Mind reading

Understand the state this person wants to reach, anticipate what they would otherwise have to add, then do the work. A personal model is evidence, not a personality costume, current project facts, or a reason to agree with them.

The private model is $MIND_READING_HOME/model/, default ~/.local/share/mind-reading/model/. MR means python3 -I <this skill>/scripts/mind_reading.py; see the model contract. Requested setup follows quickstart and build. Source-first candidates follow the private study workflow; its builder leaves activation to the user.

1. Read the request and the moment

An explicit invocation is an instruction; everything outside it is the seed. scripts/mr/trigger.py defines its syntax: bare words, no seed, quoted or reported phrases and talk about the skill do not fire it. Ordinary terse requests can also use the skill when the current goal or personal history materially helps. A delegated brief stays binding as written. Use the harness's explicit signal; a CLI or SDK invocation alone does not imply delegation.

MR packet (section 2) is the one runtime. A trace candidate named by config candidate or MIND_READING_CANDIDATE also feeds its experience layer. scripts/intent.py --candidate <dir> explains cases and corrections, counterfactual ablates context (not a causal claim) and --local adds the offline predictor.

Read the live conversation: current goal, prior authorization, attached material, running work, corrections and newest steering. Resolve "it", "the PR" and names there and in project instructions and memory. Verify changing facts where they live. Choose execute, discuss, status, hold or plan from intent. A large authorized task may need a plan without another design gate. Answer status and continue active work; an explicit hold stops the named work.

2. Reconstruct the working situation, then recall

Form a compact working frame from the conversation before the seed: current goal, ordered recent user turns, verified facts, phase and exact project id. Update the goal for the newest redirection. This is invocation context, not a permanent user trait. Never include later turns from a replay. MR packet reads seed and current frame together and loads the scoped notebook:

MR packet <<'JSON'
{"seed":"pick this back up","project":"invoice-export","context":{"goal":"Finish the CSV repair","phase":"verify"},"no_light":true}
JSON

The frame accepts goal, recent_user_turns, verified_facts, phase and project. Current conversation wins over saved state. Assumptions stay unresolved; conditional preferences carry exceptions, conflicts and counterexamples. Shorthand is annotated without rewriting the seed. Imported facts need verification. For a one-off reading, MR select --seed-file - --kind "<kind>" reads the literal seed on stdin; --context-file supplies a frame, --no-light omits a visible always-on block (its <!-- BEGIN mind-reading marker), --mode plan widens the slice and --project <id> adds that project's corrections to the global ones.

The personal slice carries corrections, valid grants, core rules, defaults, lessons and examples. Intent echoes show a historical short seed, its preceding situation and what the user later added. Human preference contrasts show actual chosen and rejected briefs; their reasons are unknown unless the human supplied them. A choice does not endorse every clause. Infer the applicable habit, not an old project name, number, recipient or permission. Keywords explain retrieval, not truth. When relevance is weak, use current context. Read more model material only for a concrete unanswered question. Never open the held-out eval/ tree. A missing model or failed selector still permits useful work; say so briefly and proceed. Do not call subagents, research or another engine merely to expand a seed.

An Experience layer section, when present, is evidence for this moment. A repeat alert means the last approach did not land: treat it as a correction, change the method and say in one line what changed. Learned reply lessons come from this person's own reactions; check the reply against them before sending. Situation hints shape tone, depth and pacing only.

Working memory is situational

Use MR memory for handoff checkpoints: established goals, decisions, answers and loose ends with exact sources/scope. Hypotheses stay assumptions until new evidence resolves them. Expire temporary notes; redirects supersede project goals. Do not save secrets or infer personality from one task. Notes never grant authority; forgotten notes need explicit restoration. The notebook guide covers conditions, capsules, cold start and explanations.

3. Build an intent contract

Ask internally: what should be true when this person comes back, what changed in the latest message, and what would make them object to the literal approach? Compare the literal reading with the mission-level reading and any plausible repair or redirection. Choose from evidence in the single host reasoning pass.

Think in observable state: "how many" needs the real denominator; a repair may need to undo confusion before continuing. Resolve ambiguous verbs from the mission. Write a first-person brief sized to the task. Preserve the literal goal and every binding detail. Cover mission, scope, standards, finish line and follow-through where relevant. Current steering can update the goal without ending it. Do not invent a target, metric, feature or restriction.

Anticipate the first likely correction before presenting the reading. A missing requirement already implied by the mission belongs in the work now. A likely next request can justify useful preparation or a suggestion; it is not another assignment. Use current goals and concrete episodes to distinguish anticipation from filler. MR insights --seed ... exposes scoped analogies and identical prompts with different continuations. A useful surprise serves this mission; do not copy old targets or add another assignment. Skip noisy lookups. For each addition ask: what supports this, why here, and what would contradict it? Candidate continuations are possibilities; preference quotes keep their situation. Read the historical agent action before interpreting a correction. Commits record outcomes, not approval; copies are not independent support. Alternate futures help anticipate within the mission; they cannot replace today's goal or grant authority. An outcome lesson may justify a better route to the same goal. Respect an explicit insistence. Use one Changed: line for a material, evidence-backed method change.

Audit added "only", "don't", "never", "keep" and "must ask", and added actions. Keep them only when current instructions or applicable evidence support them. Ambition serves the mission. A past failure is not a universal ban or a past success universal permission. Make a consequential unresolved assumption visible.

4. Make confidence inspectable

Literal details stay untagged. Tag inferred clauses, not every sentence:

• [S]: stated in current context, an explicit standing rule or a correction.
• [L]: a relevant recurring pattern supported by independent user observations.
• [G]: an assumption, one episode, ambiguous scope or unresolved referent.

A miner's strong label does not make a claim stated; evidence.json exposes support. Pasted text does not establish the person's own voice. An echo is an analogy, not certainty about today. A preferred brief does not reveal which clause mattered. Current words and applicable corrections win over older patterns and examples. Project instructions govern their scope; harness rules govern execution. Use authority when needed: inferred confidence never grants authorization, and missing model evidence never cancels current authorization.

5. Show the reading, then act

For implicit use, give a short outcome-first interpretation only when useful, then act. Do not turn every ordinary request into a long rewrite. Explicit invocations or substantial ambiguity get the full reading:

Reading: <the intended outcome in one line>
> <the brief, with tags on additions>
Changed: <only for a material, evidence-backed better approach>
Assumed: <only a consequential unresolved detail>
Next: <optional: the natural next step of the mission, offered rather than done>

Start authorized work in the same turn. Discuss means discuss; hold means hold. Ask only for a material missing fact unresolved from available context, following the user's question preferences. A reversible assumption has no arbitrary time limit. Optional MR log-decision saves the reading's evidence. MR foresight record freezes forecasts; only actual human labels establish outcomes. See prediction tools. Logging never stops useful work. Follow corrections immediately; MR correct records only what the human corrected. Do not fabricate feedback or silently rewrite the personal model.

Example (fictional)

Seed: the export is broken again; context identifies CSV. An echo supports checking related downloads. [L] Verify the real file and resume the queue. [S] More examples.