agentleFS
t1mericksont1merickson/skills/obsess-hero/SKILL.md

obsess-hero

Skill1 starsChanged 2 months agoMITView on GitHub

Build a bounded, source-accounted vault about one named person's published work and documented thinking. Use only when explicitly invoked.

The file

Show its rules and text

We split this skill into the rules it gives an agent

21 rules. About 0 are common, 10 are distinctive, and 11 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈10

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 11

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    If no person is named, ask.

    line 10
  2. Distinctive:

    Read two levers from the invocation.

    line 24
  3. Very specific:

    Read references/vault-contract.md when shaping the vault and copy the selected contract into SHAPE.md so the vault remains self-describing.

    line 58
  4. Very specific:

    Every captured artifact admitted to the hero vault belongs to exactly one sources/<slug>/ with SOURCE.md.

    line 62
  5. Very specific:

    Every sources/<slug>/ has exactly one owning wiki page whose local_path points back to it.

    line 63
  6. Distinctive:

    If work is already in flight, park its non-reconstructable state before switching.

    line 72
  7. Very specific:

    Write 00-braindump.md, or a dated sibling on an augment, before searching, opening sources, rereading the vault, or asking an agent to …

    line 77
  8. Very specific:

    After the blind dump, reread an existing vault's hero page, index, frontier, ledger, relevant notes, and actual holdings.

    line 83
  9. Very specific:

    Follow obsess's mile-marker protocol.

    line 87
  10. Distinctive:

    Read in the primary task. A selected subagent may find candidates or map a bibliography, but a source read only by an agent cannot support …

    line 100
  11. Very specific:

    Keep NOTES.md as a verdict log against the braindump: “had the date wrong,” “confirmed, and the primary work adds…,” “these accounts …

    line 106
  12. Very specific:

    Every lead not followed goes to FRONTIER.md immediately with its origin, expected value, estimated next-pull cost, status, and stop reason.

    line 108
  13. Distinctive:

    Use read-only finders and one central filer.

    line 114
  14. Distinctive:

    Every candidate clears the same gate:

    line 116
  15. Very specific:

    If a configured born-digital/scan gate exists, use only its read-only scoring or report entry point.

    line 119
  16. Distinctive:

    Never overwrite. Preserve collisions under distinct names until identity is resolved.

    line 121
  17. Distinctive:

    Do not substitute plausible works to hit a count.

    line 123
  18. Very specific:

    Read the vault contract, copy it into SHAPE.md, then apply brain-wiki.

    line 127
  19. Distinctive:

    Claims: every material claim resolves to its marked source and every quotation appears in an inspected source.

    line 149
  20. Distinctive:

    Layout: run this skill's read-only verifier:

    line 151
  21. Very specific:

    After context loss, read the hero page and INDEX.md before answering and reverify volatile claims.

    line 167

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Obsess hero

Invoke with /obsess-hero <name> or an equally explicit request to go deep on one person. The output is a durable person-shaped vault: their work, their ideas, and the documented relationship between how they lived, thought, and made things.

If no person is named, ask. Disambiguate them before creating files.

This composes three skills. Read each when its phase begins; do not replace their distinctive work with a summary here.

SkillUsed forChanged here
obsessthe braindump-first immersion and mile markersthe hero vault replaces the ordinary obsession index
offline-archivesubstantial web capturehero-owned captures land in sources/<slug>/; standalone archives use the configured archive root
brain-wikithe typed vaultthe vocabulary is already governed by this skill's contract

The anchor is a person, not a topic or a pile of files. Their life and work are one subject when sources connect them. Do not split the vault into unrelated “life” and “work” halves, and do not infer a connection merely because it would make a good story.

Levers

Read two levers from the invocation.

Depth defaults to brief.

BriefDeep
Researchone user-selected, high-yield pullseveral proposed paths, selected one pull at a time
Peoplethe hero and first-degree relationships that explain the workwider lineage, students, rivals, or influences only on selected paths
Worksthe works needed to establish the shapea route toward a fuller verified census
Frontierreport the rest after the first mile markerredraw and ask which branch to fund next

Angle is an optional direction such as “the music side” or “their teaching practice.” It orders the frontier; it does not soften the evidence standard inside that angle.

Neither lever licenses guessing. A brief run holds fewer facts, not softer ones. “Deep,” “everything,” and “exhaustive” expand the route map; they do not waive mile markers or promise completeness.

Vault layout

Resolve storage roots before creating files. When the workspace is itself the configured heroes root, place each hero directly at <heroes-root>/<person-slug>/; never introduce an additional heroes/ wrapper. Put standalone source archives at <archive-root>/<archive-slug>/, outside the hero vault. Prefer explicit user conventions, then existing sibling layout, and only then project-local defaults or a clarification question. Do not hard-code a machine-specific library.

<vault>/
  00-braindump.md
  00-braindump-<date>.md   optional blind delta on later runs
  NOTES.md
  SHAPE.md
  INDEX.md
  LEDGER.md
  FRONTIER.md
  OPEN-QUESTIONS.md
  wiki/<object-slug>.md
  sources/<capture-slug>/SOURCE.md
  tools/regen_ledger.py

Read references/vault-contract.md when shaping the vault and copy the selected contract into SHAPE.md so the vault remains self-describing.

Two admission rules defend against a junk drawer:

  1. Every captured artifact admitted to the hero vault belongs to exactly one sources/<slug>/ with SOURCE.md; no loose files or miscellaneous holding directories. Standalone archives stay at the resolved archive root instead.
  2. Every sources/<slug>/ has exactly one owning wiki page whose local_path points back to it.

Slugs are local to the vault unless the selected root is intentionally governed as one cross-person library. Do not silently impose a global namespace on unrelated vaults.

The ritual

0 · Resolve and set the contract

  • Name the person, angle, breadth, vault root, recency risk, time/request/token/storage budget, and stopping condition.
  • If work is already in flight, park its non-reconstructable state before switching.
  • An existing vault means augment, not restart. Locate it, but preserve the blind phase: write a dated braindump before rereading its index, notes, frontier, ledger, or holdings. Never overwrite an earlier braindump.

1 · Braindump, before research

Write 00-braindump.md, or a dated sibling on an augment, before searching, opening sources, rereading the vault, or asking an agent to investigate. Put down everything you already believe about the person, fast and dense. Mark it using obsess's ✓, ?, and ✗ notation.

For a person, pre-mark four fluent failure modes as ?: dates, titles of works, influence claims, and quotations. A quote that cannot later be pinned to a source actually inspected does not graduate into the vault. Do not revise the blind dump after research.

2 · Re-enter, build the agenda, and map routes

After the blind dump, reread an existing vault's hero page, index, frontier, ledger, relevant notes, and actual holdings. Notes are a map; sources/ is the territory. Re-mine relevant holdings before fetching replacements, then treat this run as a dated delta.

Open NOTES.md with the agenda built from the braindump: what to verify, fill, or update. Draft a works list as a hypothesis. Every entry begins unverified and either graduates against evidence or is struck; confidence alone never promotes it into the ledger.

Follow obsess's mile-marker protocol. Make the route chart person-specific and keep it to a few high-yield alternatives. Relevant paths may include:

  • works census and catalog identity;
  • direct study of primary works;
  • interviews and documented positions;
  • biography or working practice that explains the work;
  • reception, influence, students, and rivals;
  • recent work or changed status.

For each path show where it leads, the bounded pull, main-task token range, subagent roles and delegated-token range, requests or access dependencies, expected capture/storage growth, vault-writing cost, and stop point. Ask which path or paths to fund. Do not begin an unchosen branch or spawn its agents.

3 · Immerse and capture, one selected pull at a time

Read in the primary task. A selected subagent may find candidates or map a bibliography, but a source read only by an agent cannot support the hero synthesis.

Prioritize the person's own work and words, then long-form interviews, serious criticism or biography, and reference sources. A coherent author can reveal an arc that a crowd-edited page cannot. Use the latter for orientation, not as a substitute for the best available evidence.

Capture substantial public pages through offline-archive as you go rather than batching provenance at the end. A site or coherent corpus may be one source directory rather than hundreds of arbitrary fragments. SOURCE.md records what the capture is, where it came from, when and how it was captured, what changed in capture, and which page owns it.

Keep NOTES.md as a verdict log against the braindump: “had the date wrong,” “confirmed, and the primary work adds…,” “these accounts conflict because….” A note that merely restates a source is wasted.

Every lead not followed goes to FRONTIER.md immediately with its origin, expected value, estimated next-pull cost, status, and stop reason. Dead ends remain marked followed-and-empty so an augment does not walk them again.

At each mile marker, report what the selected path changed, coverage gained, actual observable spend, held versus merely identified works, corrections, and the redrawn frontier. Say not metered rather than inventing token usage. Then pause unless the user explicitly authorized an uninterrupted run bounded by paths, total budget, and stopping condition.

4 · Acquire files without losing their identity

Use read-only finders and one central filer. Searchers do not invent destination paths or write into the vault; the primary task applies the vault's admission rules.

Every candidate clears the same gate:

  1. It is what it claims: check format identity and, for text-bearing documents, confirm a usable text layer.
  2. If a configured born-digital/scan gate exists, use only its read-only scoring or report entry point. Never invoke a batch mode known to move files. If no gate is available, record gate: ungated rather than implying a check ran.
  3. Deduplicate by work identity and normalized title, not merely by path or filename.
  4. Never overwrite. Preserve collisions under distinct names until identity is resolved.

Do not substitute plausible works to hit a count. A work verified to exist but not held still earns a Work page with the appropriate contract status. Any actual purchase remains a separate user decision. Read file-batch-safety before a large acquisition or reorganization.

5 · Shape the vault

Read the vault contract, copy it into SHAPE.md, then apply brain-wiki. Only stop for shape approval when this hero genuinely requires a new type, relationship, or admission rule.

Three boundaries carry unusual weight:

  • Work versus Source turns on who made the artifact. An interviewer-made interview is normally a Source containing evidence from the hero.
  • Position pages require a specific statement or observed practice and a stated_in source. Unsupported psychological portraiture does not become a Position.
  • Idea versus Topic turns on evidenced ownership or advancement. Record only the influence direction the evidence supports.

The hero page is the entry point and the only page allowed to carry the long narrative synthesis. Other pages stay atomic. Do not enable an operational work-management pack; this vault is studied, not operated.

6 · Write

Serialize only claims supported by material this task inspected. In a brief run, write the pages directly. In a deeper run, optional writers must appear in a user-selected route; one writer owns each file, and writers only serialize the inspected evidence assigned to them. The primary task owns the hero synthesis, shape, index, ledger, and final evidence judgment.

Generate LEDGER.md from Work pages at the end of every pull. The first time, put the generator at tools/regen_ledger.py and name it in the ledger header; “do not hand-edit” is not a useful rule without something concrete to run.

INDEX.md lists every page grouped by type. Any hook is extracted verbatim from the page's body rather than composed afresh. A page unreachable from the index does not exist operationally.

7 · Verify

Run three passes:

  1. Claims: every material claim resolves to its marked source and every quotation appears in an inspected source. If an independent reviewer was funded in the selected route, give it an area it did not write; otherwise perform the mapping directly.
  2. Archive: replay web captures with the network blocked using offline-archive's verifier.
  3. Layout: run this skill's read-only verifier:

    python3 <skill-directory>/verify_vault.py <vault>

Fix findings where evidence permits; otherwise keep them visibly unresolved. A clean layout does not prove claim truth.

8 · Return, then teach

At the final return, report the person, angle, selected paths, pulls completed, coverage by source/work type, held versus identified works, major corrections, frontier, open questions, verifier result, observable spend, storage consumed, and stopping reason.

Then teach from the sourced pages: first principles, motivation before mechanism, real synthesis, and visible uncertainty. Never recommend one of the person's works from recall alone. When a question lands on OPEN-QUESTIONS.md, say so and offer the next bounded pull.

Housekeeping

  • After context loss, read the hero page and INDEX.md before answering and reverify volatile claims.
  • Hero vaults are durable outputs, not scratch. Never sweep or merge roots casually.
  • Non-hero material does not belong in the vault. Park it in OPEN-QUESTIONS.md until the user names a home; do not grow a miscellaneous branch.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.