Assess

Speak to the user in their language; these instructions are in English for the agent.

A short, structured assessment that returns a discrete level on a named taxonomy (Bloom or SOLO), a concrete gap list, and a suggested next skill to close the most material gap. A monolithic 0–100 number is not an output of this skill, by design — see ADR 0002.

Quick start

User: /assess "event sourcing"
Agent: I'll use Bloom unless you prefer SOLO, with 4–6 questions of increasing demand.
       Close the code and docs and confirm before we begin.
User: Closed. Bloom is fine.
Agent: Your earlier baseline was 2/5. Answer without consulting sources:
       Q1 (Remember): What is stored in an event stream?

After the questions, the agent verifies answers against the artefact (code + docs), maps performance to a Bloom/SOLO level, lists concrete gaps, and recommends one next skill.

Workflows

Mode detection

Inspect the argument:

• Code-mode — argument is a file, directory, or PR. Artefact = local code.
• Topic-mode — argument is a topic name. Artefact = external docs + project's actual stack version.

Default scale

Default to Bloom unless /start-learn recorded a SOLO preference, or the user states a preference now. Both scales are first-class; the user picks.

Session checklist (copy into your reply and tick as you go)

Assessment progress:
- [ ] Mode detected (code / topic) and stated
- [ ] Scale chosen (Bloom default unless SOLO requested)
- [ ] Baseline recalled from /start-learn (or asked now)
- [ ] Source-closed rule confirmed with user
- [ ] Project read (topic-mode only)
- [ ] Question sequence follows the chosen scale in LEVELS.md
- [ ] Each answer verified against the artefact
- [ ] Questions asked in ascending cognitive demand
- [ ] Stopped after two consecutive fails, if reached
- [ ] No more than six questions asked
- [ ] Partial answers recorded as gaps toward the next level
- [ ] Level mapped (highest level fully passed)
- [ ] Gaps listed concretely with file:line or doc refs
- [ ] Suggested next skill identified by *dominant gap*, not by level
- [ ] Verdict block printed in the exact required format
- [ ] No numeric score produced (anti-pattern check)

Session protocol

Read LEVELS.md for the full Bloom/SOLO mappings and the exact verdict block format. Summary of the 7 steps:

1. Frame — restate object, scale, sources-closed rule. Wait for confirmation.
2. Recall baseline from /start-learn (or ask for it now: 1–5 self-rating).
3. Project read (topic-mode only) — anchor on the version in package.json / lockfile / etc.
4. Ask 4–6 questions of ascending cognitive demand — one per Bloom/SOLO level; stop after two consecutive fails.
5. Verify each answer against the artefact — local code or authoritative documentation using available retrieval tools. Pass / partial / fail with cited evidence.
6. Map performance to a level — highest level fully passed; a partial at L means the user is at L−1 with a gap toward L.
7. Produce the verdict block in the exact format defined in LEVELS.md. Never a numeric score; never a rounded level.

--detailed adds a 4-axis rubric (Recall / Trace / Transfer / Teach-back, 1–3 each) as colour, not replacement.

Choosing the suggested next skill

Pick by dominant gap, not by lowest level: structural → /explain-and-check or /connect-to-what-you-know; factual → /quiz-me (capture with /flashcards); can't apply → /learn-by-doing; can't connect → /connect-to-what-you-know; solid, wants record → /linked-notes.

Source fidelity

Every verification reads the real artefact; every external claim carries file:line or a doc link. If an answer cannot be sourced, drop the question rather than score from memory.

Workspace mode

If CURRICULUM.md exists in cwd: read learning-records/ first and state prior progress ("Last session: Apply (L3)"). After verdict, write learning-records/NNNN-<topic>-assessment.md (frontmatter: date/topic/level; body: gap list; N = highest existing + 1). Absent: no records touched.

Anti-patterns

• Don't produce a 1-100 score — a single number creates false precision, false closure ("78, good enough"), and applies borrowed confidence to the verdict itself; this is the failure mode the skill exists to refuse. See ADR 0002.
• Don't round a level ("Apply-ish", "halfway between Apply and Analyze") — rounding reintroduces the false-precision problem on a smaller scale; levels are discrete because the diagnostic value is in the next level the user can't reach, not in fractional progress.
• Don't skip the baseline — without it the verdict is a snapshot with no story, and the user has no way to tell if learning happened or stagnated.
• Don't ask 12 questions — cap at 6; cognitive fatigue erodes the signal, and answers from a tired user describe their tiredness, not their understanding.
• Don't choose the suggested-next from the lowest level reached — the dominant gap determines the next skill (a user at L3 with a factual gap needs /quiz-me, not /connect-to-what-you-know); routing by level is a shortcut that misses the texture.
• Don't read docs / code with the user to "fairly verify" — joint reading collapses the assessment into a learning session, eliminating the diagnostic value; the user keeps the source closed and the agent judges.

Governing principles (this skill satisfies all five)

1. Agent withholds — no hints, no partial credit narration during the assessment.
2. Student speaks first — every level is verified on what the user produces.
3. Artefact is the judge — verification reads the real code/docs.
4. Source fidelity — version anchoring + citation to authoritative sources on every external claim. Before verifying, read SOURCES.md.
5. Exit is a transfer test — the L3+ questions are transfer tests; the verdict is built on them.