| name | LearnFrom |
| description | Extract session learnings and apply them as updates to rules, skills, and agents. USE WHEN session produced reusable patterns, corrections, or conventions worth capturing. |
| version | 0.1.0 |
| argument-hint | extract learnings from this session |
LearnFrom
Extract reusable learnings from the current session and apply them as updates to rules, skills, and agents in the current repo.
Workflow Routing
| Workflow | Trigger | Section |
|---|
| Analyze | "learn from this session", "extract learnings" | Analyze |
| Targeted | "add a rule for X", "update the agent to note Y" | Apply Changes |
Analyze the Session
First pass — find the friction points. Identify each area of the session where work cycled more than once: same problem revisited multiple turns, repeated "still doesn't work" exchanges, user pushback on the same theme across several messages, debugging loops that bounced between hypotheses. List the 1-3 biggest friction zones explicitly by topic ("QR code validation", "visual styling", "Python env"). Each friction zone is a learning gold mine — the cycling itself proves a missing rule, skill, or workflow guard would have prevented it. Skipping friction zones in favor of small surface fixes (a typo, a flag) is the most common LearnFrom failure mode. For each friction zone, ask: what specific instruction, if it had existed at session start, would have collapsed the cycle to one turn?
Second pass — scan for user-correction signals. Re-read the conversation specifically looking for: user messages containing "no", "don't", "actually", "wait", "stop", "that's wrong"; follow-up questions that reveal you missed something ("did you X?", "is it still Y?"); user edits or rewrites of your output; requests that imply a prior step should have been done differently. These encode behaviors the user actively wants changed; surface them alongside the friction zones.
Third pass — walk the scan checklist below. Walk through each category and list concrete findings before filtering.
Scan Checklist
- Corrections made — wrong assumptions that were fixed, approaches rejected by the user, things that didn't work the first time
- Tool behaviors learned — CLI flags, API quirks, platform constraints, output format surprises, command interactions (e.g., one command cleaning another's output)
- Packaging and deployment — build pipelines, release workflows, distribution patterns, CI/CD discoveries
- Cross-tool interactions — when tool A's output feeds tool B, ordering dependencies, cleanup side effects
- Patterns discovered — reusable conventions, architectural decisions, workflow improvements
- Process improvements — better ways to approach tasks discovered during work
Cross-reference against friction zones — items inside a friction zone matter more than isolated ones.
For each finding, apply the reusability test: will this come up again in a different session? If no, skip it. If uncertain, include it: a skipped learning is lost, an extra proposal can be rejected.
Pick the Right Repo First
Forge modules are one home for learnings, not the only home. Before scanning a module's rules/skills/agents, decide whether the learning even belongs in a forge module. Many "platform quirks" are actually shell-env exports or install steps that belong in dotfiles or mac-setup, not in a forge rule pile.
| Question | Target | Example |
|---|
| Can a shell-env export, alias, or shell-config change fix this? | dotfiles/ (zshenv, zshrc, zprofile) | export SSH_ASKPASS=/opt/homebrew/bin/ssh-askpass to fix YubiKey commit signing in subprocesses |
Can a brew install, defaults write, or macOS setup step fix this? | mac-setup/ | "install ssh-askpass via brew", "disable Spotlight indexing of node_modules" |
| Is this a decision about a specific project's identity or architecture? | That project's docs/decisions/ ADR | "we use EUPL-1.2", "single source of truth for default models" |
| Is this a forge-ecosystem convention or workflow? | Appropriate forge module's rules/, skills/, or agents/ | "skill bodies stay under 150 lines", "ADRs use the structured-madr schema" |
| Is this a personal preference about the user the AI should remember? | A personal-overrides module | "user prefers concise responses", "user works in Czech locale" |
| Is this an unfixable platform quirk — workaround only lives in head? | forge-core rules/KnownIssues.md (last resort) | "Obsidian Linter auto-formats frontmatter between read and write" |
KnownIssues.md is the last resort, not the default. If a config change, install step, or env export would actually fix the problem, the learning belongs at the source of that fix — never as a workaround documented in KnownIssues. Adding a fixable quirk to KnownIssues teaches future sessions to tolerate the breakage forever instead of resolving it once.
Scan Existing Artifacts FIRST
Before drafting any proposal, list every file in rules/, skills/, and agents/ of the target module. For each candidate learning, search by topic for an existing artifact that already touches the same area. The default outcome is a one-line edit to an existing file, NOT a new file.
Concrete signals you should be editing not creating:
- Topic overlap: existing rule covers the same domain (git, bash, markdown, ADRs)
- Adjacent guidance: existing skill mentions the same tool or workflow
- Sibling concept: existing rule covers the inverse or a related case
Only create a new rule or short file when the learning is genuinely orthogonal to everything that exists. If you find yourself writing a rule shorter than ~3 sentences, it almost certainly belongs as a paragraph in an existing file.
Exception — when friction reveals a missing domain skill, propose a new skill, not a paragraph in a general one. If the friction zone was about a specific tool, format, protocol, or domain operation that no existing skill covers (Czech SPAYD QR generation, GraphQL schema design, Stripe webhook handling, Apple Numbers automation), the right output is a new skill with reusable scaffolding — templates, validation steps, default config, examples — that prevents the friction next time. A bullet in SystematicDebug about "clarify failure mode" still has value, but it's the secondary capture; the primary one is the missing domain skill itself. Trigger conditions: friction took 5+ turns to resolve, the operation has multi-step structure (build → validate → render → verify), and the next user invocation would benefit from /SkillName rather than from a rule firing on every session. When in doubt, surface this option to the user alongside the edit-existing proposal so they can pick.
Determine the target artifact. The categorization decision matters — rules cost tokens every session; skills cost tokens only when invoked.
| Learning type | Target | Example |
|---|
| Always-relevant constraint | rules/ | "every text file ends with newline" |
| Task-specific guidance | skills/*/SKILL.md with paths: | "shell scripting pitfalls — auto-trigger on **/*.sh" |
| Skill workflow improvement | skills/*/ | "add note about tool limitation" |
| Agent instruction update | agents/ | "add guidance about deployment scope" |
| Existing file refinement | edit in place | "add RACI to required frontmatter list" |
If the guidance only matters when working on certain files (shell scripts, Python, Markdown), make it a skill with paths: frontmatter so it auto-triggers on relevant file edits — don't put it in rules/ where it loads on every session regardless of relevance.
Source vs Deployed; Local vs Upstream
LearnFrom always edits source files in the module root (rules/, skills/, agents/). Never edit deployed copies under .claude/, .codex/, .gemini/, or .opencode/ — those are forge install outputs that get overwritten on every sync.
Start with the source artifact relevant to the current workload, then check its provenance record (.provenance/<file>.yaml sidecar). If resolvedDependencies lists an upstream module (forge-core, forge-dev, etc.), the artifact is inherited content. Propose the change upstream; the local copy refreshes via forge install after the upstream merges. Edit the local copy directly only when the change is genuinely repo-specific (project-only convention, local override).
Skipping this check creates drift: a local edit to inherited content silently shadows future upstream improvements, and a local edit to a deployed copy is overwritten on the next sync.
Draft Proposals
For each identified learning, draft a concrete proposal in this priority order:
- Append to existing rule — one paragraph added to an in-topic file
- Append to existing skill body — for skill-scoped guidance
- Edit existing agent instructions — for agent-specific behavior
- New rule or skill — only when no existing artifact fits
Show the existing-file scan result alongside each proposal so the user can see what was considered and rejected.
Rules must fire on a concrete trigger. Abstract principles ("be careful with X") don't change behavior; concrete signals do ("when you're about to write Y, check Z"). Iterate the wording with the user — first drafts are usually too abstract OR too tied to one specific case. Filename should match the final framing — rename if the concept shifts during iteration.
Interactive Review
Present proposals in batches via AskUserQuestion (4 questions max per call).
For each proposal, show the target file and proposed change with a preview field carrying the literal content that would be written. Options: "Capture", "Adjust", "Skip".
For "Adjust": ask what to change, then re-present.
After the first batch, dig deeper before declaring done. Most sessions have 2-3 obvious learnings and several non-obvious ones surfaced only by re-scanning corrections, pushback, and stuck moments. Ask yourself "what else did the user have to correct?" before stopping.
Apply Changes
Write each confirmed proposal to its target file in rules/, skills/, or agents/.
Summary
List what was captured: rules created or updated, skills updated, agents updated, items skipped.
Constraints
- Keep rules concise (max 120 words per section per the rules
.mdschema)
- New rules follow the
.mdschema in rules/ if present
- Session-specific fixes are not rules, but tool behaviors and packaging patterns often are
- Captured artifacts describe the current truth, not where it was learned. No "learned from X", "discovered during Y", or "based on the Z codebase" (per StartFresh). The learning stands on its own.
- Validate against the target directory's
.mdschema before writing
