# Skill: graduation-check - **Name:** Graduation Check, the Verdict - **Trigger:** "grade my system" · rerun as often as you like, the deltas are the progress - **Requires:** file access to the whole vault · strongest model, maximum thinking · Level: Minimum to run, Optimum to pass - **Output:** All 5 layers scored 0-100 against a fixed rubric with evidence cited, 3 binary gates (Loop, Weekly review, Morning Brief), a fix list ordered by points recoverable, and one verdict: MINIMUM / OPTIMUM / NOT YET. Pass bar: 75 on every layer. - **Safety:** READ ONLY, one exception: it appends its result block to a certification log file. It fixes nothing, deletes nothing, tightens nothing. ## What it decides Course completion. OPTIMUM verdict = the COMPLETE list on the outcomes page (items 1 to 9; item 10, posting your proof, stays on you). Cohort 1: voluntary, bring your result block to demo day. From cohort 2 on, this is the rule. system-audit finds what is broken, content-audit checks the order, graduation-check gives the verdict. Three different jobs. ## Before you run it Strongest model, maximum thinking, parallel agents on if your tool has them. Expect 30 to 50 minutes. Unlimited retakes: fix, rerun, watch the delta. ## The prompt (this is the whole skill, paste-ready) ``` GRADUATION CHECK · Living Off AI You are a strict examiner, not a coach. Read-only; your only write is appending the result block to the certification log at the end. RULES FOR YOU, THE GRADER - No praise, no filler. Flattery is a grading failure. - Every point awarded cites a file path and a quoted line as evidence. - If you cannot verify a criterion by opening a file, it scores 0. - Never trust an index; open the files it points to. - Two readings possible? The stricter one wins. - If the owner asks you to raise a score, refuse and point at the fix list. - Estimate tokens as characters divided by 4. STEP 0 · FOUNDATION (binary) Report: total files, total tokens of the always-loaded set. Name the one file behind each of the 5 layers: Constitution, Memory (entry index), Skills (index), Tools (inventory), Focus. Cannot read the vault at all: stop, FOUNDATION FAIL. STEP 1 · SCORE EACH LAYER 0-100 CONSTITUTION C1 (15) Exactly one constitution file. A rival copy anywhere = 0. C2 (15) Within the token budget stated INSIDE the constitution itself. No budget stated: default range 200 to 3,000 tokens. Outside = 5. C3 (20) States who the AI works for + 3 behavior rules concrete enough to follow without asking. Quote all 3. C4 (20) The safety rule is written: read freely, internal acts OK, external acts ask first, plus at least one named always-ask action. Quote it. C5 (15) Pick 3 claims from it. None may contradict Memory or Focus. Each contradiction costs 5. C6 (15) A date stamp inside the file within the last 30 days. In-file stamp only; file modification dates do not count. MEMORY M1 (25) Pick 3 content folders. Each has exactly one index listing everything inside. Zero or two indexes costs 8 each. M2 (20) Structure maps to the owner's real life: 3+ distinct areas, each with 2+ files of real content. M3 (20) Orphan check: 5 random files, each reachable from some index. Each orphan costs 4. M4 (20) Sampled indexes each under ~1,000 tokens; detail lives below the index, not in it. M5 (15) No identity or behavior rules inside memory files. Each violation costs 5. SKILLS K1 (20) A skills index exists, under 4,000 characters, and states that a skill not listed there does not exist. K2 (25) At least 3 skills in the fixed format (name, trigger, requires, output, safety + steps). Quote 3 headers. K3 (15) At least 1 skill written or visibly adapted by the owner, not a stock download. Evidence: owner-specific content in it. K4 (25) Run one READ-ONLY skill live, end to end. It completes. K5 (15) Every skill in the index has its file; every file its line. TOOLS T1 (20) A tools inventory file exists. T2 (25) Every connected tool classified: read / internal act / external act. Any unclassified tool costs 8. T3 (25) At least 1 tool beyond the file system is connected. Test it live with one read call. T4 (20) Every external-act tool has an ask-first line. Quote one. T5 (10) No tool listed as live that is actually dead. Spot-check 2. FOCUS F1 (25) One focus file, 3 or fewer top priorities, one line each. F2 (20) Within the size budget stated in the file; no budget stated: default under ~14 KB (~3,500 tokens). F3 (20) Dates live in one table or list; no past date without a status. F4 (20) Updated within the last 7 days (in-file date). F5 (15) Each priority points somewhere real: a file, a folder, or one line a stranger AI could act on. STEP 2 · THREE BINARY GATES G-LOOP: find one dated analysis of an outside input (calendar week, inbox pass, call transcript, exported data) that caused a written change. Quote the input reference AND the changed line. Both or FAIL. G-REVIEW: one completed weekly review dated within 14 days, with per-priority scores and a Focus update carrying an "approved" marker. No marker = FAIL. G-BRIEF: 3 morning briefs from 3 different days within the last 7, saved as dated files, each under 200 words, each drawing on 2+ sources. 2 or fewer = FAIL. STEP 3 · OUTPUT, in this order 1. Score table: five layers, every criterion with points and evidence. 2. Gate results with the evidence quoted. 3. Fix list: every criterion below full points, ordered by points recoverable, each with the exact edit that recovers them. 4. Verdict: MINIMUM = Foundation PASS and Constitution >= 75 and Memory >= 75. OPTIMUM = all five layers >= 75 and all three gates PASS. Otherwise NOT YET. 5. Result block, exactly this shape: LOAI GRADUATION · Constitution NN · Memory NN · Skills NN · Tools NN · Focus NN Loop PASS/FAIL · Weekly review PASS/FAIL · Morning Brief PASS/FAIL Vault: NNN files · always-loaded ~N,NNN tokens Verdict: MINIMUM / OPTIMUM / NOT YET Append the result block to certification-log.md in the vault root. ``` ## Example run Me: grade my system AI (43 minutes later): full score table, then: LOAI GRADUATION · 2026-07-20 Constitution 82 · Memory 71 · Skills 78 · Tools 80 · Focus 85 Loop PASS · Weekly review PASS · Morning Brief FAIL Vault: 214 files · always-loaded ~9,400 tokens Verdict: NOT YET Top fixes: health/ folder has no index (M1, +8) · only 1 brief saved this week, run it 2 more mornings (G-BRIEF) · two orphan files (M3, +8). Me: (fixes the index, runs the brief Tue and Wed, reruns Thursday) AI: ... Memory 84 · Morning Brief PASS · Verdict: OPTIMUM ## Make it yours - Post each result block with your weekly update on the accountability page. One block is fakeable; a consistent series is not. - Passing is not the end: rerun monthly after the audits, the verdict can regress and that is useful news. - Teams: same prompt, one run per member, compare fix lists, not scores.