graduate
Skill2 starsChanged 5 months agoMITView on GitHub
Promote a project's proven lessons, skills, and workflows into the shared
In this repository46 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- forge-opencode/CLAUDE.mdCLAUDE.md
- forge-adoptSkill
- forge-arch-proSkill
- forge-archSkill
- forge-autopilotSkill
- forge-autopilot-stopSkill
- forge-build-proSkill
- forge-buildSkill
- forge-deploySkill
- forge-doctorSkill
- dreamer-runSkill
- forge-evalSkill
- forge-feedbackSkill
- flowSkill
- forge-force-advanceSkill
- health-checkSkill
- forge-initSkill
- forge-monitorSkill
- forge-orchestrateSkill
- forge-plan-proSkill
- forge-planSkill
- preflightSkill
- forge-product-proSkill
- forge-productSkill
- forge-releaseSkill
- forge-resolveSkill
- forge-resumeSkill
- forge-retroSkill
- forge-reviewSkill
- forge-rulesSkill
- forge-set-profileSkill
- skill-creatorSkill
- forge-spec-proSkill
- forge-specSkill
- forge-sprint-proSkill
- forge-sprintSkill
- forge-srs-proSkill
- forge-srsSkill
- forge-statusSkill
- forge-trace-matrixSkill
- forge-uninstallSkill
- forge-validateSkill
- watchSkill
- watch-stopSkill
- forge-whySkill
The file
We split this skill into the rules it gives an agent
5 rules. About 0 are common, 3 are distinctive, and 2 are 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.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 55
All three modes go through one thin CLI over the shared driver.
- Distinctive: line 86
Promote on request: When the user wants graduation applied, run the real scan and report the per-tier promoted summary (tier → count → the …
- Distinctive: line 88
Relay honestly: If a tier promoted nothing, say so for that tier rather than implying the whole run promoted artifacts.
- Distinctive: line 94
After running, these should hold:
- Very specific: line 124
Run /forge:graduate to apply this."
≈ “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
forge-graduate — promote proven lessons, skills & workflows to ~/.forge
/forge:graduate is the on-demand front end to Forge's graduation layer: the mechanism
that lifts what proved itself in one project into a shared ~/.forge store, then recalls it
into your other projects. It runs the same graduate() driver the session-start hook runs
silently in the background — there is no second promotion path — so running it by hand just
makes that cross-project learning observable and explicit.
The driver loops registered-projects × tiers. Each tier is fail-soft and isolated: one tier's
fault degrades only that tier, never the others, and the driver never raises.
The three tiers and their gates
- lessons (
global-lessons.yaml) — a lesson graduates when the same concept (by trigger similarity) appears in ≥3 distinct projects with summed frequency ≥2. Breadth is the signal: a lesson learned everywhere is worth carrying everywhere. - skills (
global-skills.yaml+~/.forge/skills/<slug>/) — an approved skill graduates when its ExpeL weight is positive and it has been reused ≥2 times. Skills are single deliberate artifacts, so the gate is depth (reuse), not breadth. Recalled into other projects by symlink (project/plugin always wins on a name clash). - workflows (
global-workflows.yaml+~/.forge/workflows/<name>.yaml) — a.forge/workflows/<name>.yamlthat validates clean and has ≥2 successful runs in.forge/events.jsonlgraduates. Recall is the workflow loader's search path, TTL-filtered.
A 30-day last_used TTL governs recall for every tier: stale global entries decay out of
recall (they stay in the store but are not surfaced).
When to Use
/forge:graduate— run a real graduation scan now and promote everything past its gate./forge:graduate --dry-run— preview what would promote, writing nothing./forge:graduate list— enumerate the current global store per tier.
When NOT to Use
- You just want to advance the 12-stage pipeline → that's
/forge:build,/forge:plan, etc. - You want to run a workflow → that's
/forge:flow, not graduation. - Nothing has met a gate yet (a brand-new project) → graduation is a clean no-op; that's expected, not a failure.
The CLI
All three modes go through one thin CLI over the shared driver. --global-dir defaults to
~/.forge; override it only for testing.
List the global store (per tier: entry count + each entry's key and last_used):
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/_graduation.py" listDry-run preview (per-tier would-promote summary; writes nothing — no store files, no copied skill/workflow artifacts):
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/_graduation.py" --dry-runForce a real scan (collect → gate → promote across every registered project; the default
action). This writes the per-tier global stores and copies graduated skill/workflow artifacts
into ~/.forge:
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/_graduation.py"
# 'scan' is the explicit, equivalent form:
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/_graduation.py" scanSteps
- Default to a preview. Unless the user explicitly asked to promote, run the list view
and/or
--dry-runfirst and relay the per-tier summary. A dry-run never writes, so it is always safe to show before committing to a real scan. - Promote on request. When the user wants graduation applied, run the real scan and report the per-tier promoted summary (tier → count → the keys promoted).
- Relay honestly. If a tier promoted nothing, say so for that tier rather than implying the whole run promoted artifacts. If a tier degraded (its fault is isolated), report it as skipped this run.
Verification
After running, these should hold:
--dry-runprinted a would-promote preview and wrote nothing (~/.forgeis unchanged — no newglobal-*.yaml, no copiedskills//workflows/artifacts).- A real scan that promoted a workflow/skill left a matching entry in the corresponding
global-*.yamland a copied artifact under~/.forge/. listreflects exactly the entries present in the store files (counts match).- The command never crashed: even an unreadable store degrades to a clean message, not a traceback.
Examples
User: "/forge:graduate list"
Claude: [runs the list view] "Global store at ~/.forge: lessons: 2 entries skills: 1 entry - lint-fix (last_used: 2026-06-03) workflows: 1 entry - shipit (last_used: 2026-06-02)"
User: "/forge:graduate --dry-run"
Claude: [runs the dry-run; writes nothing] "Dry-run graduation preview (nothing written): lessons: nothing to promote skills: nothing to promote workflows: Would promote 1 — shipit
Run /forge:graduate to apply this."
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.

