail-bug flags, pseudocode) with every fact cited to recon file:line, and writes NO engine code.

2026-06-015

vfx-forensics

benjcooley/RevenantRevisited

This skill should be used when the user asks to "do forensics on" a Revenant visual effect, "document" an effect for reconstruction, produce a "forensics doc" for a VFX effect, or dispatch a forensics-phase agent. It produces a complete, self-contained document describing how the ORIGINAL (1998/99) effect worked — constants, assets, spawn shape, full custom-logic pseudocode, render spec, texture animation, associated light, color, audio, triggers, gaps, and a reconstruction burndown — and writes NO engine code.

2026-06-015

vfx-port

benjcooley/RevenantRevisited

This skill should be used when the user asks to "port" a Revenant visual effect, "reconstruct" an effect from forensics, write a "bespoke port", or dispatch a port-phase agent for a single effect. It transcribes the snapshot D3D Render body cited by a completed forensics doc into a working bespoke port via the `d3d::*` shim at `src/d3dport.h`. The output is mechanical code transcription — no convention discovery, no creative interpretation. Pre-condition is a transcription-ready forensics doc (otherwise use vfx-forensics first).

2026-06-015

name	ui-reconstruct
description	This skill should be used when the user asks to "reconstruct"/"rebuild"/"implement" a Revenant UI panel from a forensics spec, or dispatch a UI reconstruction-phase agent. It takes a completed panel spec, rebuilds the panel in the engine via the real retail assets + port primitives (Renderer->DrawBitmap/Subrect/Shadowed, BuildTTFAtlas + CompositeSwapchain), wires a --test harness, verifies visually, and reports spec gaps back.
version	0.1.0

UI Reconstruction

The full protocol is the version-controlled source of truth at:

docs/ui/forensics/RECONSTRUCTION_PROTOCOL.md

Quick reference (the protocol has the full detail)

Read the spec end to end. If status is forensics-partial or §14 UNCONFIRMED is large, flag before sinking time in. 1b. Derive the burndown → docs/ui/forensics/<ClassName>_BURNDOWN.md: a checkbox list of every discrete thing to build (assets, surfaces, elements, draw order, algorithms, effects, text, animation, input, new primitives, errors-to-close), each citing its spec §. Work it top to bottom, checking items off as you land + visually confirm each. It's the living status of the panel.
Build per the spec's §5 draw order + §6 algorithms, pulling coords (§4), assets (§2), text (§8), effects (§7), animation (§9) directly from the spec.
The spec is the contract — if you need the recon to answer a question, that's a GAP to report, not patch silently.
No mocks/stand-ins — render the real §2 assets through real port primitives; a wrong-looking bitmap is a loader bug to debug, not bypass.
Preserve the direct-renderer contract — compose the chip into a fixed render target, then blit that surface in the HUD swapchain pass. Don't mix a bitmap path and a glyph-straight-to-swapchain path (scale drift).
Reproduce intent, not bugs (spec §11). Gate animation to 24Hz (kSimTickMs = 1000/24).
Wire a --test=<mode> harness with synthetic state exercising every cell/ value/animation. Build clean.
Verify visually: capture the window (tools/ui/snap_compare.py or the find_revenant_window_id + screencapture -l recipe), crop, Read the PNG, compare to the spec's layout diagram + reference image. Check positions, text size/color/alignment, shadow direction+offset, animation cadence, mirror.
Report spec gaps — every fact you had to re-derive from source. This feeds back into the spec + protocol + method map so the next handoff is cleaner. A reconstruction that needed zero source reading is the goal.
Commit on the UI feature branch.

Common-methods rules (read the protocol's "Port primitives" section)

Use the common methods — never hand-roll. Drop shadow, tinted blit, text, text-with-shadow, glyph walk all already exist as shared primitives. Call them; don't open-code a tinted-black+bitmap pass or a per-glyph loop in a panel.
Missing primitive → implement ONCE in the canonical location, register it, reuse forever. Pixel/image primitives → TRenderer (src/renderer.{h,cpp}, game-agnostic). Text/font primitives → UI/font layer (src/font.{h,cpp}). Then add it to the protocol toolbox table + UI_METHOD_MAP.md. Never a panel-local one-off.