hydraia
Skill8 starsChanged 2 months agoMITView on GitHub
Use whenever the user asks to build, add, implement, or change a feature or functionality — or brings a user story or ticket to analyze, reports a bug or unexpected behavior, asks for a branch review, or wants a new app or service designed from scratch. Phase -1 triages the intent and routes it. Runs the complete non-negotiable development pipeline end to end — deep analysis and planning, sub-agent execution, and a double code-review loop — without asking the user which model, skill, or step to use. This is the default way features get built.
In this repository51 other files it gives its agents.
- adrSkill
- agency-tier-uiSkill
- api-designSkill
- banner-designSkill
- brainstormingSkill
- brandSkill
- cavemanSkill
- coding-standardsSkill
- db-optimizationSkill
- designSkill
- design-systemSkill
- design-to-codeSkill
- dispatching-parallel-agentsSkill
- django-securitySkill
- dotnet-patternsSkill
- e2e-testingSkill
- executing-plansSkill
- explainmeSkill
- finishing-a-development-branchSkill
- golang-patternsSkill
- greenfield-architectSkill
- karpathy-guidelinesSkill
- microservices-architectSkill
- mobile-design-framesSkill
- node-patternsSkill
- performance-tuningSkill
- production-auditSkill
- python-patternsSkill
- react-patternsSkill
- receiving-code-reviewSkill
- redesign-upliftSkill
- repo-scanSkill
- requesting-code-reviewSkill
- security-bounty-hunterSkill
- security-reviewSkill
- security-scanSkill
- slidesSkill
- springboot-patternsSkill
- springboot-securitySkill
- story-analysisSkill
- subagent-driven-developmentSkill
- systematic-debuggingSkill
- test-driven-developmentSkill
- ui-stylingSkill
- ui-ux-pro-maxSkill
- using-git-worktreesSkill
- using-superpowersSkill
- verification-before-completionSkill
- web-design-framesSkill
- writing-plansSkill
- writing-skillsSkill
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 3 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 21
NEVER load two phase files at once: Do not skip, reorder, or pre-load phases.
- Distinctive: line 22
Follow each phase file exactly, the way you would a step file.
- Distinctive: line 54
Thin orchestrator: keep this session's context small — dispatches carry pointers (plan path + task heading), reviewers read the diff from a …
- Very specific: line 75
Read fully and follow phases/phase--1-triage.md to begin the pipeline.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Hydraia — Agentic Development Pipeline
This skill defines the ONLY approved way to build a feature. Run every phase your level requires, in order, automatically. Never ask the user which model, skill, or reviewer to use — those decisions are already made here. Never skip a step the level requires. Never pause for "should I continue?" between phases.
Announce once at the start: "Running the Hydraia pipeline." Then proceed silently through the phases, narrating at most one short line per phase.
Step-file architecture (how to run this pipeline)
The pipeline body lives in phases/*.md, loaded just-in-time. Read one phase
file fully, execute it, then load the next only when its ## NEXT directs you.
- NEVER load two phase files at once. Do not skip, reorder, or pre-load phases.
- Follow each phase file exactly, the way you would a step file.
phases/model-policy.mdandphases/token-discipline.mdare always-on reference (summarized below), not sequential steps — consult them as needed.
Always-loaded facts (carry these for the whole run)
- Route → phases (decided in Phase -1 triage):
feature0–6 (full) ·plan0–3 then stop ·story-1–3 (PO-first) ·perf/db-1–6 (measurement-first) ·architect-1–6 (greenfield) ·review5–6 ·graphcodegraph only. Ambiguous route →AskUserQuestion. - Levels (1 Light / 2 Standard / 3 Deep): Phase -1 picks ONE level by rule from
three stated facts — intent gaps, irreversibles, footprint (files, public surface,
security surface,
gate.yamloverlap) — and announces it; it does not ask. Level 1 = one-file spec-plan + one Sonnet executor + judge +security-scan; Level 2 = spec + plan-contract + Sonnet executors + judge + security floor (default); Level 3 = full ceremony. Security surface forces ≥2. The human overrides by saying so or viacustomize.toml[pipeline].level. Agile stays a separate mode for epic-sized work (≥2 independent shippable goals) and branches throughphases/phase-2a-decompose.md→phases/agile-orchestrator.md. - Model policy (summary): Opus judges, Sonnet builds. This orchestrator session
runs on Opus (any current generation) for triage, design, plan and grading findings.
Opus sub-agents are only the judges (
hydraia-reviewer,security-reviewer); every other agent is pinned to Sonnet, generic agents always get an explicitmodel, and executors reach Opus only under Max-quality routing.hooks/agents.shenforces it (Opus gate). Full policy inphases/model-policy.md. - Convergence contract (always on): a run ends
DONE,DONE_WITH_FOLLOWUPSorBLOCKED(evidence)— never in an open-ended loop. Pre-existing failures (Phase-0 baseline) are asked about, never silently fixed. Fixes go through executors ([fix:<slug>]), never inline edits by this session. Runtime breakers end loops: same failure 2× → NO PROGRESS, 3× → STALLED (hooks/verifyloop.sh); fix budget and review cycles (hooks/agents.sh); plan scope (hooks/blastgate.sh); plan contract (hooks/plancheck.sh). When a breaker trips, stop and report — never work around it. - Thin orchestrator: keep this session's context small — dispatches carry pointers (plan path + task heading), reviewers read the diff from a file, sub-agent reports are short, investigations return summaries. Executors already start clean per task.
- Execution routing: computed at Phase 3 and announced (asked only at Level 3) —
Balanced (Sonnet all), Economy (Haiku for mechanical + Sonnet for logic/ui),
Max quality (Opus for logic/ui), or Hand-off (freeze the plan, execute later
on a cheap external runtime). The recommendation shows what each option costs/gets. Each plan task carries an
Exec class(mechanical/logic/ui/qa) that Phase 4 maps to a model.customize.toml[executor].routingcan pin a choice and make the run non-interactive. - Customization: an optional
customize.tomloverrides the executor routing / per-class model / dispatch recipe (read in Phase 4) and the Phase-5 pass-2 reviewer panel (read in Phase 5). Precedence: repo<artifacts-base>/custom/hydraia.toml> global~/.config/hydraia/custom/hydraia.toml> shippedskills/hydraia/customize.toml. Unparseable override → warn and fall back to shipped defaults. The always-on security gate is NOT customizable. - Token discipline is background and always on (summary in
phases/token-discipline.md): thecavemanskill compresses internal comms only — never code, specs, or plans.
FIRST STEP
Read fully and follow phases/phase--1-triage.md to begin the pipeline.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

