Agent Resource Discipline

How this skill is organised (progressive disclosure)

This skill follows the three-level progressive disclosure pattern codified by Anthropic's skill-creator (see "Adjacent prior art + lineage" below):

• Level 1 (always in context once the skill is loaded): this

SKILL.md, ~250 lines. Contains the universal rules + the decision-procedure for which references to consult.

• Level 2 (loaded on demand by name): the references/*.md files

-- one per discipline. Loaded only when a session actually exercises that discipline.

• Level 3 (planned future work): enforcement hooks (SessionStart /

PreToolUse / Stop) that mechanise the protocols here so they fire reliably without depending on agent discipline alone. Specification in section "Planned future work: enforcement hooks" below.

This is the same pattern the skill itself preaches: load only what you are about to use; defer the rest.

When to load this skill

Load this skill at the start of any session that will involve any of:

• reading or grepping more than ~5 files;
• handling PDFs (literature survey, related work, supplementary

material);

• editing or creating files in multiple project sub-directories;
• web fetching (publisher pages, arXiv, GitHub, doc sites);
• working across multiple agent sessions on the same project (where

cross-session memory matters).

In practice that's most non-trivial sessions in this ecosystem. The universal one-liners in ~/.scicomp-research-skills/AGENTS.md Section 6 cover the basics so cheap-and-fast rules fire even without this skill loaded; this skill expands them with the full how-to.

Why this matters

Agent tokens / quota / context-window are the scarcest resources in any non-trivial session. Beyond raw cost, the attention budget -- the agent's ability to pick the right detail out of its context -- degrades faster than the nominal context window suggests. Empirical work (Chroma's "context rot" study, cited by Anthropic in their Sep 2025 Effective Context Engineering for AI Agents post) shows that recall quality drops well before the window fills. Heavily-loaded contexts also introduce recency bias and goal drift. This skill therefore optimises for both raw token cost AND for keeping the working set small enough that the agent's attention stays sharp.

Default agent behaviour wastes resources in predictable ways:

• Tool mis-selection (bash grep instead of the dedicated Grep

tool, bash cat instead of Read, ...) costs tokens AND loses features (paging, structured results).

• Bulk reads (Read with no offset/limit on a 2000-line file when

50 lines would do) burn context window for no gain AND accelerate context rot.

• Re-derivation (re-reading a PDF that was already summarised in a

survey note last session) wastes tokens AND risks contradicting the prior summary.

• Forgetting (not reading PLAN.md / collection log / notes index

at session start) causes the agent to either re-do work or to make decisions inconsistent with prior sessions.

• Re-fetching (calling WebFetch on the same URL twice in one

session) burns external quota and adds latency.

• Goal drift in long sessions -- the original PLAN.md fades from

recent attention as conversation length grows. Manus calls this the "lost-in-the-middle" failure mode and addresses it via recitation (re-reading the plan into recent context).

These are all preventable with explicit rules. This skill codifies them as a research-flavoured operationalisation of the broader file-as-memory + just-in-time retrieval patterns now standard in the agent-engineering literature.

The five disciplines

This skill loads a small SKILL.md (you are reading it) and provides five per-topic reference files, each loaded on demand. Each codifies one resource-management discipline:

| Discipline | Reference file | When to load | |:-------------------------|:----------------------------------------|:--------------------------------------------------| | Tool selection | references/tool-selection.md | First time in this session you need a non-trivial file/search/edit operation. | | Targeted reads | (covered in tool-selection.md) | (same) | | PDF lifecycle | references/pdf-lifecycle.md | Whenever a session involves PDF intake or re-reading. | | Persistent memory | references/persistent-memory.md | Start of any session on a project with PLAN.md / collection log / notes index. | | Context-window budget | references/context-window-budget.md | When loading multiple skills, multiple reference files, or multiple PDFs simultaneously. | | Web-fetch discipline | references/web-fetch-discipline.md | Whenever WebFetch is called in this session. |

Load only the references relevant to the current session. Do NOT load all five at once -- that defeats the purpose.

Critical rules (apply unconditionally; do not require loading a reference file)

These are also in ~/.scicomp-research-skills/AGENTS.md Section 6, so they fire even if this skill is not loaded. Restated here for in-skill reference:

1. Use dedicated tools, not Bash equivalents.

• File search: Glob (not find / ls -R).
• Content search: Grep (not bash grep / bash rg).
• File read: Read (not cat / head / tail).
• File edit: Edit (not sed / awk).
• File create: Write (not cat <<EOF / echo >).
• User communication: response text (never echo / printf).

1. Batch independent tool calls into a single message. A message

with three independent Reads costs less and finishes faster than three sequential messages.

1. Read targeted, not bulk. For files >300 lines, use Grep first

to locate the relevant section OR Read with explicit offset+limit. The default 2000-line Read is for skimming, not routine consumption.

1. Re-use prior work before generating new work. Before re-reading

a PDF, check notes/survey_<citekey>.md. Before re-deriving a fact, check the audit log / notes / PLAN.md.

1. Indices are the persistent memory. Read PLAN.md status +

_collection_log.md + notes/README.md at session start; update them at session end if work was done.

1. Recitation in long sessions. For sessions exceeding ~50 tool

calls, re-read PLAN.md (or the relevant section thereof) every ~30-50 calls to combat goal drift. The Manus team identified this as the simplest defence against the "lost-in-the-middle" failure mode in long agent runs. Recitation is cheap; goal drift is expensive.

1. Do not edit AGENTS.md or system-prompt-equivalent files

mid-session. If the agent client uses prompt caching (Claude Code does, OpenCode does for Claude models), editing the cached prefix invalidates the cache and silently 10x's the per-token cost of all subsequent calls in the session. Restart the session if you genuinely need to change agent-facing rules.

1. Keep errors in the conversation; do not silently retry. When a

tool call fails (dead URL, rate limit, file not found), let the error sit in the conversation so the model adapts. Silent retry loops both burn quota and hide useful failure signal. For structural failures (a citation's PDF really is unobtainable, an arXiv ID is wrong), log to the appropriate audit entry (_collection_log.md "Items not found / left for user", PLAN.md "Open Questions") so the failure becomes part of the persistent record.

Note on prompt caching

OpenCode (and Claude Code, and Cursor) on Claude models supports prompt caching of stable prefixes (system prompt + tools + typically the most recently loaded skill content). Cached tokens are ~10x cheaper than uncached. Implication: re-loading a small skill via Read mid-session is cheaper than carrying its content forward in conversation, because the cached version pays cached-rate on every subsequent turn. This is part of why the progressive-disclosure model above works: levels 2 + 3 can be loaded fresh when needed without worrying that they'll dominate cost.

Common rationalizations + rebuttals

The agent will, in real sessions, invent plausible-sounding reasons to skip the disciplines above. The pattern is sufficiently consistent that we name + rebut the common ones explicitly. When the agent catches itself thinking one of these, it should treat that thought as a signal to STOP and re-evaluate.

| Rationalization | Why the agent thinks it | Rebuttal | |:------------------------------------------------------------------|:-------------------------------------------------------------------|:---------------------------------------------------------------------------------------------------------------| | "I already read this file last turn; I'll trust my memory." | Avoids the cost of re-Read-ing. | The file might have been edited (by you or the user). Read is cheap; recall is not always reliable. | | "It's just one extra bash cat, no big deal." | The override feels small in isolation. | This is the rationalization that turns a 200-token session into a 20k-token session. One bash-cat is fine; the habit isn't. | | "Let me re-read the PDF to make sure the survey note is right." | Healthy scepticism + low confidence in your own past summaries. | If you have specific reason to doubt the note, target-grep the .txt for the suspect fact. If not, trust the note; that's what it's for. Re-reading the whole PDF "to be safe" is the most expensive single action in this ecosystem. | | "I'll load all the section references now so I have them ready." | Tidy-up instinct; wants to "set up" before working. | Loading speculatively is the failure mode the context-window-budget exists to prevent. Load when you actually use. | | "I'll fetch the publisher page to confirm the year." | Wants external verification; doesn't trust local data. | The user verified the bib entry; that's what verification IS. Trust the bib unless you have specific reason to doubt it. | | "I'll skip updating notes/README.md; it's just an index." | The deposit feels like overhead at the end of a session. | The deposit funds the next session's withdrawal. Skipping it is the most expensive bug in this ecosystem. | | "I'll process all 14 PDFs now while I have momentum." | Wants to batch-finish a sub-task. | Process one at a time; close each before opening the next. The context-window cost of 14 simultaneous .txt files is much larger than the round-trip cost of 14 separate Reads. | | "Let me just retry that fetch, it might work this time." | Hope-based rather than evidence-based. | Twice per session is the cap. After that, log to "Items not found" and move on. | | "I'll silently fix this typo in the bib." | Helpful instinct; wants to clean up. | Silent fixes break the audit trail. Add a "Corrections to apply" entry; let the user batch-apply. | | "It's a small task; the protocol overhead would dominate." | Wants to skip first-action / last-action for speed. | A genuinely small task (one file edit, one question answered) is fine. Anything multi-file or multi-step earns the protocol's overhead back several times over. |

If you (the agent) find yourself thinking ANY of the left-column phrases mid-session, stop and re-read this table.

First-action protocol (every non-trivial session)

At the start of any session that touches a project with the standard layout (paper-skeleton or similar):

1. Load (in parallel, single message): AGENTS.md, PLAN.md

(status fields + open questions), references/_collection_log.md (verification status), notes/README.md (which surveys exist + their status). Total: 4 small reads.

1. Decide which skills the session actually needs (research-paper-

writing? literature-survey? human-facing-doc-authoring? this skill? often only 1-2 are relevant -- not all of them).

1. Decide which references this session needs from each loaded

skill (e.g. just references/introduction.md from research-paper-writing, not the whole references/ tree).

1. Then start the user's actual task.

Step 1 is cheap (4 small reads) and prevents the most common waste mode: doing work the previous session already did, or doing work inconsistent with what the previous session decided.

Last-action protocol (every session that produced work)

Before declaring the session done:

1. Update the indices that record this session's output:

• new survey notes -> add row to notes/README.md.
• new bibliography entries / verifications -> append to

references/_collection_log.md.

• status change -> update the relevant PLAN.md status field.
• new section drafted -> mark in PLAN.md outline + maybe add

notes/section_<N>.md.

1. Surface contradictions explicitly. If something this session

discovered contradicts prior notes / plan / bib entries, do not silently proceed; add a "Corrections to apply" entry to the relevant log.

1. Report to the user what was done + what indices were updated.

Steps 1+2 are the "deposit" that funds the next session's cheap "withdrawal" via the first-action protocol.

Output contract

When this skill is loaded, every action the agent takes should be auditable against the rules above. If the agent finds itself about to:

• run a Bash command that has a dedicated-tool equivalent -> stop and

use the dedicated tool.

• do a bulk Read of a >300-line file -> stop and either Grep first

or use offset+limit.

• re-read a PDF that has a survey note -> stop and read the note first.
• start work without reading PLAN.md / _collection_log.md /

notes/README.md -> stop and read them (in parallel).

• finish work without updating those same indices -> stop and update.

The goal is no avoidable waste, not "minimise tokens at the cost of correctness". When the rules conflict with correctness, correctness wins -- and the conflict gets logged as a "Corrections to apply" entry so the rule can be refined.

Tool-availability assumptions

This skill assumes the agent has tools approximately equivalent to OpenCode's Read, Grep, Glob, Edit, Write, Bash, and WebFetch. For agents with more limited toolsets:

• Shell-only agents (some Claude Code tool configs): use

pdftotext, rg, fd, sed/awk carefully (quote everything; prefer here-docs over echo chains; cap output with head/tail EXPLICITLY rather than relying on the agent's truncation).

• Agents without WebFetch: load references/web-fetch-discipline.md

for the protocol of caching fetches into the repo via shell commands (curl -> references/_cache/<hash>.html).

• Agents without parallel tool calls: serialise; the parallelism

rule simply does not apply, but the targeted-read and re-use-prior-work rules still do.

Adjacent prior art + lineage

This skill is a research-flavoured aggregation of patterns that have crystallised across the agent-engineering literature since mid-2025. Citations are given so users (and future maintainers) know what we borrowed, what we adapted, and where the genuinely novel pieces are.

Foundational sources (cited in the rules above):

• Manus team blog post -- Context Engineering for AI Agents:

Lessons from Building Manus (https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus). Source for the file-as-memory pattern (rule 5 of their post), recitation against goal drift (rule 4 -> our Critical Rule 6), KV-cache stability (rule 1 -> our Critical Rule 7), and keep-errors-in-context (rule 5 -> our Critical Rule 8).

• Anthropic engineering: Effective Context Engineering for AI

Agents (Sep 2025, https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents). Source for the "attention budget" framing in "Why this matters", the "context rot" empirical finding (Chroma research, https://research.trychroma.com/context-rot), and the hybrid pre-load + just-in-time retrieval pattern that underlies our first-action protocol.

• Anthropic engineering: Equipping agents for the real world with

Agent Skills (Oct 2025, https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills). Source for the three-level progressive-disclosure pattern this skill follows; informs the SKILL.md length budget guidance in references/context-window-budget.md.

• Anthropic engineering: Writing effective tools for agents -- with

agents (Sep 2025, https://www.anthropic.com/engineering/writing-tools-for-agents). Tool-author-side counterpart of our discipline 1; informs why dedicated Read/Grep/Glob are higher-signal than Bash equivalents.

• MemGPT paper (Packer et al. 2023, arXiv:2310.08560,

https://arxiv.org/abs/2310.08560). Academic ancestor of all subsequent file-as-memory patterns; introduces the OS-memory-hierarchy framing.

Closest comparable agent-rule skills:

• OthmanAdi/planning-with-files

(https://github.com/OthmanAdi/planning-with-files) -- 3-file workflow (task_plan.md / findings.md / progress.md) enforced via SessionStart / PreToolUse / Stop hooks. The hook-enforcement pattern is the model for our planned future work below; their evidence-based 96.7% pass-rate (vs 6.7% baseline) is what motivates shipping hooks rather than relying on agent discipline alone.

• thedotmack/claude-mem

(https://github.com/thedotmack/claude-mem) -- automated observation capture into SQLite + Chroma vector DB with lifecycle hooks. We deliberately diverge: our markdown-on-disk approach is git-diffable, blameable, and human-readable, which matters more for research reproducibility than the higher recall of an embedding store. Both approaches are valid; pick by use case.

• addyosmani/agent-skills

(https://github.com/addyosmani/agent-skills) -- 22 production- engineering skills with anti-rationalization tables in every skill; we adopted that pattern explicitly in "Common rationalizations + rebuttals" above.

• anthropics/skills/skill-creator

(https://github.com/anthropics/skills/blob/main/skills/skill-creator/SKILL.md) -- the authoritative how-to-write-a-skill skill; informs our three-level progressive-disclosure structure and the "pushy description" anti-undertriggering convention applied to this skill's own description.

• Future-House/paper-qa (PaperQA2)

(https://github.com/Future-House/paper-qa) -- programmatic RAG over PDFs (parse + cache + embed + retrieve). Solves the same problem as our PDF lifecycle (avoid re-reading PDFs) with a fundamentally different mechanism (vector embeddings vs human- curated survey notes). Our references/pdf-lifecycle.md discusses the trade-off explicitly; both approaches are valid.

Where we are genuinely novel (per the prior-art audit, no publicly-available agent-rule skill found that codifies these):

• PDF lifecycle as a per-session protocol for a research project,

with notes/survey_<citekey>.md as the agent-readable cache.

• Web-fetch discipline grounded in a research-paper context (bib

fields and survey notes consulted before WebFetch; arXiv abstract preferred over PDF page; cache to references/_cache/).

• The aggregation itself: no other skill we found packages all

five disciplines together with a coherent first-action / last-action protocol grounded in a research-paper-skeleton's specific file layout.

Adjacent ecosystems (mentioned for context, not directly borrowed):

• Cline memory bank

(https://docs.cline.bot/prompting/cline-memory-bank) -- one-task- one-goal + auto-compact + .clineignore patterns. We have no .agentignore analogue yet; possible future addition.

• MemGPT / Letta -- programmatic memory backend; conceptually

parallel but different abstraction layer.

• LangChain / LlamaIndex memory -- programmatic backends; out of

scope for an agent-operation skill.

Planned future work: enforcement hooks

The disciplines above currently rely on agent self-discipline. The single biggest reliability improvement available is to mechanise the first-action / last-action protocols as hooks invoked by the agent client, so the protocol fires regardless of whether the agent remembered to follow it.

OthmanAdi/planning-with-files reports a 96.7% pass-rate on Anthropic's skill-creator eval after introducing PreToolUse + PostToolUse + Stop hooks vs 6.7% without. Our expectation is similar gains for the first/last-action protocols, because the same failure mode -- agent skipping the bookkeeping when context fills up -- is what the hooks defend against.

Hooks are deferred for now (kept the repo simple at this stage). When implemented, the design should be:

Hook spec (for future implementation)

Layout (probably bin/hooks/ shipped in this repo, with users copying or symlinking into their per-project .opencode/hooks/):

text
bin/hooks/
├── session-start.sh       reads AGENTS.md + PLAN.md (status section)
│                          + _collection_log.md (Last updated +
│                          Corrections-to-apply) + notes/README.md
│                          (status section). Echoes a compact
│                          summary into the agent's startup context.
├── pre-tool-use.sh        on tool calls that risk losing work
│                          (Edit, Write, Bash with `git commit`),
│                          checks PLAN.md mtime > AGENTS.md mtime,
│                          warns if AGENTS.md was edited mid-session
│                          (Critical Rule 7 violation).
├── post-tool-use.sh       on Write / Edit calls under notes/ or
│                          references/, prompts the agent to update
│                          notes/README.md or _collection_log.md if
│                          the modified file is a new survey note or
│                          a new bib entry.
└── stop.sh                before declaring session done, checks
                           that PLAN.md / _collection_log.md /
                           notes/README.md have been touched if
                           survey notes / bib entries / experiment
                           dirs changed during the session. If not,
                           refuses to stop and prompts for the
                           last-action update.

Compatibility: the hooks should be CLI-agnostic shell scripts (no agent-client-specific assumptions in the hook bodies); each agent client (OpenCode, Claude Code, Cursor) wires them via its own hook mechanism, but the script bodies are portable.

Implementation order when picked up:

1. session-start.sh -- highest payoff, lowest implementation risk.

Single read-only pass over 4 files; echoes a status summary.

1. stop.sh -- second-highest payoff. Walks git status for changes

under notes/ / references/ / experiments/ and verifies the matching index was updated. Refuses to stop if not.

1. post-tool-use.sh -- soft prompt; less critical.
2. pre-tool-use.sh -- defensive against Critical Rule 7 violation;

only relevant for users iterating heavily on AGENTS.md.

Open questions to resolve at implementation time:

• Where exactly to install for OpenCode (.opencode/hooks/ per repo?

~/.config/opencode/hooks/ user-global?). Per-repo is more correct but requires users to copy explicitly.

• How to make the prompts actionable rather than annoying (the user

should never feel the hook is in the way of legitimate work).

• Whether to log hook fires to ~/.scicomp-research-skills.hooks.log

for debugging (probably yes, very small).

• Compatibility testing matrix: OpenCode + Claude Code + Cursor at

least.

When to revisit: when at least 3 real research-paper sessions have shown the agent skipping the first-action or last-action protocol despite the rules being loaded. Until then, the cost of maintaining hooks across multiple agent clients exceeds the benefit.

See also

• references/tool-selection.md -- dedicated-tools-vs-bash + targeted

read rules + parallelism rules.

• references/pdf-lifecycle.md -- one-shot pdftotext extraction;

survey-note-first lookup; section-targeted reads of .txt files.

• references/persistent-memory.md -- first-action / last-action

protocols for cross-session memory; the indices as memory.

• references/context-window-budget.md -- max-N skills / references /

PDFs simultaneously; when to summarise + close.

• references/web-fetch-discipline.md -- cache-first; bib-fields

before publisher page; arXiv abstract over PDF.

• The universal one-liners in

~/.scicomp-research-skills/AGENTS.md Section 6 are a strict subset of the rules above; that section is what fires for agents that have not loaded this skill.

Created 2026-05-13 by A. Attia. Distilled from observed waste modes across multiple agent sessions on the rl-oed paper + this skills repository's own bootstrapping. Revised 2026-05-13 (post-prior-art audit): adopted progressive-disclosure framing + pushy description (Anthropic skill-creator); added "Why this matters" framing around attention-budget + context-rot (Anthropic context-engineering post + Chroma); added Critical Rules 6-8 (recitation, no-mid-session-prompt- edits, keep-errors-in-conversation) from Manus; added prompt-caching note; added "Common rationalizations + rebuttals" table (addyosmani/agent-skills pattern); added "Adjacent prior art + lineage" section citing Manus / Anthropic / planning-with-files / claude-mem / addyosmani / paper-qa / MemGPT; added "Planned future work: enforcement hooks" with full spec (deferred implementation, specification kept here so future work has the design ready).