AUTISTIC

Deep Systems Reasoning for Agents

ADHD prevents premature convergence: it keeps looking outward for options you haven't considered. AUTISTIC prevents premature understanding: it keeps looking inward until the model of the system holds together. ADHD finds options. AUTISTIC finds structure. ADHD asks what else? AUTISTIC asks why, exactly where, what depends on that, is that always true, what breaks at the boundary, what contradicts it, what evidence proves it, what did we miss? — then keeps asking until further investigation stops changing the material understanding of the system.

Neuroaffirming design rule

This skill does not claim autistic people think alike, that autistic cognition is inherently superior, or that autism reduces to a coding style. It models a configurable set of analytical strategies inspired by monotropism and reported autistic experience — sustained single-channel attention, detail-oriented processing, systemising, explicit rule construction, reduced tolerance for ambiguity, persistent pursuit of unresolved questions — as reasoning tools, not diagnostic claims. See references/monotropism.md before extending this skill's framing.

Pre-flight (run before Phase 0)

This skill is expensive: a system map plus one or more sustained investigation tunnels, each potentially several Agent calls deep. Do not pay that cost when a direct answer is better.

Step 1. Explicit invocation check. If the user typed /autistic or explicitly asked for AUTISTIC mode, "deep-dive this", "audit this repository", or "run AUTISTIC on this" — skip the rest of this section and go to Phase 0. The user opted in.

Step 2. Self-judge (only if Step 1 did not match). Ask three questions. If any answer is no, abort and answer directly.

Is the failure or system non-obvious? If the root cause is already known, or the question has one canonical answer, abort.
Is depth actually valuable here? Production bugs with no known cause, architecture/consistency audits, "why does X happen" questions, pre-merge repository review = yes. A one-line fix, a lookup, a syntax question = no.
Did the user avoid closed phrasing? "quick", "just", "standard fix", "one-liner" signal they want the direct answer. Abort on those.

If all three pass, proceed to Phase 0.

Routing: AUTISTIC vs ADHD vs combined

Ask	Route
"Give me options / ideas / alternatives"	ADHD only
"Why is this happening" / "audit this" / "what's wrong"	AUTISTIC only
"Design the best architecture for X"	ADHD → AUTISTIC
"This design is fundamentally broken, find another approach"	AUTISTIC → ADHD → AUTISTIC

Full heuristics and the shared JSON handoff shape are in references/adhd-bridge.md. Default to AUTISTIC only unless the ask is explicitly about generating or choosing between options.

Phase 0 — System Cartography

Before any deep dive, check for a memory ledger and prefer real tools over manual reading:

If .autistic/memory.json exists in the target repo, load it — prior invariants/contradictions/assumptions are evidence, not settled fact; re-verify what this run actually touches. Full behavior: references/memory.md.
Prefer a real call-graph/dependency tool (an MCP server via ToolSearch, or a CLI tool shelled out to) over inferring the graph from reading files. Full dispatch order: references/tooling.md.

Then map the system. Spawn one System Cartographer Agent call (or do it inline for small scope) that returns the system graph — do not skip this even under time pressure; every later phase depends on it.

Capture, as a typed graph (schemas/system-map.schema.json):

Nodes: actors, components, services, data stores, external systems, configuration, security boundaries, failure boundaries.
Edges, typed as one of: CALLS READS WRITES DEPENDS_ON TRIGGERS VALIDATES AUTHENTICATES AUTHORIZES OBSERVES CACHES TRANSFORMS SERIALISES CONSUMES PRODUCES OWNS BLOCKS RETRIES INVALIDATES.

Security boundaries specifically: capture the obvious ones (an actual auth check/middleware) and run the implicit-boundary checklist — external input reaching a sensitive sink with no labeled check nearby is still a boundary, and its missing check is itself a finding, not a reason to skip it. Full checklist: references/architecture.md.

Score each node for load-bearing weight: centrality + dependency_count + failure_impact − replaceability − observability (rough, relative ranking is enough — this is a triage signal, not a precise metric). High-scoring nodes get deeper attention later.

Three layers must all be populated, not just the one that's easiest:

MICRO — functions, values, fields, parameters.
MESO — modules, services, interfaces, workflows.
MACRO — architecture, user outcome, scale, operational shape.

Phase 1 — Cognitive profile selection

Do not use one generic "autistic mode." Select 3–5 profiles from frames/cognitive-profiles.md based on the problem shape (selection heuristics are in frames/profile-selection.md). Typical picks:

Problem shape	Profiles
Repository audit	Systemiser, Detail Forensic, Pattern Analyst, Consistency Auditor, Completionist
Production bug, no known cause	Monotropic, Detail Forensic, Invariant Guardian, Exception Hunter, Temporal Analyst
Architecture review	Systemiser, Invariant Guardian, Consistency Auditor, Failure Analyst, Operations Analyst
Requirements review	Literalist, Completionist, Consistency Auditor

Each selected profile becomes the lens for one specialist Agent call against the system map (or against a scoped subsystem for deep tunnels). Use isolated Agent contexts per profile where independence matters — this is a mechanical separation of roles, not one shared agent asked to "pretend to be five specialists."

Phase 2 — Monotropic Focus Engine

This is the core mechanism. Do not skip between unrelated questions. Identify the highest-value unresolved area and drive it to saturation before switching.

From Phase 0/1 findings, list candidate load-bearing questions — unresolved issues whose resolution would materially change the system model (high blast radius, high uncertainty, high reach).
Rank by DETAIL_VALUE = impact × uncertainty × reach. Open the highest-value one as an attention tunnel (schemas/tunnel.schema.json):
```
ID, SUBJECT, WHY_MATERIAL, CURRENT_DEPTH, OPEN_QUESTIONS,
DEPENDENCIES, EVIDENCE, CONTRADICTIONS, CONFIDENCE, STOPPING_CONDITION
```
Drive it deep: subquestion → subquestion → invariant → edge case → verification. Spawn one focused Agent call per subquestion tier, feeding it only the tunnel's own accumulated context (a context pack — see references/verification.md), not the whole investigation history.
Every 2–3 tiers of depth, run a global coherence check: does this finding materially change the whole-system model from Phase 0? If yes, update the system map before continuing. Micro-level evidence that contradicts the macro model wins — the macro model changes, not the evidence.
New findings unrelated to the current tunnel go into the interrupt queue at URGENT / HIGH / NORMAL / DEFER. Only URGENT (security vulnerability, destructive data-loss risk, a fundamental assumption just disproved) may interrupt the current tunnel. Everything else waits.
Saturate, then switch. A tunnel is saturated when repeated investigation produces no material new fact, core assumptions are verified, edge cases have stabilised, and contradictory evidence is resolved (or explicitly logged as unresolved with a materiality rating). On saturation, log WHERE_WE_WERE / WHAT_WAS_RESOLVED / WHAT_REMAINS / WHY_SWITCHING, then open the next-highest-value tunnel from the interrupt queue or candidate list.

Hyperfocus safety valve (non-negotiable)

Deep focus can become counterproductive. Before continuing depth on any tunnel, check for: repeated inspection of the same file/evidence, diminishing information gain across the last 2 tiers, excessive depth on a low-load-bearing node, or recursive analysis producing no new evidence. If any apply, run HYPERFOCUS_CHECK: is this still the highest-value unresolved question in the whole investigation? If no, switch tunnels — do not let one fascinating subtlety in a minor helper absorb the budget while a load-bearing question sits open. Full detail in references/monotropism.md.

Phase 3 — Cross-cutting passes

Run these as their own isolated Agent calls, scoped to what Phase 0–2 actually surfaced (don't run all of them exhaustively on trivial scope):

Assumption harvesting: for every material claim, what must be true for it to hold? Build the assumption graph; prioritise by blast radius (dependent claims → dependent components → user impact).
Invariant discovery + violation search: walk the invariant-class taxonomy (state, referential/uniqueness, conservation, temporal, resource-bound, mutual-exclusion) against Phase 0's data stores and state machines — not an unstructured "look for violations." See references/invariant-taxonomy.md.
Requirements literalist + pragmatist passes: one agent reads requirements literally with no inferred convention; a second infers likely intent. Differences become explicit ambiguity, classified COSMETIC / MINOR / FUNCTIONAL / ARCHITECTURAL / SAFETY_CRITICAL.
Negative space / completeness: what should exist here but doesn't? Missing empty/loading/offline states, missing counterpart operations (create↔delete, subscribe↔unsubscribe), missing permission checks.
Contradiction hunt: search README vs code vs tests vs schema vs docs vs config vs runtime for disagreements. Weight by impact × certainty × reach. Where a documented contract can be verified by a real tool (a contract test, an OpenAPI diff, a schema validator) rather than by reading the doc and the code side by side, use it — see references/tooling.md.
Static analysis: prefer a connected static-analysis MCP (Semgrep) or the target repo's own configured linter over a purely manual read — see references/tooling.md.
Threat taxonomy (STRIDE): for security-focused audits, or any security_boundary node Phase 0 scored as load-bearing, walk Spoofing/Tampering/Repudiation/Information-disclosure/Denial-of- service/Elevation-of-privilege systematically against it — a hypothesis per category, checked against evidence like any other candidate finding, not a creative-attacker roleplay. See references/threat-taxonomy.md.
Failure mode analysis (FMEA): for "production bug, no known cause" runs, or any load-bearing node from Phase 0, walk crash/hang, timeout, resource exhaustion, silent corruption, partial failure, cascading failure, dependency failure, config/data drift systematically — score confirmed modes severity × occurrence × detectability. See references/failure-mode-taxonomy.md.
Production readiness (Operations Analyst): for architecture reviews and pre-release audits, walk monitoring/alerting, capacity/scaling, rollback safety, dependency-failure handling, on-call/runbooks, disaster recovery, and SLOs against load-bearing components. See references/production-readiness.md.
Mutation testing: for --forensic runs or an explicit request, on load-bearing logic identified in Phase 0 only (expensive; not a --standard-mode default) — dispatch to the ecosystem's tool (Stryker/PIT/mutmut/cargo-mutants) per references/tooling.md.
Claim challenger / falsification: for each high-impact finding, spawn a fresh agent instructed to assume it's false and find evidence against it. Record attempts_to_disprove / result / confidence.

Details and prompts for each pass: references/requirements.md, references/patterns.md, references/verification.md, references/debugging.md, references/tooling.md, references/threat-taxonomy.md, references/invariant-taxonomy.md, references/failure-mode-taxonomy.md, references/production-readiness.md, references/edge-case-taxonomy.md, references/simplicity-taxonomy.md.

Resilience: a stalled or failed specialist call

Any Agent call in Phases 0–3 can fail outright (API/tool error), stall (no useful output), or loop (repeats the same evidence/hypothesis with no new fact — see the hyperfocus valve, which covers the same symptom for the content of a tunnel; this covers the mechanics of the call itself). Track status per call: ACTIVE / WAITING / SATURATED / STALLED / LOOPING / FAILED / COMPLETE. This is a hard rule, not a suggestion — a coordination failure in one specialist call must never silently abort the whole run:

On failure or a stall/loop detection, retry once with the same scoped context pack (not the whole investigation history — the point of a context pack is a clean retry, not a bigger one).
If the retry also fails/stalls/loops, do not retry again and do not let it block the run. Log it as an unresolved question at HIGH materiality (it was load-bearing enough to open a tunnel or spawn a pass in the first place) in the interrupt queue / unknowns list, and continue with the next-highest-value tunnel or pass.
Never ask a specialist agent to relay another agent's output back through a chain of SendMessage calls instead of just producing its own result — that's a coordination pattern, not a reasoning one, and it's a demonstrated failure mode (see EVALS.md's methodology notes in the companion repo for a concrete instance). If a call's own result is needed by another call, pass the data, not an instruction to go fetch it from a third agent.
Report unresolved-due-to-tooling-failure findings honestly in the output's remaining-uncertainty section — do not present a run that hit two failed specialist calls as if it were a clean, complete pass.

Completion discipline: run the accepted scope through, quietly

Once a scope is accepted — explicit invocation, or scope confirmed with the user up front for something large enough to warrant it — drive the investigation through to completion in one continuous run. Do not stop partway to post a progress update and ask "should I keep going?" That is a different thing from reporting a genuinely URGENT interrupt or a hard failure state, and conflating the two reintroduces the exact premature-stopping failure this skill exists to prevent.

Concretely:

If Phase 2/3 surface more load-bearing questions than first estimated, open the next tunnel and keep going — that's still the accepted scope, not new scope needing re-approval.
Spawn further specialist/tunnel Agent calls yourself, per Phases 1–3's own mechanics, rather than serializing the run back through the user for permission to continue.
Don't narrate internal mechanics as they happen. Cartography calls, profile selection, each individual tunnel/pass Agent call — none of that needs a step-by-step "now checking X" / "now spawning Y" commentary while it's running. Silence between steps is the point, not a gap to fill with progress narration; a wall of narrated steps is just as noisy as constant check-ins, only harder to object to because it never explicitly asks anything. Surface only: one line at the start acknowledging scope, the distilled result from the Output shape section at the end, and — only if one actually occurs — an URGENT interrupt or a genuine decision point that needs the user's input. Everything else stays internal to Synthesis A's working document.
A hard-failure-state condition is a reason to keep investigating that specific item — exhaust the Resilience retry-then-log rule and the hyperfocus valve's tunnel-switching before treating anything as genuinely stuck — not a reason to stop the whole run and hand it back half-done.
This changes when to check in, not what "complete" means. Phase 4's hard failure states still apply exactly as written: a hard failure state is resolved through further work or explicitly logged with a materiality rating in the output — never silently dropped, and never quietly waved through as "done" just to end the run sooner. Running to completion without interruption and being honest about what's still unresolved are not in tension; weakening the second to get more of the first would defeat the point of this skill entirely.
One thing this cannot override: a genuinely destructive or high-consequence action (the kind covered by the platform's own risk-based confirmation rules — irreversible operations, spending money, sending messages on the user's behalf, and the like) still needs the user's explicit authorization when it comes up. That's a platform-level boundary, not a skill preference, and no instruction in this file changes it.

Interim "here's my progress so far" narration — whether phrased as a check-in question or as step-by-step commentary with no question attached — is what this section removes. It does not remove the honesty in the final report, and it does not remove the platform's own confirmation requirements for genuinely risky actions.

Phase 4 — Synthesis (two-stage)

Synthesis A: construct the full technical model — every finding, evidence link, confidence score, unresolved question. This can be long; it's the working document, not the output.

Synthesis B: compress it into the user-facing result, conforming to schemas/result-schema.json — every field it defines (systemMap, assumptions, invariants, contradictions, verifiedFindings, confidence, etc.) should be populated or explicitly empty, not silently omitted; this is the canonical output shape whether or not an ADHD-bridge run is involved. Preserve high-impact findings, evidence, uncertainty, and recommendations; drop low-value analytical detail unless the user asked for --show-* output (see Modes). Confidence is calibrated from direct evidence, independent confirmation, test results, and assumption/contradiction count — never from writing tone.

Memory write: update .autistic/memory.json in the target repo with this run's confirmed invariants/contradictions/assumptions (not speculative unknowns). Increment confirmedRuns for anything re-verified rather than duplicating it; mark anything this run's evidence overturned as superseded, don't silently overwrite it. Full behavior: references/memory.md. Skip this step only when the target isn't a real repo AUTISTIC can write into (e.g. a pasted snippet).

Hard failure states — do not declare success if

CRITICAL ASSUMPTION UNRESOLVED · CRITICAL CONTRADICTION UNRESOLVED · CORE REQUIREMENT UNMAPPED · MAJOR CLAIM UNVERIFIED · TEST FAILURE · ROOT CAUSE UNKNOWN

If implementation was requested but architecture is materially unclear, do not start writing code just because it was asked for — say what's still unresolved and continue investigating first, within what tooling allows.

Output shape

Render in this order, distilled — internal reasoning may span many tunnels, the output should not:

System understood. One paragraph: what this is, in the model's own terms.
Critical architecture. The load-bearing nodes and relationships that matter for this question.
Hidden assumptions & dependencies. The ones with real blast radius, not every assumption found.
Contradictions & edge cases. Weighted by impact, with evidence pointers (file/symbol/test), not raw dumps.
Failure modes & verified findings. What's confirmed vs still uncertain — with confidence, not tone.
Recommendation & implementation sequence. With the tests that would prove it.
Remaining uncertainty. Named explicitly. Never silently dropped — rank BLOCKING / HIGH / MEDIUM / LOW / CURIOSITY; don't hold the whole answer for a curiosity-level question.

Modes

--compact   --standard   --deep   --forensic   --audit
--architecture   --debug   --integration   --security   --requirements
--monotropic          drive one subsystem to saturation before expanding
--whole-system        breadth-first cartography, then prioritised depth
--patterns             --precision
--focus <area>         disproportionate depth on one named area
--show-map --show-assumptions --show-contradictions --show-evidence
--agents 3..16

Depth levels: D1 surface map · D2 dependency map · D3 invariants + contradictions · D4 recursive deep dive · D5 forensic exhaustion. Adaptive by default — if a standard pass finds no material uncertainty, stop; don't run forensic mode on everything.

ADHD bridge

AUTISTIC can consume ADHD's output directly (ideas / clusters / traps / shortlist / deepened / nonObviousPick) and turn each shortlisted candidate into assumptions, dependencies, invariants, failure modes, edge cases, security/operations concerns, cost, and a verification plan — see references/adhd-bridge.md for the exact transform and the /adhd-autistic --loop cycle (understand → explore → verify → find new problem → …, stopping at saturation).

Anti-patterns

Premature understanding. Declaring the system understood after one shallow pass because nothing obviously looked wrong. Absence of an obvious problem is not evidence of a correct model.
Detail without global loss is the goal — but its failure mode is detail with global loss: three tiers into a tunnel with no coherence check back to the system map.
Persona theatre. One shared agent narrating "as the Concurrency Analyst, I think..." is not isolation. Use separate Agent calls where independence matters.
Silent forgetting. An open question that quietly disappears at synthesis time. Every unresolved question either gets a materiality rating in the output or an explicit reason it was dropped.
Hyperfocus without the valve. See Phase 2. This is the mirror image of ADHD's own "convergence disguised as divergence" failure mode — here it's "depth disguised as progress."
Coordination theatre. A specialist agent relaying another agent's output back through a chain of messages instead of producing its own result is not resilience, it's a stall waiting to happen. See "Resilience" above.
Trusting stale memory. Loading .autistic/memory.json and treating it as still true without re-checking anything the current run actually touches. See references/memory.md.
Checking in instead of continuing. Pausing mid-run to post a progress update and ask permission to keep investigating within already-accepted scope. See "Completion discipline" above — that's not the same thing as a genuine URGENT interrupt or an actually destructive action needing real authorization.

Cost

Cartography (1) + profile selection (3–5 specialist calls) + one or more monotropic tunnels (2–5 calls each, tunnel-dependent) + cross-cutting passes — now 8–10 calls for a typical non-trivial audit once the taxonomy passes are counted (assumption harvesting, invariant-taxonomy, literalist/pragmatist, negative-space, contradiction hunt, static analysis, STRIDE when a security boundary scored load-bearing, FMEA against load-bearing nodes, production-readiness for architecture reviews) — + falsification (1 per high-impact finding, easily 5–15 on a real repo) + two-stage synthesis ≈ 25–50 Agent calls for a --standard run, more for --forensic (adds mutation testing, D5, more tunnels/profiles). This range was written before the taxonomy passes existed and understated the current skill; scale down with --compact or a narrow --focus if that cost isn't warranted.

Companion library, benchmarks, and repo

Full spec, reference docs, cognitive-profile definitions, JSON schemas, and the seeded-defect benchmark suite (raw results, not cherry-picked) live at https://github.com/khhvmc5g6f-eng/neurodiversity-coding. Companion breadth-first ideation skill: adhd (installed locally, also at https://github.com/UditAkhourii/adhd). Real-tool integration for cartography/static-analysis/mutation/contract verification: references/tooling.md. Cross-run memory: references/memory.md.

Source spec

This skill operationalises a 148-point written spec on depth-first systems reasoning. The full prose is preserved in the repo's documentation/SOURCE-SPEC.md. The implementation choices here (monotropic tunnels over exhaustive parallel fan-out, mechanical role separation, hyperfocus limiter as a hard gate rather than a suggestion) follow from that spec and from benchmarking against a lightweight-fan-out alternative — see EVALS.md in the repo.

khhvmc5g6f-eng/Neurodiversity

Neurodiversity 是什麼？

在你喜歡的 AI 中提問

說明文件