Manual Curation Pipeline

The exact commands to take a skill from discovery to a closed intake. No AI required. Every phase with every command, typed by a human.

Pipeline overview

#	Phase	What you produce	Gate
1	gaia-curate	A `discovery-packet-v2` JSON in `registry-for-review/discovery-packets/`	Human L4
2	L4 Review	You append `l4Resolution` to the packet	Hard stop — must be you
3	gaia push → PR	Intake issue opened; draft PR on a `review/meta` branch	CI green
4	ev-pipeline	Verified evidence lake ready for ingest	Human approves evidence
5	gaia-ingest	Evidence rows written, TM appraised, stars calibrated	Human approves calibration
6	Close	PR merged, intake issue closed with findings comment	Done

gaia-curate — Discover & package the skill

You have a real repo with a SKILL.md. Run prefill, apply the 6-rule mapping decision, write the packet, validate. Nothing touches the registry yet.

Step 1 — Preflight

bashclick to copy

# confirm gaia CLI is installed
gaia --version

# confirm gh CLI is authenticated
gh auth status

# confirm python3 available
python3 --version

Step 2 — Snapshot the generic list

Save this file. It's the trust anchor for packet validation. Don't overwrite it mid-run.

bashclick to copy

gaia dev list --generic --json > /tmp/generic-snapshot.json
wc -l /tmp/generic-snapshot.json

Step 3 — Run prefill against the SKILL.md URL

bashclick to copy

gaia dev prefill contributor/skill-name \
  --name "Skill Name" \
  --description "Verbatim from SKILL.md frontmatter" \
  --url "https://github.com/owner/repo/blob/main/SKILL.md" \
  --stdout > /tmp/prefill-output.json

cat /tmp/prefill-output.json | python3 -m json.tool | head -80

ℹ

Drop --stdout and gaia dev prefill writes the packet straight to registry-for-review/discovery-packets/<candidate-id-slug>.json — there is no --output flag.

Step 4 — Apply the 6-rule mapping decision

Rule (first match wins)	Decision	When
1	`NOT_A_SKILL`	`artifactGate != "valid-skill"`
2	`DUPLICATE`	`exactDedupe != null`
3	`DEFER`	`ambiguity != null`
4	`NEW_GENERIC`	`nTotal == 0`
5	`MAP`	`nStrong == 1` — use that genericId
6	`DEFER`	All other cases (weak only, or nStrong ≥ 2)

ℹ

Count nStrong = entries with matchTier == "strong" in mappingOptions[]. Count nTotal = total entries. Do not re-rank.

Step 5 — Write and validate the discovery-packet-v2

Update the packet at registry-for-review/discovery-packets/owner--skill-name.json with your Step 4 decision. Key fields (matching what validate_discovery_packet.py actually checks — decision is an object, not a string, and the top-level field is contractVersion, not schemaVersion):

json — minimum required fieldsclick to copy

{
  "contractVersion": "discovery-packet-v2",
  "candidateId": "owner/skill-name",
  "source": { "canonicalUrl": "<blob-url>", "sourceLane": "source-repository" },
  "normalized": { "name": "verbatim from frontmatter", "description": "verbatim" },
  "exactDedupe": { "matched": false },
  "mappingOptions": [ /* copy from prefill-output.json — each needs genericId, rationale, similarity, matchTier */ ],
  "decision": { "value": "MAP", "reasonCode": "MAP_EXISTING_GENERIC", "genericId": "<id from strong option>" },
  "flags": [ /* copy from prefill-output.json */ ],
  "genericSnapshot": { "capturedAt": "<ISO timestamp>", "command": "gaia dev list --generic --json", "contentSha256": "<hash>", "mappingOptionsSha256": "<hash>", "generics": [ /* snapshot contents */ ] },
  "lifecycle": [ "discovered", "fetched", "parsed", "normalized", "deduped", "mapped", "review-ready" ]
}

⚠

The "review-ready" lifecycle end state requires the full six-stage prefix shown above. gaia dev prefill (Step 3) only ever emits ["discovered", "deferred"] — that's by design, not a gap: prefill just ranks mapping options, it never fetches or reads anything. Advancing to "fetched" and "parsed" is on you, the curator who actually opened the URL. Add source.hostRepository (the repo URL), source.fetchedAt (an ISO timestamp for when you fetched it), and source.contentSha256 (a SHA-256 of the raw SKILL.md text — sha256sum <file> or shasum -a 256 <file>) for "fetched"; add source.frontmatter.name and source.frontmatter.description — the same verbatim values you already passed to Step 3's --name/--description — for "parsed". Run the validator after every edit rather than guessing.

bash — validateclick to copy

python3 .agents/skills/gaia-curate/scripts/validate_discovery_packet.py \
  --generic-snapshot /tmp/generic-snapshot.json \
  registry-for-review/discovery-packets/owner--skill-name.json

# Must exit 0 with no MALFORMED_PACKET or MISSING_REQUIRED_FIELD errors

L4 Review — You ratify the topology

No one else does this step. Check the packet, append l4Resolution, re-validate.

What to check

Mapping decision is correct — does selectedGenericId match the skill's actual capability?
If NEW_GENERIC: is the proposed name and description vendor-neutral and falsifiable?
source.url is a real blob/ GitHub URL — not tree/ or a listing page.
normalized.description is verbatim from the SKILL.md frontmatter — not paraphrased.
Similarity score (mappingOptions[].similarity) and matchTier make sense.

Append l4Resolution

json — append to packetclick to copy

"l4Resolution": {
  "generic": { "id": "final-generic-id", "name": "Human Readable Name", "description": "Vendor-neutral, falsifiable.", "type": "basic", "prerequisites": [] },
  "named": { "contributor": "github-handle", "skillName": "kebab-case-name" },
  "upstreamSkillFileUrl": "https://github.com/owner/repo/blob/main/SKILL.md"
}

bash — re-validate after appendingclick to copy

python3 .agents/skills/gaia-curate/scripts/validate_discovery_packet.py \
  --generic-snapshot /tmp/generic-snapshot.json \
  registry-for-review/discovery-packets/owner--skill-name.json

gaia push → Draft PR — Open the intake issue

bashclick to copy

git checkout -b review/meta/github-handle--skill-name

# Dry run first
gaia push --from-file registry-for-review/discovery-packets/owner--skill-name.json --dry-run

# Real run — opens intake issue (auto-labels: intake + needs-triage)
gaia push --from-file registry-for-review/discovery-packets/owner--skill-name.json

git add registry-for-review/discovery-packets/owner--skill-name.json
git commit -m "feat(intake): curate contributor/skill-name [closes #ISSUE]"
git push -u origin review/meta/github-handle--skill-name
gh pr create --draft --title "feat(intake): ..." --body-file /tmp/pr-body.md

intake:* label lifecycle

These are machine gates — intake-approval.yml watches for them and fires automatically.

Label	Who applies	What it triggers
`intake`	Auto (`gaia push`)	Routes into intake queue
`needs-triage`	Auto (`gaia push`)	Marks as awaiting review
`intake:topology-approved`	Maintainer only	Workflow fires: adds `evidence-review`, removes `needs-triage`
`intake:evidence-review`	Auto (workflow)	Signals agent to prepare Stage-1 seed
`intake:evidence-ready`	Agent / you	Evidence plan ready for human approval
`intake:evidence-approved`	Maintainer only	Opens the single draft promotion PR
`intake:rejected`	Maintainer only	Closes intake; no automation fires

⚠

intake:evidence-approved requires both intake:evidence-review and intake:evidence-ready to already be present. The workflow errors if either is missing.

bash — approve topology when ready (maintainer only)click to copy

gh issue edit <ISSUE_NUMBER> --add-label "intake:topology-approved"
# Fires intake-approval.yml — watch it:
gh run list --workflow=intake-approval.yml

ev-pipeline — Verify evidence before ingesting

Operates on evidence/by-type/ files only. Never touches registry/. Each phase produces review artifacts; you decide what to approve.

Phase 0 — ev-discovery (skippable)

The only phase that searches the web for new evidence. Skip for routine Stage-1 intakes. Run when you need benchmark-result, arxiv, peer-review, or richer social-signal rows.

⚠

No gaia CLI command runs this phase — there is no gaia dev discover. Trigger the ev-discovery agent skill directly (/ev-discovery in Claude Code), naming the skill id, source repo, and evidence types needed. It requires FIRECRAWL_API_KEY and skips gracefully if absent. This is the one phase on this page that needs an AI agent — every other phase is a plain command.

Discovered rows get appended to evidence/by-type/<type>.md under the same source-input format Phase 1 already reads — see Phase 1 below.

Phase 1 — ev-collection

⚠

This runs on the entire registry — not just your new skill. generate_source_dump.py reads all named-skill evidence rows across every skill in registry/named/. Your new skill's rows appear alongside all others. That is expected.

bashclick to copy

python3 evidence/scripts/generate_source_dump.py \
  --output-dir evidence --by-type-dir evidence/by-type

python3 evidence/scripts/compile_data_lake.py \
  --sources evidence/by-type --lake evidence

Phase 2 — ev-star-verification

bashclick to copy

# Check auth first — zero stars from an unauth'd call is poison data
gh auth status
gh api repos/OWNER/REPO --jq '.stargazers_count'
# Flag rows where recorded stars differ from live count. Do NOT silently update.

Phase 3 — ev-adversarial-audit

bash — check for common issuesclick to copy

# Find github tree/ URLs that should be blob/
grep -rn "github.com/.*/tree/" evidence/by-type/

# Find subjective wording not backed by source
grep -rni "elite\|high-quality\|best-in-class\|leading" evidence/by-type/

Phase 4 — ev-link-validation

bashclick to copy

python3 evidence/scripts/validate_sources.py
# Or a 10-URL sample first:
python3 evidence/scripts/validate_sources.py 10

gaia-ingest — Write evidence to the named skill

First time mutating registry/. Docs rebuild is skipped by default — no flag needed. Run gaia dev build once after all rows, or pass --build on a single row instead. You approve calibration before it runs.

Ingest each approved evidence row

bash — repo-ownclick to copy

GAIA_OPERATOR_OVERRIDE=1 gaia dev evidence contributor/skill-id \
  "https://github.com/owner/repo" \
  --type repo-own --commits 847 --contributors 12 \
  --notes "Official upstream repo. Counts verified via GitHub API." \
  --source-started-at 2024-03-01

bash — github-stars-ownclick to copy

GAIA_OPERATOR_OVERRIDE=1 gaia dev evidence contributor/skill-id \
  "https://github.com/owner/repo/blob/main/SKILL.md" \
  --type github-stars-own --stars 3421 --skill-count-in-repo 1 \
  --notes "Official SKILL.md. Star count verified live via gh api." \
  --source-started-at 2024-03-01

bash — build + appraise + calibrateclick to copy

# Single build after all rows are written
GAIA_OPERATOR_OVERRIDE=1 gaia dev build

# Appraise Trust Magnitude
PYTHONPATH=src python3 scripts/trust_appraise.py --skill contributor/skill-id

# Only after you approve the proposed calibration:
GAIA_OPERATOR_OVERRIDE=1 gaia dev calibrate contributor/skill-id 2★
GAIA_OPERATOR_OVERRIDE=1 gaia dev validate

TM grade reference

Trust Magnitude is the sole promotion gate under Yggdrasil II. Thresholds sourced from src/gaia_cli/grading.py and META.md §1.1.

Grade	TM floor	Stars it gates	Notes
ungraded	< 20	1★–2★	Stage-1 minimum (Awakened / Named baseline)
C	≥ 20	up to 3★	Badge unlock floor at 2★ Named
B	≥ 50	up to 3★	Solid adoption signals (Evolved)
A	≥ 100	gates 4★	4★ Extra/Unique gate — Origin + TM ≥ 100
S	≥ 250 + diversity	gates 5★–6★	Diversity gate: ≥3 evidence types + non-self-producible row. 5★ = TM ≥ 250. 6★ Apex adds 6-predicate gate.

ℹ

The per-star Evidence Floor is retired. TM is the sole gate. But the Star Bar (installability, 3★+ requires a verified GitHub blob/ link) still applies.

Close — Merge PR & post intake comments

bashclick to copy

gh pr ready <PR_NUMBER>
gh pr checks <PR_NUMBER>

# When CI is green:
gh pr merge <PR_NUMBER> --subject "feat(registry): ingest contributor/skill-name [closes #ISSUE]"

# Post findings comment on PR (evidence table + TM + badge status)
gh pr comment <PR_NUMBER> --body-file /tmp/pr-close-comment.md

# Post closing comment on intake issue (path-to-promotion, @contributor tag)
gh issue comment <ISSUE_NUMBER> --body-file /tmp/issue-close-comment.md
gh issue close <ISSUE_NUMBER>

# Regenerate Class S site artifacts
GAIA_OPERATOR_OVERRIDE=1 gaia dev docs
git add docs/graph/ && git commit -m "chore: regenerate Class S artifacts" && git push

Common errors

UNTRUSTED_GENERIC_SNAPSHOT

The contentSha256 in the packet doesn't match the snapshot file passed to --generic-snapshot.

bash — recompute the hashclick to copy

python3 -c "import json,hashlib; d=json.load(open('/tmp/generic-snapshot.json')); print(hashlib.sha256(json.dumps(d,sort_keys=True,separators=(',',':')).encode()).hexdigest())"

INVALID_GENERIC_SELECTION — mapped ID not in snapshot

You used a genericId that wasn't in the snapshot you captured. Re-capture the snapshot after adding the new generic, or use an ID that's actually there.

gaia dev evidence fails — "unknown evidence type"

bashclick to copy

gaia dev evidence --help
# Lists all supported --type values and required flags

gaia dev validate fails after ingest

bashclick to copy

GAIA_OPERATOR_OVERRIDE=1 gaia dev validate --verbose
# Read the first error — usually a missing field or schema mismatch

Cheat sheet — every command in order

bash — full pipeline sequenceclick to copy

# ── Phase 1: Curate ──────────────────────────────────────────
gaia dev list --generic --json > /tmp/generic-snapshot.json
gaia dev prefill contributor/skill-name --name "Skill Name" --description "..." --url "<blob-url>" --stdout > /tmp/prefill-output.json
# ... write packet JSON ...
python3 .agents/skills/gaia-curate/scripts/validate_discovery_packet.py --generic-snapshot /tmp/generic-snapshot.json <packet>.json

# ── Phase 2: L4 ──────────────────────────────────────────────
# Append l4Resolution to packet, re-validate

# ── Phase 3: Push + PR ───────────────────────────────────────
git checkout -b review/meta/handle--skill
gaia push --from-file <packet>.json --dry-run
gaia push --from-file <packet>.json           # auto-labels issue: intake + needs-triage
git add . && git commit -m "feat(intake): ..." && git push -u origin <branch>
gh pr create --draft --title "..." --body-file /tmp/pr-body.md
gh issue edit <ISSUE> --add-label "intake:topology-approved"  # maintainer only
gh issue edit <ISSUE> --add-label "intake:evidence-approved"   # maintainer only, after evidence-ready

# ── Phase 4: ev-pipeline ─────────────────────────────────────
# Phase 0 (skippable, no CLI equivalent — run the ev-discovery agent skill instead)
# /ev-discovery contributor/skill "<repo-url>" arxiv,peer-review,social-signal
# Phase 1 (whole-registry lake rebuild)
python3 evidence/scripts/generate_source_dump.py --output-dir evidence --by-type-dir evidence/by-type
python3 evidence/scripts/compile_data_lake.py --sources evidence/by-type --lake evidence
gh api repos/OWNER/REPO --jq '.stargazers_count'   # phase 2 star check
grep -rn "github.com/.*/tree/" evidence/by-type/      # phase 3 adversarial
python3 evidence/scripts/validate_sources.py          # phase 4 link health

# ── Phase 5: gaia-ingest ─────────────────────────────────────
GAIA_OPERATOR_OVERRIDE=1 gaia dev evidence contributor/skill "<url>" \
  --type repo-own --commits N --contributors N --notes "..." --source-started-at YYYY-MM-DD
# ... repeat for each evidence type ...
GAIA_OPERATOR_OVERRIDE=1 gaia dev build
PYTHONPATH=src python3 scripts/trust_appraise.py --skill contributor/skill
GAIA_OPERATOR_OVERRIDE=1 gaia dev calibrate contributor/skill N★  # after your approval
GAIA_OPERATOR_OVERRIDE=1 gaia dev validate

# ── Phase 6: Close ───────────────────────────────────────────
gh pr ready <PR> && gh pr checks <PR>
gh pr merge <PR> --subject "feat(registry): ... [closes #ISSUE]"
gh pr comment <PR> --body-file /tmp/pr-close-comment.md
gh issue comment <ISSUE> --body-file /tmp/issue-close-comment.md
gh issue close <ISSUE>
GAIA_OPERATOR_OVERRIDE=1 gaia dev docs

Want to automate this? Become a contributor.

Once you've curated a skill manually, you can contribute back to the registry directly using gaia dev commands — the same CLI the pipeline runs under the hood. No intake queue. Direct mutations with your approval.

What you want to do	Command
Add a generic skill node	`gaia dev add "Name" --id kebab-id --type basic --description "..."`
Add a named implementation	`gaia dev named --contributor handle --skill slug --generic id --upstream <blob-url>`
Add an evidence row	`gaia dev evidence contributor/skill "<url>" --type repo-own --commits N ...`
Appraise Trust Magnitude	`PYTHONPATH=src python3 scripts/trust_appraise.py --skill contributor/skill`
Calibrate star level	`gaia dev calibrate contributor/skill N★`
Fuse a suite capstone	`gaia dev fuse <id> --name "..." --prereqs id1,id2 --named-capstone contributor/slug`
Validate registry	`GAIA_OPERATOR_OVERRIDE=1 gaia dev validate`
Regenerate site artifacts	`GAIA_OPERATOR_OVERRIDE=1 gaia dev docs`

✓

All gaia dev mutations require GAIA_OPERATOR_OVERRIDE=1 or maintainer permissions. Read Contributing §1C before making direct registry changes. Or run gaia curate <url> to automate the full pipeline with 2 human gates.

Manual Curation Pipeline

On this page

Pipeline overview

gaia-curate — Discover & package the skill

Step 1 — Preflight

Step 2 — Snapshot the generic list

Step 3 — Run prefill against the SKILL.md URL

Step 4 — Apply the 6-rule mapping decision

Step 5 — Write and validate the discovery-packet-v2

L4 Review — You ratify the topology

What to check

Append l4Resolution

gaia push → Draft PR — Open the intake issue

intake:* label lifecycle

ev-pipeline — Verify evidence before ingesting

Phase 0 — ev-discovery (skippable)

Phase 1 — ev-collection

Phase 2 — ev-star-verification

Phase 3 — ev-adversarial-audit

Phase 4 — ev-link-validation

gaia-ingest — Write evidence to the named skill

Ingest each approved evidence row

TM grade reference

Close — Merge PR & post intake comments

Common errors

UNTRUSTED_GENERIC_SNAPSHOT

INVALID_GENERIC_SELECTION — mapped ID not in snapshot

gaia dev evidence fails — "unknown evidence type"

gaia dev validate fails after ingest

Cheat sheet — every command in order

Want to automate this? Become a contributor.