Guided build of a database feature
Paladin · Full-stack websites
Anthropic Feature Development
Supabase Postgres Best Practices
Supabase MCP Server
Microsoft Playwright MCP
Seven-phase Claude Code workflow that explores the repo, designs options, then implements after you approve.
How it works
Foe: Fog Wraith, missing facts, outdated docs or thin context.
Foes
- Fog WraithHaze Poison Ward Insight Flicker Pinned Enraged
Party · Context
- PaladinYour agent100/100Haze Poison Ward Insight Flicker Pinned Enraged
- Code explorerSubagent0/40Haze Poison Ward Insight Flicker Pinned Enraged
- Code explorerSubagent0/40Haze Poison Ward Insight Flicker Pinned Enraged
- Architect: minimalSubagent0/40Haze Poison Ward Insight Flicker Pinned Enraged
- Architect: cleanSubagent0/40Haze Poison Ward Insight Flicker Pinned Enraged
- Architect: pragmaticSubagent0/40Haze Poison Ward Insight Flicker Pinned Enraged
The Fog Wraith blocks the way.
What really happensIt stands for missing facts, outdated docs or thin context.
Paladin: Context 100/100.
The Fog Wraith breathes haze. Swings start to miss.
What really happensWithout current docs, the agent guesses at APIs that may have changed.
Paladin takes 8. Paladin: Context 92/100. Paladin: Haze.
The Feature Development relic unrolls before a line is written.
What really happensYou run /feature-dev with a feature request. Phase 1 asks what problem you are solving and confirms it.
Fog Wraith takes 17.
Scouts walk the fog and mark paths already cut.
What really happensPhase 2 launches 2 or 3 code-explorer agents, then reads the files they name. Phase 3 waits for your answers.
Code explorer: Context 40/40. Code explorer: Context 40/40.
The Fog Wraith lashes at Code explorer.
What really happensThe subagent takes the hit in its own context. The main context stays clean.
Code explorer takes 8. Code explorer: Context 32/40. Code explorer: Haze.
The summons work the problem, each in its own context.
What really happensEach subagent does its own model and tool work, apart from the main conversation.
Fog Wraith takes 7. Fog Wraith takes 7.
The summons return with short reports.
What really happensOnly the subagent's final summary comes back to the main agent.
Paladin takes 2. Paladin takes 2. Paladin: Context 88/100. Paladin: Insight. Code explorer returns. Code explorer returns.
Architects offer blueprints and wait for your pick.
What really happensPhase 4 launches 2 or 3 code-architect agents (minimal, clean, pragmatic) and asks which approach you want.
Architect: minimal: Context 40/40. Architect: clean: Context 40/40. Architect: pragmatic: Context 40/40.
The summons work the problem, each in its own context.
What really happensEach subagent does its own model and tool work, apart from the main conversation.
Fog Wraith takes 7. Fog Wraith takes 7. Fog Wraith takes 7.
The summons return with short reports.
What really happensOnly the subagent's final summary comes back to the main agent.
Paladin takes 2. Paladin takes 2. Paladin takes 2. Paladin: Context 82/100. Architect: minimal returns. Architect: clean returns. Architect: pragmatic returns.
Building starts only after you approve the plan.
What really happensPhase 5 waits for explicit approval, then implements the chosen design using patterns found in the repo.
Fog Wraith takes 21.
A short chronicle names what was built and which files moved.
What really happensPhase 6 runs three code-reviewer agents and asks whether to fix findings. Phase 7 summarizes decisions and files changed.
Fog Wraith takes 27. Fog Wraith falls.
The Fog Wraith falls.
1 / 13
Learn this spell
Claude Code
/plugin install feature-dev@claude-plugins-officialNotes
Claude Code only. `/feature-dev` runs seven phases. It waits for your answers before design, and for your approval before it writes code. It needs an existing codebase to learn from and a git repository for the review phase. The README says not to use it for single-line fixes, trivial changes, or urgent hotfixes. Phase 6 launches three code-reviewer agents and asks whether to fix now, fix later, or proceed.
Proving Grounds record
Controlled quest results, separate from recorded usage. Each setup pools runs across its quests. The per-quest table below has casts, cast time, mana, incantation and runes; medians are not pooled.
Hero: Claude Code 2.1.305; weapon (model): claude-haiku-5-5; charge (effort): low; arena: sealed; invoked cast (the prompt named the skill). macOS 27.0.1, arm64. Curator.
Clear rate100%1 of 1 runs passed every check; Unarmed: 3 of 3 (matched quests)
Steadiness21%reliability: Wilson 95% lower bound
Buff+0 ppvs Unarmed, only matched quests (1 equipped runs; 3 control runs)
Instant casts1 of 1runs cleared on the first prompt; Unarmed: 3 of 3 (matched quests)
Attunement100%1 of 1 runs opened an equipped skill
Upkeep+13% manaextra mana per run vs Unarmed, median of 1 quest
IncantationNot measuredoutput tokens per run vs Unarmed
Runes-11%lines added per run vs Unarmed, median of 1 quest
Hero: Claude Code 2.1.305; weapon (model): claude-sonnet-5-5; charge (effort): medium; arena: sealed; invoked cast (the prompt named the skill). macOS 27.0.1, arm64. Curator.
Clear rate100%6 of 6 runs passed every check; Unarmed: 28 of 32 (matched quests)
Steadiness61%reliability: Wilson 95% lower bound
Buff+13 ppvs Unarmed, only matched quests (6 equipped runs; 32 control runs)
Instant casts6 of 6runs cleared on the first prompt; Unarmed: 27 of 32 (matched quests)
Attunement100%6 of 6 runs opened an equipped skill
Upkeep+9% manaextra mana per run vs Unarmed, median of 2 quests
IncantationNot measuredoutput tokens per run vs Unarmed
Runes-6%lines added per run vs Unarmed, median of 2 quests
Hero: Claude Code 2.1.305; weapon (model): claude-sonnet-5-5; charge (effort): medium; arena: sealed; ambient cast (the agent chose). macOS 27.0.1, arm64. Curator.
Clear rate100%4 of 4 runs passed every check; Unarmed: 28 of 32 (matched quests)
Steadiness51%reliability: Wilson 95% lower bound
Buff+13 ppvs Unarmed, only matched quests (4 equipped runs; 32 control runs)
Instant casts4 of 4runs cleared on the first prompt; Unarmed: 27 of 32 (matched quests)
Attunement0%0 of 4 runs opened an equipped skill
Upkeep+0% manaextra mana per run vs Unarmed, median of 2 quests
IncantationNot measuredoutput tokens per run vs Unarmed
Runes+1%lines added per run vs Unarmed, median of 2 quests
| QuestFixed task | Clear ratePassed runs | SteadinessReliability lower bound | BuffClear-rate points vs Unarmed | Instant castsFirst-prompt clears | CastsPrompts to clear | Cast timeAgent runtime | ManaCost at API list prices | IncantationOutput tokens | RunesLines added / removed | AttunementSkills opened | SetupApp, model, effort and sandbox |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Paladin: expense claims migrationQuest v1 | 100%3 of 3Unarmed: 76% (13 of 17) | 44% | +24 ppvs Unarmed | 3 of 3100%Unarmed: 13 of 17 | 1 | 38sUnarmed: 37s | 1,210 Unarmed: 1,072 | Not reportedUnarmed: 4.4k | +145 / -0Unarmed: +157 / -0 | 100% | Hero: Claude Code 2.1.305; weapon (model): claude-sonnet-5-5; charge (effort): medium; arena: sealed; invoked cast (the prompt named the skill). macOS 27.0.1, arm64. Curator.0 mishaps excluded. 2026-10-21 to 2026-10-21.Scrying (screenshot) Try this loadoutRecorded with quest v1 and spell commits |
| Paladin: expense claims migrationQuest v1 | 100%2 of 2Unarmed: 76% (13 of 17) | 34% | +24 ppvs Unarmed | 2 of 2100%Unarmed: 13 of 17 | 1 | 38sUnarmed: 37s | 1,119 Unarmed: 1,072 | Not reportedUnarmed: 4.4k | +167 / -0Unarmed: +157 / -0 | 0% | Hero: Claude Code 2.1.305; weapon (model): claude-sonnet-5-5; charge (effort): medium; arena: sealed; ambient cast (the agent chose). macOS 27.0.1, arm64. Curator.0 mishaps excluded. 2026-10-21 to 2026-10-21.Scrying (screenshot) Try this loadoutRecorded with quest v1 and spell commits |
| Paladin: RSVP Capacity and WaitlistQuest v7 | 100%1 of 1Unarmed: 100% (3 of 3) | 21% | +0 ppvs Unarmed | 1 of 1100%Unarmed: 3 of 3 | 1 | 2m 59sUnarmed: 2m 36s | 392 Unarmed: 347 | Not reportedUnarmed: Not reported | +476 / -35Unarmed: +537 / -37 | 100% | Hero: Claude Code 2.1.305; weapon (model): claude-haiku-5-5; charge (effort): low; arena: sealed; invoked cast (the prompt named the skill). macOS 27.0.1, arm64. Curator.0 mishaps excluded. 2026-10-21 to 2026-10-21.Scrying (screenshot) Try this loadoutRecorded with quest v7 and spell commits |
| Paladin: RSVP Capacity and WaitlistQuest v7 | 100%3 of 3Unarmed: 100% (15 of 15) | 44% | +0 ppvs Unarmed | 3 of 3100%Unarmed: 14 of 15 | 1 | 2m 10sUnarmed: 2m 4s | 3,836 Unarmed: 3,624 | Not reportedUnarmed: 19k | +439 / -33Unarmed: +457 / -33 | 100% | Hero: Claude Code 2.1.305; weapon (model): claude-sonnet-5-5; charge (effort): medium; arena: sealed; invoked cast (the prompt named the skill). macOS 27.0.1, arm64. Curator.0 mishaps excluded. 2026-10-21 to 2026-10-21.Scrying (screenshot) Try this loadoutRecorded with quest v7 and spell commits |
| Paladin: RSVP Capacity and WaitlistQuest v7 | 100%2 of 2Unarmed: 100% (15 of 15) | 34% | +0 ppvs Unarmed | 2 of 2100%Unarmed: 14 of 15 | 1 | 1m 49sUnarmed: 2m 4s | 3,492 Unarmed: 3,624 | Not reportedUnarmed: 19k | +435 / -32Unarmed: +457 / -33 | 0% | Hero: Claude Code 2.1.305; weapon (model): claude-sonnet-5-5; charge (effort): medium; arena: sealed; ambient cast (the agent chose). macOS 27.0.1, arm64. Curator.0 mishaps excluded. 2026-10-21 to 2026-10-21.Scrying (screenshot) Try this loadoutRecorded with quest v7 and spell commits |
Scrying (screenshots of runs)
The latest run of each setup. Local paths, names and ids are blurred; credentials are removed.
This spell has no recorded sessions in the curator's logs. Tried it? Send a spell report.