Lore

Lore

Your sessions are your lore. This skill turns a user's real coding history into installed skills. A deterministic engine (scripts/mine-skills.mjs) parses their SpecStory transcripts - from every agent SpecStory captures (Claude Code, Codex CLI, Cursor CLI, Gemini CLI, Factory Droid, DeepSeek, Antigravity, ...) - into a persistent corpus of beats, and returns corroborated candidates. You - the calling agent, whichever harness you are - supply all judgment: name them, discard the generic ones, and interactively forge the good ones into SKILL.md packages grounded in the user's own commands.

The engine does the retrieval and counting; you do the synthesis. Do not try to read raw transcripts yourself - they can be hundreds of thousands of lines. Run the engine and work from its output.

This skill is harness-portable (agentskills.io format). Where it names a specific tool (e.g. AskUserQuestion), treat that as "use your harness's equivalent; fall back to plain chat."

Voice: when narrating to the user, talk about mining their lore and skill candidates - e.g. "I'll mine your lore here in <project> for skill candidates." Reserve the word forge for the final act only: creating the skills the user selected (Step 4). Never describe mining, judging, or candidates as "forge-…" anything.

OUTPUT CONTRACT - three LAWS, read before emitting anything to the user

Named failure mode #1 (2026-06-09, BearClaude run): the agent indexed, deep-mined four skills, then jumped straight to AskUserQuestion with bare option labels - ZERO dossiers rendered in chat. The user declined every question because they had nothing to judge by. The entire mining run was wasted.

Named failure mode #2 (2026-06-09, BearClaude run, SAME DAY, fresh session, LAWs in effect): the agent narrated phases correctly, did verification reads, then asked again with NO dossier message - its last message before the question was process narration ("CodeMirrorBundle is alive in today's repo…") - and the question text falsely claimed "dossiers above". Lesson: a felt self-check is not a check. Compliance must be MECHANICAL: the sentinel line below is the check, not your impression.

Named failure mode #3 (2026-06-10, teammate's machine, Opus 4.8, plan-mode path): the agent DID use plan-mode curation but presented a THIN plan - skill names and skip reasons with the dossiers summarized away - so the user approved a forge they never saw the evidence for. The plan UI makes skipping the display step impossible, not skipping the content. Lesson: the plan body must BE the engine's plan render artifact (Step 3), which embeds every card verbatim and ends with the LAW 1 sentinel. In Claude Code this is now HOOK-ENFORCED: a PreToolUse hook in this skill's frontmatter denies any ExitPlanMode whose plan is not that artifact.

LAW 1 - DOSSIERS BEFORE CANDIDATE QUESTIONS, PROVEN BY SENTINEL. This law governs candidate decisions - any prompt where the user chooses which skills to forge, skip, or update. (Navigation questions like the Step 0.25 guided start, or scope confirmations, are exempt - they decide nothing about candidates.) Before any candidate prompt you must emit one chat message that contains a full dossier block (### <name> …, per Step 3) for EVERY candidate, and that message must END with this exact line:

=== dossiers above: N ===

where N equals the number of candidates you are about to offer. At the moment of asking, the check is mechanical: "Does a prior message of mine end with === dossiers above: N === and does N match my option count?" No sentinel → you have not rendered dossiers, whatever you remember - STOP and write them. Process narration between tool calls does NOT count; interim notes do NOT count.

The strongest form of LAW 1 is plan-style curation (Claude Code, see Step 3): present the dossiers AS the plan via ExitPlanMode - then showing the evidence and asking for the decision are the same act, and skipping the display is structurally impossible. But the plan only enforces that something is shown, not what (failure mode #3): the plan body must embed the engine-rendered dossier cards verbatim and end with the sentinel, same mechanical check as chat. The sentinel path alone is the fallback for harnesses without plan mode.

LAW 2 - RENDER THE ENGINE'S VISUALS VERBATIM, IN A REAL MESSAGE. After the report, you must emit a user-facing mining summary MESSAGE (tool output alone does not count - the user should not need to expand collapsed tool results). It opens with the engine's 📜 lore · … badge line and ends with the <!-- PASS-THROUGH FOOTER --> block, both verbatim. The same rule covers every PASS-THROUGH block the engine emits (STATUS, THEMES, DOSSIERS). Going tool → tool → question with no synthesis message in between is failure mode #2.

LAW 3 - NARRATE PHASES. Before every long-running engine or deep-mine call, emit one short status line so the window always shows what is happening: 📜 indexing BearClaude (253 sessions)…, 📜 deep-mining 4 clusters (this runs subagents; a few minutes)…, 📜 checking the forged-skill registry…. Never leave the user staring at a silent tool call.

What makes a candidate skill-worthy

A reproducible skill is a behavior that recurs, is regular, and has a clear trigger. The engine scores for recurrence/span/recency/specificity/outcomes; you apply the judgment it cannot:

• Keep it when the procedure is distinctive and specific to how this user/project works (e.g. supabase link → supabase db → supabase migration, gh run watch CI-watching, "write a comprehensive commit", "fix git divergence against origin/main", a read-only diagnosis).
• Discard it when it is generic to all coding and carries no project-specific procedure (e.g. bare git status → git diff, a lone "yes"/"do 1,2,3" confirmation). High session counts alone do not make a skill - ubiquity is not a trigger.

In cross-project mode the engine splits candidates into PORTABLE (recurs across ≥2 projects) and PROJECT-SPECIFIC (one project). Portability is the strongest signal of a real transferable skill: forge PORTABLE ones to the personal canonical dir and PROJECT-SPECIFIC ones into that repo.

Authorship (shared repos): committed histories carry their session owner - the engine attributes every session (git add-author > home-dir sniff > machine user) and candidates show 👥 N authors when several people exhibit the behavior. Use it:

• Multi-author candidate = a TEAM practice, the strongest forge signal of all - propose it at project scope (committed .claude/skills) so the whole team benefits.
• Single-author, and it's the current user = personal candidate, personal scope.
• Single-author, a TEAMMATE's = say so plainly in the dossier ("mined from Jake's sessions") and recommend team scope or checking with them before forging it as the user's own practice. Never present a teammate's workflow as the user's.
• Privacy: teammate names may appear in team-scoped (committed) skills; scrub them from personal-scope skills.

Note on the evidence: every command candidate comes from an actually-executed shell <tool-use> block (detected by the provider-set data-tool-type="shell" attribute, so it works for Bash, Shell, run_shell_command, exec_command, and every other provider's runner). It is real agent activity, not a pasted example. Single-line commands (inline backtick or in the tool <summary>) and multi-line

## Process

### Step 0 - Locate the history directory(ies)

Default to `.specstory/history` in the current project - but **check for nested histories first**
(monorepos keep them in sub-packages too):

```zsh
find . -type d -path '*/.specstory/history' -not -path '*/node_modules/*' 2>/dev/null | head

If more than one shows up, use --scan . (any-depth discovery, includes the root's own history). For cross-project trends across sibling repos, pass several --dir flags, one --projects <parent>, or --scan <parent>. If no history exists anywhere, tell them SpecStory records sessions and stop.

Step 0.25 - Guided start (when invoked with NO arguments)

A bare /lore means the user wants to be walked through it. Ask ONE structured question round (AskUserQuestion with three questions; plain numbered lists on harnesses without it), then proceed - do not make them learn the argument grammar:

1. Scope (header "Scope"): "This project (Recommended)" → cwd history, auto---scan . if nested histories exist · "All my repos under a folder" → ask which parent, then --scan <parent> · "Just the existing corpus" → skip indexing, report on ~/.specstory/lore.db directly.
2. Window (header "Window"): "All time (Recommended)" · "Last 30 days" → --days 30 · "Last 90 days" → --days 90.
3. Goal (header "Goal"): "Find & forge skills (Recommended)" → full pipeline · "Just show me candidates" → stop after dossiers, no forging · "Status / what has Lore done" → run status and render it verbatim (LAW 2), nothing else · "Reset my lore" → confirm, then reset.

This is a navigation question, not a candidate decision - LAW 1 does not apply to it. After the answers, echo the resolved interpretation in one line (per Step 0.5) and run. If the user typed ANY arguments, skip this step entirely and interpret them via Step 0.5.

Step 0.5 - Interpret the user's input

Map what the user typed to engine flags / process modes. If they gave nothing, Step 0.25 already collected the choices.

Echo back the resolved interpretation in one line before running (e.g. "Mining ~/code cross-project, last 60 days, command runbooks about supabase").

Step 1 - Index, then report (two engine commands)

The engine keeps the user's lore at ~/.specstory/lore.db (override with --db). It segments every session into beats (intent → agent method → outcome, where the outcome label comes from the user's NEXT reply: approval = success, steering correction = corrected). Indexing is incremental - unchanged sessions are skipped, so re-running is cheap. Transcripts from ALL agents accumulate into the same lore; each session is tagged with its agent (claude-code, codex-cli, cursor-cli, ...).

# 1. index (repeat --dir per project, or --projects <parent> to scan many repos)
node "<skill-dir>/scripts/mine-skills.mjs" index --dir <history-dir>
node "<skill-dir>/scripts/mine-skills.mjs" index --projects <parent-dir-of-repos>

# 2. report candidates (filters: --days, --min-sessions, --top, --kind cmd,task,meta,corr, --filter <substr>)
node "<skill-dir>/scripts/mine-skills.mjs" report --min-sessions 3 --top 10

(Legacy one-shot --dir without a subcommand does index + report together.) The report is wrapped in <!-- EVIDENCE FOR SYNTHESIS --> markers - raw evidence for you, not the user (--emit=json for structured output). It has four sections; read CORROBORATED first:

• CORROBORATED - intent × procedure pairs co-occurring in the same beats, with outcome rates. These are pre-verified deep-skill seeds: the user asked for X and the agent did Y, repeatedly.
• RUNBOOKS - executed command procedures (single-channel).
• INTENTS - recurring prompt task-types (single-channel).
• META-SKILLS - ways-of-working detectors.

Each evidence line carries path:line [outcome] intent=… cmds=… - an beat you can open directly.

Re-running is safe and expected. Indexing is idempotent: unchanged sessions are skipped (fingerprint = size + mtime + parser version); new sessions are appended; grown/edited sessions are replaced whole; engine upgrades re-parse the whole corpus automatically (one-time). --force re-indexes everything; prune drops sessions whose transcript files no longer exist and flags duplicate project identities (e.g. a repo that later gained a git remote and thus a new git_id). Run prune if the user has deleted or reorganized histories.

Step 2 - Synthesize candidate skills

From the evidence block, produce a shortlist of real skill proposals. For each kept candidate:

• name - a kebab verb-phrase (e.g. verify-go-changes, comprehensive-commit, fix-git-divergence).
• description (the trigger) - a one-line "Use when…" matched to how the behavior actually shows up in the evidence quotes. This is the most important field; it is what makes the skill discoverable.
• procedure - the steps, taken from the user's real command sequence or task shape. Do not invent steps; ground them in the evidence. Open the cited path:line refs (read the file at that span) if you need to confirm the exact commands before writing them into a skill.
• kind - runbook (a command procedure / task type) or meta-skill (a way of working).
• scope - personal (canonical ~/.agents/skills, fanned out to all harnesses) or project (<repo>/.claude/skills or the repo's equivalent, committed for the team).

Discard generic and weak candidates explicitly; tell the user what you dropped and why.

Skill-level idempotency - consult the registry first. Lore remembers what it has already forged and what the user has declined. Before proposing anything, run:

node "<skill-dir>/scripts/mine-skills.mjs" forged check --emit json

and obey each row's recommendation:

• up-to-date - the forged skill's cluster is unchanged: exclude it from candidates entirely.
• update: N new corrected beat(s), sessions A→B - the evidence grew materially since forging: propose an update to the existing skill (deep-mine the cluster, diff the new failure modes / steps into the installed SKILL.md), never a duplicate.
• update-carefully (user hand-edited the file) - same, but present the diff and let the user apply; do not overwrite their edits.
• suppress: user declined... - do NOT re-propose; mention it only in the discard list ("declined previously, evidence unchanged").
• re-engage: evidence grew materially since the user declined - you MAY re-propose, saying exactly what changed since they said no.
• orphaned - the skill file was deleted; offer to re-forge or forget it.

Also ls ~/.agents/skills/ for skills NOT authored by Lore (no registry row) - match those by name and skip duplicates. Re-running /lore today, tomorrow, or next month must never produce duplicate skills; it should produce updates as the lore grows.

Step 2b - Verify each candidate against the source (the truth check)

The engine finds recurrence of surface forms - it does not understand meaning, so a high count can be a coincidence (unrelated commands that happen to sit adjacent) or a parsing artifact. Before forging, confirm each shortlisted candidate is a TRUE pattern using corroboration, not just its score:

1. Start from CORROBORATED. The engine already computed the strongest truth signal - intent × procedure co-occurring in the same beats, with outcome rates. A corroborated pair with a healthy success rate needs only a light read; a single-channel runbook or intent needs more scrutiny.
2. Re-open the evidence. Each evidence line is an beat (path:line [outcome] intent cmds). Open 1–2 spans (grep/sed the file at that line range; never the whole file) and confirm the arc is coherent: the commands serve that intent, the outcome label is plausible, it is one procedure rather than accidental neighbors.
3. Weigh outcomes honestly. Outcome labels are conservative - most beats are neutral because the next prompt is a new task. Treat ✗ corrected as a strong negative signal; treat a few ✓ as suggestive, not proof (small denominators).
4. Portability & distinctiveness (already scored): recurring across projects, or built from project-specific tooling, beats universal-command ubiquity.
5. Refute the cheap explanation. Would a skeptic say this is tool noise, one busy afternoon, or the agent flailing? If you cannot answer with evidence, drop it.

Only candidates that survive this check proceed to curation. This is the same discipline the 25-patterns extraction used (adversarial verification against real transcripts), applied to a handful of finalists so you never read whole transcripts - only the evidence behind the candidates that already cleared the bar.