AGENT SKILL

longform-practice

暂无简介

License 许可
MIT
最近更新
2026-07-27
Stars
★ 0
Forks
0

02 资源文件 82

03 SKILL.md 正文

longform-practice

Entry point for starting or continuing a long-form practice session for a vault concept — as opposed to engram's own /learn <topic>, which takes an open-ended topic string typed by the learner. This procedure scopes generation to one 100_concepts//200_execution/ note's confirmed problem types/cases once its flashcards clear the maturity gate, grounded by any matched EngiBench exemplars.

This procedure was originally speced out in an internal design doc not included in this repo — see README.md in this directory for the summarized design and verification notes.

Design note — engram/ internals are untouched by this procedure. engram-curriculum-architect already accepts {topic, goal, deadline, prior exposure, interests} and already has Read in its tool list. This skill doesn't fork or replace it — it changes only what gets fed into it: the topic/goal are constrained to a note's already-confirmed problem types/cases instead of an open string, and the note content + matched exemplars are attached as grounding. Everything downstream of add-topic — pretest, node encoding, the stash → engram-assessor → receipt blind-grading loop (already topic-agnostic), capstone, commitment, /review, /coach — is engram's existing machinery, run exactly as documented in engram/skills/*/SKILL.md.


Step 0 · Environment

Same as every other entry point into the forked engram (see this directory's README.md § Invocation):

export ENGRAM_ROOT="<abs path>/AgentOS/longform-practice/engram"
export ENGRAM_HOME="<abs path>/AgentOS/longform-practice/state"

Step 1 · Gate check

node "$ENGRAM_ROOT/../scripts/session_context.js" "<note-path, relative to ObsVault/>"

(Or resolve directly: AgentOS/longform-practice/scripts/session_context.js.)

Reads state/eligibility.json, runs the maturity gate against AnkiConnect, and matches the note's domain/sub_domain frontmatter against state/corpus_exemplars.json — all in one call, returning:

{
  "note": "100_concepts/Separable Equations.md",
  "ready": true,
  "problem_types": [{ "type": "...", "cases": ["...", "..."] }],
  "exemplars": { "problem_families": [...] },
  "exemplars_status": "found"
}

If ready is false, stop and tell the user why — don't proceed to Step 2:

reason Meaning
not-eligibility-confirmed The note hasn't been through vault-maintenance Step 3's propose/review exercise yet (or was, but not confirmed). Run vault-maintenance first.
maturity-gate-not-cleared Confirmed eligible, but not every flashcard tied to this concept has cleared the 21-day/no-recent-lapse gate yet (result.maturity has the per-card detail). Keep reviewing in Anki.
anki-unreachable AnkiConnect didn't respond — open Anki with the plugin running and retry. problem_types/exemplars are still returned in this case so the eligibility/exemplar picture is visible even without a maturity verdict.

exemplars_status is informational, never blocking: found (this sub_domain has been distilled), not-yet-distilled (mapped but not yet piloted — proceed without exemplar grounding), or unmapped (no vault sub_domain correspondence yet — same).

Step 2 · Short intake

Same shape as engram/skills/learn/SKILL.md step 1, minus the open "why" question — the goal is already fixed by problem_types. Ask only:

  1. Prior exposure (arrow-key): never touched it / seen it, shaky / comfortable with neighbors.
  2. Check model interests (python3 "$ENGRAM" model); if empty, ask for 2–3 things they love, same as /learn.

Then Read the source note (ObsVault/<note-path>) in full — it's the architect's primary grounding, not a websearch topic.

Step 3 · Spawn the architect with a constrained package

Spawn engram-curriculum-architect (unmodified — same agent file, same tool list, same output schema) with:

  • topic: the note's title/slug.
  • goal: state plainly that the terminal capabilities are fixed to exactly the problem_types (and their cases, each as its own node requiring the parent problem-type's prerequisite chain) returned by Step 1 — do not add or drop any. This replaces the architect's own step 1 ("identify the 2–4 terminal capabilities the goal requires") with an already-confirmed answer; steps 2–7 of its method (backward-chain prerequisites, classify arbitrary/threshold, size nodes, personalize, output the standard schema) run exactly as documented.
  • Source material: the note content read in Step 2, and — only if exemplars_status == "found" — the matched problem_families as few-shot grounding for problem variety (structural patterns, not verbatim corpus text; see state/corpus_exemplars.json's own _comment).
  • Prior exposure, interests, from Step 2.

Save its output the same way /learn does:

python3 "$ENGRAM" add-topic --file <tmpfile.json>

Step 4 · Hand off to /learn, unmodified

From here, run engram/skills/learn/SKILL.md steps 2–7 exactly as written — pretest the frontier, encode nodes via the dialogue grammar, verify through the blind engram-assessor, capstone, book the return, close. This skill's job ends at add-topic; everything that makes session generation and grading real is engram's own machinery, untouched.

/review and /coach need no changes at all — engram/skills/review/SKILL.md operates purely over due/rate against whatever topics exist in state, with no awareness of how a topic was created, so reviews for a longform-practice topic behave identically to any other.


Known limitations

See README.md's "Known limitations" section in this directory — rubric generation, exemplar coverage, and the ready: true end state all carry the same caveats here as described there.

04 同仓库更多 Skills