learning-extract
Skill2 starsChanged 4 months agoMITView on GitHub
After human wave-acceptance / tip fixes on a wave PR, extract structured programme learning (SPEC / SKILL / HARNESS / ENV) with stable L-* ids for Gateflow to persist and for ground-spec to cite. Closeout hop — does not write the Ground Report. Use Enter-at or /learning-extract after Pass-1 stop at wave-acceptance.
In this repository22 other files it gives its agents.
- ground-specSkill
- initiative-feasibilitySkill
- loop-specSkill
- pre-implementSkill
- purge-initiative-artifacts-appSkill
- spec-draftSkill
- spec-implementation-planSkill
- spec-technical-reviewSkill
- ensure-repo-graphSkill
- post-product-questionsSkill
- prd-codebase-mapSkill
- review-product-questionsSkill
- commit-workspaceSkill
- create-board-ticketsSkill
- open-draft-prSkill
- prd-impact-mapSkill
- prd-qualitySkill
- prd-thinkSkill
- purge-initiative-artifacts-metaSkill
- review-findingsSkill
- update-documentsSkill
- validate-requirementsSkill
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 1 is distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- could be automatic: a setting could enforce it
- Distinctive: line 38
Do not open auto-merge codify PRs.
- Very specific: line 48
Do not call Gateflow HTTP / DB as skill success.
- Very specific: line 103
When handoff_path is bound, overwrite that path with the same handoff: envelope before exit.
- Very specific: line 109
Follow prayog-skills/references/forge-side-effects.md#content-producers when pin forge.commit_workspace is not disabled.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 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
Learning extract
Producer of structured learning after Pass-1 (implement) and human
wave-acceptance / tip fixes. Gateflow may persist records in a global DB; this skill
only writes a workspace artifact + handoff. /ground-spec still owns the
wave Ground Report and §Contracts produced.
Do not require humans to write learning essays. Infer from repo + bind context. Ask taxonomy chips only when classification is ambiguous.
NON-NEGOTIABLE
- Resolve layout from
.harness/profile.yamlor references/layout-defaults.md. - Inspect the wave under test: bound ticket/initiative/wave/PR/run context when
present; Pass-1 tip vs human-fix window (
gitlog/diff); spec/plan/TASK rows; verify scripts /tests_readme; prior Ground Reports as needed. - Emit learning items with closed taxonomy:
SPEC,SKILL,HARNESS, optionalENV. One primary class per item. Prefer SPEC over SKILL when both fit. - Assign stable ids
L-01,L-02, … (seeprayog-skills/references/id-conventions.md). - Each item includes: summary, evidence (paths/commits), codify hint
(suggested skill / spec area / harness home), status
open|codified. Do not open auto-merge codify PRs. - Write durable artifact
{reports_dir}/Learning-Extract-{initiative}-W{N}.mdwith a human table and a fencedlearning_extract:YAML block (machine payload). This file is PURGE at initiative closure (see artifact-write-contract). - Empty
items: []only when there is no human-fix signal and the tip matches intent — state that rationale explicitly. If human fixes clearly exist and zero items → do notpass(usefindings/ fail-closed). - Do not author the full Ground Report REQ matrix or §Contracts produced —
that remains
/ground-spec. - Do not call Gateflow HTTP / DB as skill success. Worker ingest is the
consumer (H6). Follow
prayog-skills/references/forge-side-effects.md#content-producersfor optional workspace publish. - Ids / paths:
prayog-skills/references/id-conventions.md,prayog-skills/references/artifact-write-contract.md.
Taxonomy
| Class | Use when |
SPEC | Upstream product/TDD/acceptance incomplete or wrong |
SKILL | Playbook/prompt/procedure miss (agent followed process, still wrong) |
HARNESS | Verify/assert/setup wrong — harness would mislead |
ENV | Infra/secrets/cluster only — not product learning |
Inputs
- Initiative + wave — (REQUIRED)
- Workspace / PR tip — run head after human wave-acceptance
- Optional bind — ticket, PR/run ids from orchestrator
Prerequisite
- Pass-1 complete enough that a tip exists (typically after
wave-acceptancepark or equivalent human prove + patch) - Pin next on
pass:ground-spec
Process
- T0 Gather — initiative, wave N, plan/spec paths, tip SHA, fix window
- T1 Diff signal — detect human_fix_detected yes/no
- T2 Classify — draft
L-*rows; chip-ask only if ambiguous - T3 Write artifact — markdown + YAML fence per references/output-template.md
- T4 Handoff —
pass→ground-spec; dual-write baton when bound
Chain position
Illustrative — transitions SSOT: pinned workflow.yaml.
Pass-1: … → loop-spec → wave-acceptance (human) → wave-awaiting-closeout
Pass-2 Enter-at /human:
/learning-extract ← YOU ARE HERE
↓
/ground-spec → wave-done-action → wave-signoff (merge only)
Pass-2 closeout (`learning-extract` → `ground-spec`) closes the wave;
`wave-signoff` is merge/publish only — not a second human approve.Workflow handoff
- Append/emit envelope from
prayog-skills/references/handoff-envelope.md. Stagelearning-extract. - When
handoff_pathis bound, overwrite that path with the samehandoff:envelope before exit. - Derive
next_candidatesandhuman_checkpointfrom pinnedworkflow.yamlfor(stage: learning-extract, outcome). Sethuman_checkpoint: trueonly when the resolved next node'stypeishuman-checkpoint. - Happy path:
outcome: pass→ nextground-spec→human_checkpoint: false. - Follow
prayog-skills/references/forge-side-effects.md#content-producerswhen pinforge.commit_workspaceis notdisabled.
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.

