Engineering note ↗

Evidence-backed agent capabilities

Gaia Skill Tree

See what agents can do, who proved it first, and the evidence behind every claim.

Install the Gaia CLI curl -fsSL https://gaiaskilltree.com/install.sh | sh
Submit a skill Send public repository evidence for review Get a README badge Carry your Named Skill and earned stars
— skills / — named / — ultimates / updated
scroll

Registry honors

Hall of Heroes

The highest-ranked named implementations in the registry, kept on the public record.

Apex-first Evidence-backed Contributor-owned
View Hall of Heroes →
Populated by page-ia.js
Skills → Contributors →
— of — skills
Origin skills — the canonical implementation by the contributor who first published the technique. Marked with the golden avatar wreath inside the bar.
Browse Named Skills Trending this week Trust Leaderboard
Gaia Skill Tree
Loading…

Gaia is three names

Research proves it, the registry records it, the launcher runs it. The Skill Tree is the record.

Read the ecosystem explainer
Recommended

Install the CLI

curl -fsSL https://gaiaskilltree.com/install.sh | sh
iex (irm https://gaiaskilltree.com/install.ps1)
pipx install gaia-cli
Path A — DIY

Submit Your Repo Skills

Run three commands yourself.

  1. Initialise

    Non-interactive defaults; add --force only when replacing existing config.

    gaia init --yes
    py -m gaia_cli init --yes
    gaia init --yes
  2. Scan

    Detect demonstrated skills.

    gaia scan
    py -m gaia_cli scan
    gaia scan
  3. Push

    Preview with gaia push --dry-run, then submit.

    gaia push
    py -m gaia_cli push
    gaia push
Continue with Gaia Optional follow-ups
  1. Bond your agent

    Give an MCP-compatible agent access to the Registry.

    claude plugin install skill-heaven@gaia-skill-heaven
    💡Skill Heaven bundles its own summon MCP server, so there is nothing to configure. Standalone @gaia-research/mcp was decommissioned on 2026-08-19. See the MCP server guide for details.
  2. Open the Terminal UI

    Browse Gaia without memorising commands.

    gaia
    py -m gaia_cli
    gaia
    ◇The Terminal UI is experimental: keyboard-native, fuzzy searchable, and built for agent-driven work.
Path B — Agent

Prompt Your Agent

Your agent handles the mechanics and asks before submitting.

Work in my current repository and register it with Gaia. The Gaia CLI operates inside the target repository. If a target URL is provided and it is not already available locally, clone a temporary copy and work there.

1. Read the current instructions at https://github.com/gaia-research/gaia-skill-tree. Install Gaia if missing; if it already exists, upgrade `gaia-cli` with its package manager before continuing. Print `gaia version` and report the install method used.
2. Run `gaia init --yes` inside the target repository. If `.gaia/config.toml` already exists, inspect it first and ask before replacing it; only then use `gaia init --force --yes`.
3. Run `gaia scan`. If scanning fails, stop and show the error rather than continuing with stale state.
4. Run `gaia push --dry-run`. This must only print the proposed batch; do not open an issue or write an Intake submission.
5. Summarise detected skills, proposed evidence, attribution, warnings, and exactly what the real push will submit.
6. Ask: “Submit this Gaia Intake now? [Y/n]”
7. Only after I confirm, run `gaia push --yes` and return the resulting Intake issue or pull-request URL. Never use `git commit` or `git push` on the target repository, and do not edit Registry data by hand.
8. If a temporary clone was used, optionally ask if I want to delete it.
Prepare one standalone skill for Gaia Intake on my behalf. The Gaia CLI operates inside the target repository. If a target URL is provided and it is not already available locally, clone a temporary copy and work there.

Use https://github.com/gaia-research/gaia-skill-tree as the source of truth and follow `.github/ISSUE_TEMPLATE/new_skill_intake.yml` exactly.

1. Determine my authenticated GitHub username with `gh api user --jq .login` and show it to me.
2. Clarify the skill name, falsifiable capability, source repository, and any evidence I provide.
3. Check the existing Registry for overlap before proposing a new kebab-case ID.
4. Curate real, resolvable evidence. Include my supplied evidence, preserve its source, use GitHub `blob/` links rather than `tree/`, and never invent metrics or claims.
5. Treat this as `self-made` and attribute the named submission to my authenticated GitHub username only if the sources confirm I am the Origin Contributor. If they indicate another origin, stop and explain the attribution conflict.
6. Draft the complete Intake YAML, including `basic` or `fusion`, prerequisites, attribution, Grade B-or-better evidence, and the optional named block when valid. Validate it against the issue template; if Gaia is installed, also run `gaia push --from-file skills.yml --dry-run` and treat any failure as blocking.
7. Before opening anything, summarise the proposed skill, attribution, evidence, exact issue title, and the validated YAML. Ask: “Open this skill Intake as @<my-handle>? [Y/n]”
8. Only after I confirm, submit through the Skill Intake issue template—or run `gaia push --from-file skills.yml --yes`—and return the issue URL. Never use both routes, never commit or push to the target repository, and do not modify canonical Registry files directly.
9. If a temporary clone was used, optionally ask if I want to delete it.

Always review an agent's proposed plan before giving it permission to push.

How Your Skills Get Reviewed

Every contribution undergoes rigorous programmatic and contributor checks before merge, ensuring registry integrity. Click "Simulate Review" below to see it in action.

gaia-local-cli.sh
> Local command-line terminal. Awaiting simulation trigger...
github-actions-reviewer.sh
> Remote review gateway terminal. Awaiting simulation trigger...
Read Full Curation Codex →

The Ascension Cycle

Ascension

Rank is earned. Branch is chosen.

Awakened rank medallion

Act I · The trunk

Awakened

A real skill enters the tree. The first seal is set.

  • verified as real skill
  • standard branch
Named rank medallion

Act I · The trunk

Named

Given an identity, a signet, and a place in the atlas.

  • named implementation
  • standard branch
Evolved rank medallion

Act I · The trunk

Evolved

Reproducible, installable, and traceable to its source.

  • verified source
  • standard branch

Act II · The fork

One trunk. Two branches.

At three stars, every named skill has walked the same trunk. Awakened, Named, Evolved. The fourth star is where the tree forks. A named skill with suiteComponents ascends the Suite branch. A named skill without suiteComponents ascends the Unique branch. The choice is not made at promotion time; it is derived from the named skill's structure at read time. Read the derivation in the Suite or Unique branch reference in the Codex.

Suite · ordered structure Unique · standalone distinction

→ Unique

Extra rank medallion

Suite branch · ceremonial

Extra

The Suite rank is earned. The reinforced seal is set.

  • suiteComponents present
  • ordered fusion
Unique rank stellar medallion

Unique branch · void

Unique

The path bends. The standard structural rule does not apply.

  • named skill
  • no suiteComponents

→ Unique

Ultimate rank stellar medallion

Suite branch · ceremonial

Ultimate

The Suite 5★ rank name. Ultimate is the ceiling of the standard promotion path when the branch is Suite.

  • Suite branch
  • 5★ rank
Unique Ultimate black-hole medallion

Unique branch · void

Unique Ultimate

The Unique 5★ rank name. Ultimate is the shared 5★ label across both branches; the Unique prefix marks the branch.

  • Unique branch
  • 5★ rank

Ultimate is a rank, not a class. Every 5★ skill is Ultimate; the branch prefix tells you which lane it rose through.

→ Unique

Apex white-hot supernova medallion

Suite terminal · six predicates

Apex

The Suite terminal. Six predicates, all required. The Suite Apex Gate is unchanged from Yggdrasil I.

  1. aGradedOriginsGte5Five A-graded origins
  2. directNestedSuiteGte1One directly nested Suite
  3. depth2OnlyReachableGte1One depth-two-only origin
  4. overallGradeSOverall Trust Grade S
  5. sourceTenureDaysGte180AorSA or S source held for 180 days
  6. apexPromotionPrSignedPromotion pull request signed
Unique Impossible ember-void medallion

Unique terminal · five predicates

Unique Impossible

The five-predicate gate is an explicit bridge state. The Suite/Unique branch distinction is ratified under Yggdrasil III; the five-predicate gate itself remains provisional.

The Unique terminal. Five predicates, provisional. The Unique Impossible Gate is the Suite Apex Gate minus the directNestedSuiteGte1 predicate. The gate itself remains provisional.

  1. aGradedOriginsGte5Five A-graded origins
  2. depth2OnlyReachableGte1One depth-two-only origin
  3. overallGradeSOverall Trust Grade S
  4. sourceTenureDaysGte180AorSA or S source held for 180 days
  5. apexPromotionPrSignedPromotion pull request signed

Two terminals. Two gates. Both are Apex-grade trust, earned through opposite structural paths.

Evidence Grade

A skill earns its Overall Trust Grade by the strength of the evidence behind it. Four tiers (Bronze, Silver, Gold, Platinum) mark the aggregate Trust Magnitude that a skill's evidence pool reaches.

Platinum (S) ≥ 250 trust
Gold (A) ≥ 100 trust
Silver (B) ≥ 50 trust
Bronze (C) ≥ 20 trust
Ungraded

Each piece of evidence (a repo, a benchmark, a peer-reviewed paper) earns its own per-row Evidence Grade. The numbers above are the aggregate Trust Magnitude thresholds the rows must combine to clear. See the methodology report for the full formula.

Evidence Library → Read Trust Methodology → Benchmarks → Weekly Reports →

Meta Reports

Recent registry curation, schema migrations, and audit passes.

Engineering note

Discovery Is Not Admission: What 95 Agent-Skill Candidates Taught Us ↗

Five upstream release streams produced 95 candidate skills. A reviewed integration accepted 38, mapped 40 to existing capabilities, omitted 17, and deferred none.

 Registry Update

September Mid-Cycle Update: Frozen Ranks, Upstream Sync, and the Four-Tier Audit ↗

What changed in the registry between v8.7 and v8.12: two new skills admitted (trailhq/graft at 4★), four skills disentangled from an unrelated purge and restored — Matt Pocock's trio to 3★ and rico-favor/implement-with-discernment to 1★ — six suites synchronized to upstream releases, 48 rank recalibrations, and the Four-Tier Audit Taxonomy that now governs how skills are evaluated.

 Registry Update

September 9 Rank Calibration ↗

Updated named-skill ranks to match current Trust Magnitude grades and added Steward drift detection.
Open All Reports →
Permalink copied!